This week, I thought I would share a philosophical perspective on the Intelligence Economy vs. the regular intelligence frameworks. The reason is that it’s a bit urgent. Although the thought process originated from my personal experience, I can sense it may apply to you as well.

I spent the vast majority of my time in front of devices. My morning begins with planning my day using Sameer OS around 6 am, reviewing my work calendar, and then attending meetings starting at 7 am also in front of a device.

Then driving to work, and apparently most cars now come with screens, so you are technically never offline, as we are always using some type of device.

Once I am back home, besides small time spent with family, I do the workout and still have something on my device so I can listen to when working out. Finally watching something on Netflix before going to bed.

All of my daily reading is also on devices as I crunch through the output produced by my AI agents to keep me up to date in the digital/AI world.

I review the information, develop ideas of key topics for the week, resolve any problems, and quickly move from one topic to another.

The AI tools are valuable, and I will be the first to tell you it has significantly made me more productive than before 2023. I am learning more and faster than ever before. I feel very confident knowing what I know now, and I am making data-backed decisions.

But I have been paying attention to a contradiction. The more information I consume, the less room I have to hear my strategic mind.

How did I find this?

The day before yesterday, I was driving without any radio or podcast, using my phone or any device. I also dimmed my car screen to eliminate any distraction.

In that apparently unproductive space, ideas began connecting, and I intentionally didn't capture anything on a device. I let the ideas flow, and at the end of my drive, there was this thought that captured my attention.

Boredom is wisdom!

Before you object, I want to clarify that boredom itself is not wisdom in its literal meaning.

The reason I coined that phrase is that it captures something important: When external inputs stop, internal synthesis can begin.

You may be better informed than at any previous point in your career, while having less uninterrupted space to think.

That is the paradox that carries the article.

Table of Contents

1. Boredom Is Wisdom

2. We Eliminated Boredom Without Understanding Its Function

3. The Intelligence Economy Has Changed What Is Scarce

4. What Happens When We Step Away

5. The Risk of Outsourcing Too Much Thinking

6. The Wisdom Loop (Exact Steps to Take Now)

7. What This Means for Leaders and Organizations

8. Conclusion: Creating Space for Wisdom

We Eliminated Boredom Without Understanding Its Function

After paying attention to my own routine, I started noticing that we have designed almost every empty moment out of our daily life.

Think about it for a second. The elevator ride, waiting at a doctor's office, or taking a lunch break. All of that time is filled with something on a device.

You're always going to be distracted by email, LinkedIn, useless videos, or, worst of all, the news. We have reached a point where even a few minutes of your time cannot go by without “wasting” it. Our phone gave us access to unlimited information but at the same time removed the natural pause that existed between one task and another.

Before 2008 (launch of iPhone), I used to walk from one office building to another without checking anything. I used to drive without listening to a podcast, and it’s funny, but waiting for someone was the only thing to do.

Today, the stream of information never ends. I don’t think this is only a customer behavior problem where people are spending too much time on social media and videos.

Executives and office staff suffer from the same problem, but the environment is different, and the screens are replaced by larger screens.

Instead of social media, you now have constant meetings, a barrage of unwanted emails within which actual useful emails are hidden in plain sight, teams/slack/zoom messages, alerts from your dashboards, and, to top that AI generated summaries.

Don’t forget we also have AI research assistants and agents that technically never stop working, producing reports overnight, monitoring the market, summarizing industry news or trends.

I am personally using these capabilities every day, so I am not saying it’s wrong.

But, the result is that there is almost never a natural stopping point anymore. One meeting leads to another one, one email leads to another, and so on.

So the problem is not necessarily that we don’t have enough information.

The problem may be that we have too much information moving through our mind without enough space in between.

Research on attention residue also shows that when we move from one task to another, part of our attention can remain attached to the previous task, which may impact how effectively we focus on the next one. ScienceDirect

I think this is what many executives may be experiencing without necessarily calling it that.

You may leave one meeting but still be thinking about something that happened there while already joining the next meeting. Although sometimes it can help connect the dots, it is not always making us more strategic or productive.

By the end of the day, you may have consumed a significant amount of intelligence, but your mind never really received the opportunity to fully process any of it.

But, at the same time, you may also be cognitively saturated.

This is the contradiction that is worth talking about. We solved information scarcity by creating integration scarcity.

The Intelligence Economy Has Changed What Is Scarce

In the previous economy (aka industrial or semi digital), access to information itself was a major advantage since the launch of the internet. If you had the original idea, then using industry research, leveraging experts, accessing analytics and data allowed you to make decisions that others could not make easily.

Information was not readily available, or at least not cheap. That’s no longer the case.

Right now, I can ask AI a question and get more of a summary of the entire industry or compare different companies and identify trends in the market. I can even challenge existing business models or poke holes in a scenario.

The information may not always be 100% accurate, and it still needs to be validated, but the amount of intelligence available to us has increased significantly.

That’s just the tip of the iceberg.

As AI agents become more capable, they will analyze internal data (inference), constantly monitor markets, identify any risks, and generate recommendations without any human intervention. I mean 0% touch from anyone.

The competitive advantage is no longer only about having access to intelligence because everyone has access to it as intelligence gets cheaper and compute is accessible even more.

The real question becomes: what do you do with all of that intelligence?

Because AI can provide you many options, but it cannot automatically tell you which one matters most. For example: it can summarize data from the latest research, but the data may be telling a different story. Someone needs to understand the context, risk, and impact after the decision in real life scenario.

This is where human judgment becomes more important, not less important.

I think we sometimes assume that as AI becomes more intelligent, there will be less need for human thinking because the technology will eventually provide the answers, recommendations, and probably tell us what decision we should make.

I don’t think that is what is going to happen.

The amount of intelligence available to us will continue increasing, but that may actually make the decision more difficult because now there will be more information, more options, and probably several different directions that all look reasonable.

AI can help me identify those options, compare them, and probably show me something I may have missed.

But someone still needs to decide what is important in that specific situation, what can be ignored and what the organization should actually do.

This is where I think judgment becomes more important.

Because judgment is not only based on the amount of information you have but also based on your experience, context, people involved, what happened before, and hidden bottlenecks that the data is not showing.

Two leaders can have access to the same information, AI tools, and analytical capabilities and still make completely different decisions. So the difference may not be what they know but how they understand what they know and what decisions they make after that.

This is also where Boredom is wisdom connects with the Intelligence Economy.

I am not saying that we should stop using AI or reduce the amount of intelligence available to us. I will continue using AI aggressively because it has significantly improved how I work, learn, and make decisions.

But I also need some space between receiving all of that information and deciding what I personally think about that situation or use case.

Otherwise, I may continue consuming more intelligence, but my own point of view may simply become a combination of the last report or my last meeting or the last AI generated response.

To me, that is the real risk. The Intelligence Economy is making intelligence more available and probably much less expensive than before, but our attention is still limited and judgment is not something that automatically increases because more information is available.

This is why I don’t see Boredom is wisdom as something outside of the Intelligence Economy. It may actually be part of the human side of it.

AI can continue producing the intelligence, but leaders still need some time where they are not consuming anything new so they can understand what all of that intelligence means for the decision in front of them.

What Happens When We Step Away

After the drive, I started thinking about what exactly was happening during the time when I was not consuming anything, because obviously my mind was not doing nothing. Right?

I was driving and paying attention to the road, but at the same time different ideas were coming into my mind without me intentionally thinking about any specific topic. One thought connected with another one, and eventually there was this phrase: Boredom is wisdom.

Honestly, I did not start the drive with the intention to develop a topic for this newsletter, nor was I trying to solve any particular problem. The only thing I did differently was that I stopped the information coming in deliberately.

This made me think about the number of times I was trying to solve something while sitting in front of my computer or discussing it during a meeting, and the answer did not come at that time. Then later I am driving, walking or doing something completely different, and suddenly the solution appears.

How did that happen?

The information was probably already in my mind, but for some reason it did not connect when I was directly trying to find the answer using AI.

I later found that researchers sometimes call this incubation. You spend time thinking about a problem, and when you step away from it, the mind may continue processing it without you intentionally focusing on it. I think something similar happened during my drive.

The thoughts were not 100% new because they were probably coming from my own experience with devices, AI, information overload, and everything else I have been thinking about as part of the Intelligence Economy.

But, frankly, I had not connected them in this way before.

Normally, when an idea comes to my mind, I immediately try to capture it. I will open my phone, send myself a message, put something in Sameer OS, or ask ChatGPT or Claude to help me develop it further.

This time I intentionally did not capture anything because I thought if I took out the phone, I would probably see another notification or something and get distracted.

So I just continued driving and let the thoughts go wherever they were going. This is also where I realized that taking a break from work is not always a break from information.

If I finish my work and start watching Netflix or browsing LinkedIn, I may be taking a break from work, but I am still consuming something.

So I was taking the required breaks from work, but almost no breaks from information.

I am not saying every time we stop using the devices an important idea will appear. There may be times when nothing useful happens, or the mind goes into completely random directions.

But if there is always something coming in, then there is probably less opportunity for the mind to do anything with what is already there.

That is what I was missing. Maybe it also needs some time where I am not asking it to process anything new.

The Risk of Outsourcing Too Much Thinking

As I was thinking about all of this, another thought came to my mind which was probably more uncomfortable.

How much of my own thinking am I giving to AI without realizing it?

Many times something happens during a meeting, or I have a topic in my mind, and before I have fully thought about it, I have already opened ChatGPT or Claude. I ask it to provide me with research or maybe a different perspective or identify what I may be missing, i.e., blind spots.

The response is normally very good, and then I continue from there. I ask another question, challenge something in the response, ask it to provide another point of view, or tell me what may be the best direction.

Within a few minutes, I have more information than I may have collected in several hours before generative AI. As I mentioned before, I find this extremely valuable, and I use it almost every day.

But, recently I started asking myself what exactly I was thinking before I received the first response.

Once the response appears, I immediately start thinking about what it provided. I may agree with it, disagree with it, or ask another question, but now the response is already part of my thought process vs generated.

So technically I am still thinking and making the final decision, but I don’t know what direction my mind may have taken if I did not receive the response immediately.

This does not mean I think we should stop using AI or first struggle with every problem for several hours before asking for help. That would not make sense and personally I dont want to work that way because that’s going in reverse.

But, I think I may be moving too quickly from having a question to receiving an answer. There is almost no space between the two.

It also matters more when the question is not only about finding information, and there is an actual decision involved.

The AI may provide the research, different options and probably a recommendation, but it does not know everything that may be going through my mind from previous experience, discussions, people involved and many other things that may not be part of the question I asked.

This is where the connection with boredom came back to me.

Maybe the empty space is not only helping different information connect. It may also be allowing me to understand what I think before asking another intelligence to tell me what it thinks.

I am not saying my own initial thought will always be right. No one can claim that for all of their thoughts.

But, at least I will know there was an original thought before the AI response arrived.

Otherwise, I may be asking better questions, receiving better answers and making faster decisions, while slowly becoming less clear about which part of the thinking actually started from me.

The Wisdom Loop

After thinking about this more, I started asking myself how I actually use this?

The issue was that I was constantly consuming it without creating enough time to understand what I personally think about it. So I went back to what happened during the drive and started looking at the process.

I did not plan any of these steps at that time, but apparently there were few things happening.

Acquire

Before the drive, I had already consumed a lot of information, i.e., 16/7 device access, AI agents working 24/7, emails, media, and all.

So the information required for the thought was probably already there. This is why I dont think the answer is to stop consuming information. You still need the intelligence before you can do anything with it.

But, I think it helps to know what question you are trying to answer or what problem you are thinking about because otherwise the research can continue forever. Every AI response gives you another five areas you can explore and there is technically no ending point.

Disconnect

At some point, I stopped adding more. There was no podcast, radio, phone call, or AI response coming in during the drive. I was not intentionally following a process. I just decided I did not want anything playing in the background.

Apparently this was the important part because the information will probably never stop by itself. There will always be another report or article or problem I can ask AI.

At some point I have to decide I have enough information for now and stop consuming something new.

Synthesize

After some time, the different ideas started connecting. I was not intentionally trying to connect them, and I don’t remember everything that was going through my mind.

Some of it was probably random. But the devices and AI tools overload, and the feeling that I was not hearing my strategic mind enough were already somewhere in my thinking. During the drive, they started coming together.

This is what I mean by synthesis.

It was my own mind trying to understand what all of the information and experience meant together.

Distill

At some point, there was this phrase that captured my attention.

Boredom is wisdom.

As I mentioned before the phrase is not literally correct, but it helped me capture what I was trying to understand. When external inputs stop, internal synthesis can begin.

Without the phrase I may have continued thinking about several different things and probably forgot most of them by the end of the drive.

The phrase gave the thought some shape.

Decide

Then there was the question about what I should change based on this.

Because having an interesting idea during a drive does not mean much if I return to exactly the same routine the next day.

For me the change is probably very simple and hidden in plain sight.

Not every drive needs a podcast, phone call or music playing in the background. I dont need to use every few minutes between meetings to check another email or review another AI output.

Maybe sometimes the workout can happen without me trying to learn something at the same time. By doing this, I am not doing less work, but I am simply creating a few places in the day where I am not adding more information.

Validate

After the drive I came back and started questioning whether the thought was actually correct. Something coming to my mind during a quiet moment does not automatically make it wisdom.

So I looked at the research, discussed the idea with AI, and started testing whether there was something real behind my personal experience. This is where AI becomes useful again because I can ask it to challenge the thought, identify what may be wrong and provide a different point of view.

So AI is still inside the process. It still helps me gather the intelligence and then later helps me challenge what I concluded.

The difference is that there is some time in the middle where I am not asking it to produce another answer.

I am currently calling this the Wisdom Loop.

Acquire the information, disconnect from the inputs, allow the thoughts to connect, capture what appears important, decide if anything should change and then go back to challenge it.

I am still thinking through the framework, and it may change over time.

But, for now, it helps me understand what happened during that drive and why constantly consuming more intelligence may not automatically create better thinking.

What This May Mean for Leaders and Organizations

Once I started paying attention to this in my own routine, I also noticed the same thing may be happening during the regular workday.

Most of my work is based on consuming information.

I attend a high number of meetings or review presentations, discuss problems, check dashboards, and receive updates from my org. I also have AI generated email summaries, meeting notes, research, and outputs produced by agents while I am probably attending another meeting.

All of this is useful, and I will be the first to tell you it has helped me understand things much faster than before. But the same contradiction is there as before.

At the end of the day I may have attended eight or ten meetings, reviewed several documents, resolved some problems and consumed a significant amount of information.

But, when exactly did I think about all of it?

I think this becomes more important as you move into senior leadership because the amount of information coming to you continues increasing, and at the same time people are waiting for your decision, opinion, or direction.

There is normally no shortage of information, but actually be too much of it coming from different directions and not enough time to understand which part matters the most.

But, if AI saves me thirty minutes and I immediately fill those thirty minutes with another meeting, report, or problem, then I am not sure I created any additional space to think.

I only used the time to consume something else.

I dont think organizations are intentionally creating this problem.

We are simply taking advantage of the additional capacity and apparently using every available minute because there is always more work that can be done. The result is that a leader may be more informed than ever before while also moving too quickly between different issues to develop a clear point of view on any of them.

This is where the few minutes after an important meeting may matter more than I previously thought.

Now I am questioning whether sometimes I should not do anything immediately.

Maybe I need to let the discussion stay in my mind for few minutes and see what part continues coming back.

The same applies when I have a difficult strategic question.

Instead of immediately asking AI for another perspective or listening to something during the drive, I may take only that question with me and see where my own mind goes.

I dont know if every question will produce an important answer. Most likely it will not.

But, I also know that if I continue providing more inputs every few minutes, I am not really giving it a chance. This may look unproductive from the outside because there is no email being answered, document being produced or meeting happening.

Maybe that is why it is difficult to protect inside organizations because organizations vote for visible activity. Visible activity is much easier to recognize than someone apparently doing nothing and thinking.

But, executive work is not only about consuming the information and responding faster.

I think we are spending significant amount of time improving how quickly the intelligence can be produced, but probably not the same amount of time thinking about what happens after it reaches the leader.

• Does the additional intelligence reduce the uncertainty or simply create another five questions?

• Does the AI summary create clarity or another output that needs to be reviewed?

• When every tool is helping us move faster, at what point do we decide that moving faster may not be the only thing required?

I don’t have the complete answer yet. But, I am starting to pay more attention to the small gaps between those activities.

I will continue using AI, devices, and all of the intelligence available to me because I dont want to go back to the way things were before 2023.

But, I am also starting to understand that constantly consuming more information may not automatically help me think better.

Sometimes I may need to stop, allow the thoughts to connect and give my strategic mind enough time to tell me what it already knows. Maybe boredom is not wisdom itself, but apparently it may be the space where wisdom begins.

Talk soon,

Sameer Khan

Intelligence Economy

Here is a question I have been thinking about.

If every constraint on your AI project disappeared tomorrow except one, which one would still delay it?

Most people would probably say GPUs.

I would have said the same thing a year ago. But something changed my thinking.

We were in a meeting about expanding our AI infrastructure when someone raised a concern about our significant dependency on one of the major AI providers behind the copilots we use.

It was a valid concern. But it also made me think about dependency differently.

We experience a copilot as software. We open it, type a prompt, and expect an answer.

Underneath that simple experience is a lot of physical infrastructure.

• Data centers.

• GPUs.

• Power.

• Transmission.

• Cooling.

• Transformers.

• Permits.

And people who know how to build and operate all of it.

Our dependency was not limited to one AI provider.

We also depended on whether that provider could keep expanding the physical infrastructure underneath its software.

That is a very different problem.

Table of Content

1. The bottleneck has moved below the compute layer

2. What’s Actually Happening

3. Why “The Stack” Isn’t Ordered, and Why That Matters

4. The Constraint Stack Audit (Download the Constraint Audit Calculator)

5. What This Means Depending on Where You Sit

6. Where This Leaves You

The bottleneck has moved below the compute layer

One lesson the energy industry has taught me is that physical reality always wins.

Technology teams often assume infrastructure will scale as demand grows because that is how the cloud era worked.

Need more storage?

Buy more.

Need more compute?

Increase your cloud capacity.

The physical complexity stayed mostly hidden from the customer. Energy does not work that way.

Power, transmission, permitting, and supply chains move on multi-year timelines.

No amount of software can compress them.

That is why I spend less time asking what the best technology is and more time asking what the real constraint is.

Because the constraint is what determines how fast anything can scale.

I have been in attendance at these conversations. The conversation always narrowed to GPU availability and compute contracts. Nobody was talking about interconnection queues.

It’s also increasingly wrong, and the market is now telling on itself.

H100 cloud rental prices have fallen 64 to 75 percent from their 2024 peak. Chip supply, in the narrow sense of raw GPU availability, has genuinely eased. If chips were still the binding constraint, that easing should be translating directly into faster AI deployment timelines.

It isn’t.

Source: SemiAnalysis H100 Rental Price Index, IntuitionLabs and Thunder Compute 2026 GPU pricing trackers.

Between 30 and 50 percent of scheduled 2026 U.S. data center capacity is now expected to be delayed or canceled, up sharply from a historical baseline where project cancellations went from 6 in 2024 to 25 in 2025. Something is gating delivery that has nothing to do with chip availability.

That something is the physical and regulatory layer underneath the compute layer, and it is currently the least-modeled part of almost every AI capital plan I’ve seen.

What’s Actually Happening

For the last two years, almost every AI infrastructure conversation has focused on chips.

That focus made sense. GPU supply was tight. Prices were high. Companies were competing for allocations.

But access to some types of compute has started to improve.

The other timelines have not. A data center still needs access to a large and reliable block of power.

It may need a new grid connection. It may need a substation, transformers, switchgear, cooling equipment, permits, and utility upgrades. Those things are not delivered at software speed.

Think about it.

Interconnection queues, the process by which a new facility gets connected to the grid, are now running 4 to 7 years in major hubs like Northern Virginia, Phoenix, and Dallas.

PJM data shows projects spending roughly 3 years just reaching an interconnection service agreement, then another 4 years waiting to actually go live.

Even after approval, you’re not done. High-power transformers now carry 3 to 5 year delivery schedules. Switchgear is sold out through 2028. Substation transformer lead times have climbed from about 140 weeks in 2023 to over 160 weeks in 2026.

Compare that to chip lead times.

Data-center GPU lead times currently run 36 to 52 weeks. Do the math, and you get something close to a five-to-one gap between the constraint everyone is managing and the constraint that actually governs when the facility goes live.

Five to one. Most project plans don’t account for that ratio.

CBRE’s 2026 Data Center Outlook makes clear that demand in 2026 is no longer gated by capital, connectivity, or land. It’s gated by the ability to deliver large blocks of power on an aggressive timeline.

Northern Virginia’s data center vacancy has fallen to 0.3 percent. Atlanta sits at 1 percent. There is functionally no slack left in the markets everyone wants to build in.

Then there’s water, which is the constraint almost nobody is pricing into their models yet.

Hyperscale campuses can consume 1 to 5 million gallons a day. U.S. data centers used an estimated 17 to 20 billion gallons directly in 2024 and 2025, and that figure is projected to roughly quadruple by 2030.

More than 20 states are now weighing data center water bans or moratoriums. In the first quarter of 2026 alone, more than $130 billion in projects were delayed or abandoned over water-related issues, more than all of 2025 combined.

Source: Photo by Erik McGregor/LightRocket via Getty Images

When I think about this from inside a Fortune 100 energy company, the physical constraint problem isn’t new.

We’ve been dealing with multi-year infrastructure lead times in oil and gas for decades. What surprises me is watching the tech industry rediscover what heavy industry already knew if you don’t control the physical inputs, you don’t control the timeline.

But if the site cannot be energized, you do not have an operating AI facility.

You have expensive equipment waiting for electricity.

That is why the difference between the GPU timeline and the power timeline matters.

Your procurement team may be managing to one clock while the project is actually governed by another.

I’ll be straight with you about the counter-signal, because ignoring it would make this argument weaker, not stronger. Sam Altman said it plainly this year: “It goes back and forth. Right now, again, it’s chips.”

He’s not wrong. HBM memory lead times for data-center GPUs still run 36 to 52 weeks, and TSMC’s CoWoS advanced packaging capacity remains sold out into 2026, with Nvidia alone holding over 60 percent of that output.

The honest picture is that chip and packaging constraints govern how fast you can procure compute.

However, power, water, and permitting constraints govern how fast you can physically build the place that houses it.

Those are two different bottlenecks gating two different things, and most boards are still treating them as one undifferentiated “AI scarcity” problem. That conflation is exactly why capex gets approved against timelines that were never achievable.

Why “The Stack” Isn’t Ordered, and Why That Matters

I am not saying chips are no longer a constraint.

They are.

High-bandwidth memory, advanced packaging, and demand for the latest GPUs can still slow a project down. But chips govern how quickly you can acquire compute.

Power, water, construction, and permitting govern how quickly you can operate it at a particular site. Those are not the same question.

And the constraint can change over time. Today it may be GPUs. Tomorrow it may be power.

After that, it may be cooling, labor, or a permit.

So I do not think this is a fixed stack that we solve from top to bottom.

It is a network.

When I started outlining this piece, I wanted to give you a seven-layer diagram: power, interconnection, water, chips, labor, permitting, capital, stacked in a fixed sequence, solve them top to bottom. I used to think that kind of clean sequencing was the most useful thing I could hand an executive.

I’ve revised that view, and I want to tell you why, because the reasoning matters more than the diagram would have.

A fixed sequence implies that once you clear layer one, layer two is next in line, waiting its turn.

The Southern Nevada example alone disproves that. Water and power aren’t queued; they’re coupled. Fixing one changes the other. Altman’s own comment tells you the same thing from a different angle: which constraint binds hardest isn’t fixed over time, it oscillates depending on where the market is in a given cycle.

A rigid stack order is the single easiest thing for a sharp reader, or a McKinsey partner, to dismantle, because it will be visibly wrong the moment they bring their own project to it.

So here’s the more accurate, more useful model.

These constraints aren’t a stack you climb but a network or process you manage.

Power, interconnection, water, chip and packaging capacity, labor, permitting, and capital all interact, and resolving one can tighten another.

Although the visual metaphor of a stack is still useful for communicating the idea of “underneath the layer you can see, there are more layers,” the moment you treat it as a strict sequence, you’ve built something fragile.

Treat it as a system of tradeoffs you manage continuously, and it holds up against scrutiny from people who actually run these projects.

This is also, not coincidentally, a lesson heavy industry already learned.

For example, Aluminum smelters were sited next to hydropower before roads existed to reach them, because power was cheaper to build around than to transport.

Nobody in that era thought of “power access” as one item on a sequential checklist behind capital and permitting. It was the organizing constraint the entire site decision was built around, and everything else, roads, labor, logistics, got engineered to follow it.

What looks like a 2026 AI infrastructure innovation, on-site gas plants, fuel cells, small modular reactors, is really a rediscovery of a century-old industrial siting discipline.

Oracle’s partnership with Volta Grid to deploy a 2.3 gigawatt modular gas fleet.

Crusoe Energy building dedicated on-site gas plants.

Equinix’s agreement with Bloom Energy for over 100 megawatts of fuel cells across 19-plus sites.

Oracle’s own plans to power a 1 gigawatt facility with small modular reactors.

Source: VoltaGrid

These are not ad-hoc decisions. It’s the predictable response of sophisticated operators who’ve concluded that owning generation beats leasing grid access.

Watching who owns power generation versus who’s still waiting in an interconnection queue is now a better predictor of who wins in AI infrastructure than watching who has the most GPUs.

This is also where this week’s argument connects directly back to the framework underneath everything I wrote in the Intelligence Supply Chain post, and its core claim that intelligence is manufactured, and value migrates to the bottom of the stack.

Now you can see precisely why that migration happens. Value moves downward because the bottom of the stack is where the multi-year, non-substitutable lead times live.

Whoever controls that longer, harder-to-substitute layer, or can work through the lead times to secure it, captures a disproportionate share of the value the whole system creates.

I wanted a simple way to make this problem visible to executives.

So I came up with something I call the Constraint Multiple.

The calculation is simple:

Constraint Multiple = longest critical-path lead time ÷ lead time used in the current project plan

Let’s say your team is planning around a one-year GPU procurement timeline. But the realistic power-delivery timeline for the site is five years. Your Constraint Multiple is five.

That number does not tell you everything, but it tells you something important.

You are managing the project against the wrong timeline.

In this market, that number is currently running somewhere around 5.

The Constraint Stack Audit

Before approving the next AI infrastructure investment, look at every major constraint.

• Power.

• Interconnection.

• Water.

• Chips and packaging.

• Labor.

• Permitting.

• Capital.

Score each one in three areas.

1. Lead time: How long will it realistically take to resolve?

2. Severity: Will it stop the project, reduce capacity, or just increase cost?

3. Mitigability: Can you address it through investment, contracting, redesign, or a different location?

Then rank the constraints by their impact on the go-live date. The output should show you which constraint is really controlling the project. It should also show you which constraint is likely to appear next. That matters because a completed checklist can create false confidence.

Seven green status boxes do not mean the facility will open on time. The real question is whether the slowest critical dependency supports the date in the presentation.

Here is a Constraint Multiple calculator I created.

Below you will find a link to download this calculator.

Read more

When you ask a room full of executives what’s stopping their AI rollout, almost everyone points at the same thing. Model quality. Talent. Governance. Maybe budget. Almost nobody says power.

That’s the problem. Because in 2026 and beyond, power is the answer.

For the last decade, every technology rollout you’ve ever run followed the same quiet sequence.

• Build the product.

• Prove the demand.

• Scale the infrastructure underneath it once you know it works.

Cloud made that sequence feel almost automatic. Need more compute? Spin up another instance. Need more storage? Click a button.

Infrastructure became elastic, invisible, and something you paid for after the fact, not something you planned for years in advance.

AI just broke that sequence.

How?

Here is the uncomfortable truth sitting underneath every AI roadmap right now: the companies that will have usable AI capacity in 2028 already made that decision in many cases, back in 2024 and 2025, when the workloads that will eventually run on that capacity weren’t even fully specified yet.

Power, land, cooling, permitting, and grid interconnection now have to be secured years before the intelligence layer on top of them is designed.

The infrastructure comes first, and the product comes later. That is a full inversion of the sequence you learned, and most executives are still planning against a reality that no longer exists.

I call this the Infrastructure Inversion, and once you see it, you can’t unsee it in every AI budget conversation you’re having right now.

Table of Content

1. The Infrastructure Inversion

2. Why Power Became the Bottleneck

3. Intelligence Is Manufactured

4. The Readiness Curve

5. Enterprise Playbook

6. Mid-Market Playbook

7. Key Takeaways

What Changed, and Why It’s Happening Now

The bottleneck in AI stopped being about model capability a while ago. It is now, almost entirely, about physical capacity. Megawatts. Substations. Water rights. Buildout timelines measured in years, not sprints.

This might be an unusual statement, as you normally hear about models and compute.

Grid interconnection queues in major U.S. markets now average 5-7 years. In Virginia, the epicenter of American data center growth, Dominion Energy’s queue for large commercial load additions stretches beyond three years just for substation service, with some 100 megawatt connections reportedly waiting seven years to get power flowing.

ERCOT in Texas is sitting on a 410-gigawatt large-load interconnection queue as of this year, and data centers now account for roughly 73% of it. Nationally, the U.S. interconnection backlog has ballooned to around 2,600 gigawatts.

Compare that to hyperscale data center construction itself, which now runs eighteen to twenty-four months, up from about twelve months before the current squeeze. Even that number keeps stretching, because the equipment inside these buildings has its own bottleneck.

Large transformers, the kind that step generation down to usable voltage, now take up to 144 weeks to deliver. That’s 3 years just for transformers.

This is why hyperscaler capital spending looks the way it does. Combined 2026 capex across Amazon, Google, Meta, and Microsoft is around $ 725 billion, up roughly 72% year over year.

Analysts expect that figure to cross a trillion dollars in 2027. Numbers that large only make sense if these companies believe capacity, not demand, is the thing standing between them and their next competitive move.

The four largest, most capital-rich, most technically sophisticated companies on earth have stopped waiting on the grid entirely.

If I say they are becoming power companies, it will not be incorrect.

• Meta signed a 20 year power purchase agreement with Vistra plus forward commitments with TerraPower and Oklo, adding up to 6.6 gigawatts of firm, round the clock nuclear power.

• Microsoft signed a 20 year 1.6 billion deal with Constellation Energy to restart Three Mile Island, targeting 2027.

• Amazon committed 50 billion dollars to a partnership deploying 960 megawatts of small modular reactors with X-energy.

• Google secured a 615 megawatt power purchase agreement to help restart a decommissioned nuclear plant.

Across the sector, disclosed nuclear commitments now total roughly 9.8 gigawatts, and industry watchers expect close to a third of all data centers to be fully off grid by 2030.

Think about what that means for a moment.

They committed tens of billions of dollars to secure power for workloads that don’t fully exist yet. That is a company betting its capital on the belief that the physical layer, not the software layer, is where the real constraint lives.

Microsoft’s own CEO confirmed on a recent earnings call that internal workloads and the largest enterprise commitments are now being prioritized over on demand customers.

Capacity allocation, not just pricing, is now a lever vendors are actively pulling. If you’re a mid-market company buying AI as a service, you are, by definition, further back in that line than you think you are.

Applying the Framework: Intelligence Is Manufactured

If you’ve been reading this newsletter, you already know the core thesis of the Intelligence Economy is that Intelligence is manufactured. Value migrates to the bottom of the stack.

The Infrastructure Inversion is Lens #1 of the Intelligence Supply Chain. It’s the operational mechanism behind the claim we’ve been building toward. Value migrates downward, toward power, land, and grid access, because that’s the layer with the longest lead times and the highest capital intensity.

Long lead times plus capital intensity create a moat. An actual, physical, multi-year moat that a competitor with more capital and more urgency cannot simply buy their way around, because the constraint isn’t money but a queue position.

That distinction matters more than it sounds like it should.

Most executives hear power shortage and think in market terms. Pay more, get more. But grid interconnection doesn’t work that way. It is a queued, non-parallelizable resource. You cannot outbid your way to the front of a 7 year interconnection line the way you can outbid a competitor for scarce chips.

This reframes the entire executive question. It’s not how much should we budget. It’s when did we get in line.

Here’s a second piece of this that even sophisticated readers usually miss.

For years, the assumption was that semiconductor capacity could always scale faster than physical grid capacity, because chip fabs are the fast, flexible lever and the grid is the slow one.

That assumption just quietly died.

U.S. semiconductor fab construction now runs 36-60 months for a fully operational facility, whereas grid interconnection for large industrial loads runs 3-7 years. Those two numbers have converged.

The AI industry has lost its fastest scalable infrastructure lever at exactly the moment it needs it most, and almost nobody outside specialist supply chain circles has connected those two facts out loud.

Fab capacity is parallelizable. You can build five fabs in five different places at once if you have the capital, the way the CHIPS Act tried to encourage.

Grid capacity in a given region is shared and queued. Every data center, every manufacturer, every electrification project in that region is competing for the same limited interconnection slots.

That’s the term worth carrying forward from this lens: parallelizable versus queued capacity. It’s a more honest, more useful distinction than the old fast chips, slow grid story, and it applies to more than just power.

Water rights, specialized construction labor, permitting staff capacity. Anywhere a bottleneck is shared and queued rather than something you can simply buy more of, this lens applies.

I want to be straight with you about something, because a framework that’s never confronted its own weak points isn’t worth much.

There’s a real counterargument here, and it deserves airtime.

PJM, the largest grid operator in the country, has publicly walked back parts of its own AI demand forecast, saying its project pipeline included data center announcements without firm construction commitments attached.

That’s the same pattern that produced the fiber glut of 2001, when 1990s telecom carriers spent roughly 2 trillion dollars and laid 8-9 million miles of fiber, betting on internet traffic doubling every ninety days.

Actual traffic growth was closer to doubling annually, a real number, just far below the story used to justify the spending. 85% of that fiber sat dark for four years after the crash.

Here’s the nuance worth holding onto, because it’s more sophisticated than either the hype version or the skeptic’s version of this story.

Being wrong about the growth curve and being wrong about the need for the infrastructure is a double whammy.

The 1990s carriers were wrong about the speed of demand. They were not wrong that the internet would eventually need that fiber. It became the physical backbone of the modern internet economy, just years later than the pitch deck promised, and only after a brutal capital destruction cycle wiped out the companies that moved first and misjudged the curve.

That’s the honest version of the Infrastructure Inversion.

Building ahead of proven demand is directionally necessary. It is also genuinely risky for whoever moves first and gets the growth curve wrong. Both things are true at once, and any framework that only tells you one side of that isn’t giving you a real decision tool. It’s giving you a pep talk.

Where Do You Actually Stand

Before we get to what to do about this, it’s worth asking a more honest question first.

Where does your organization actually sit on this, right now, today?

I think about readiness here on a simple 4 curve, and most companies I talk to can place themselves on it in about ten seconds once they hear it described.

1. Stage one is buying spot capacity with zero visibility into your vendor’s own infrastructure position, which is where most mid-market companies and a surprising number of enterprises still sit, often without realizing it.

2. Stage two is asking the right questions: you’ve started probing vendors about capacity commitments during renewal conversations, but you haven’t yet built any contractual protection around the answers.

3. Stage three is where you’ve secured some contractual capacity or priority guarantees, at least for your most critical workloads, so you’re not entirely exposed to a single vendor’s internal prioritization decisions.

4. Stage four is where the four hyperscalers already sit, multi-year capacity secured directly, with energy strategy fully integrated into the technology roadmap rather than treated as someone else’s problem.

Almost nobody starts at stage four, and that’s fine.

What matters is knowing honestly where you are, because the gap between stage one and stage three is closeable in a single contract renewal cycle if you ask the right questions now, while you still have leverage.

The gap between stage three and stage four takes years and, for most organizations outside the hyperscaler tier, simply isn’t a realistic or necessary destination.

I want to make it super clear that your job isn’t to become a power company but the last one in line when the company you depend on has to decide who gets served first.

What This Means for Enterprise

Look at how the four hyperscalers actually answered this question, because each of them chose a different path and each path is now a named, dated, real world case study you can map your own organization against.

• Meta bought long-term power through a PPA.

• Microsoft partnered by restarting an existing asset at Three Mile Island.

• Amazon built, taking a direct ownership stake in small modular reactor capacity.

• Google bought again, through a PPA that helped restart a decommissioned plant.

Four of the most capital rich, best informed companies in the world, with every reason to simply wait for the grid to catch up, instead chose to become power companies.

That decision, made with tens of billions of dollars rather than commentary, tells you more about the true state of this constraint than any forecast, including PJM’s own walk-back.

The practical move: get infrastructure, energy, and finance functions into your AI strategy planning as standing members this quarter, not as a downstream approval step brought in after the roadmap is already set.

Then pick one of those four hyperscaler strategies, PPA, restart and partner, direct build, or forward contract, and evaluate it as a live option for your organization, not a hypothetical you’ll revisit someday.

What This Means for Mid-Market

You don’t need to build a power plant. You need to know whether your vendor already has.

Practically, that means treating vendor infrastructure exposure as a named risk category in every AI and cloud contract renewal this year, the same way you already treat cybersecurity or data residency.

A short checklist is worth running through before your next renewal conversation: has the vendor disclosed multi-year power or compute capacity commitments, or are they buying capacity on the spot market the way everyone else is?

• Does the contract include actual capacity or priority guarantees, not just an uptime SLA that means nothing during a genuine crunch?

• Is your pricing fixed, or exposed to the kind of usage-based increases we’re already seeing show up quietly in enterprise software bills?

• Are you single-sourced on a vendor whose infrastructure position you honestly can’t answer questions about right now?

None of that requires capital most mid-market firms don’t have. It requires asking questions before you sign, while you still have negotiating leverage, instead of after you’ve been throttled and have none.

The Takeaway

Here is what I want to leave you with.

The gap between the companies with usable AI capacity in 2028 and the ones queued behind them is being decided right now, in 2026, not in some future budget cycle you haven’t planned yet.

You are not early to this decision. You are either on time, or you are already late, and the difference was set months ago, not years from now.

So here’s the question worth bringing to your own leadership team this week. If someone asked you right now when your organization got in line for the capacity it will need in 2028, what would your honest answer be? Not your vendor’s answer. Yours.

That question is Lens #1 of the Intelligence Supply Chain.

There are more lenses coming, each one taking a different piece of this stack and turning it into a decision you can actually act on. Next time, we go one layer up.

Talk soon,

Sameer Khan

Intelligence Economy

Every few months, a headline tells you AI just got dramatically cheaper. A new model launches at a fraction of the previous price.

Executives read it as good news, because in almost every other part of business, falling input costs are good news. Cheaper steel helps the carmaker. Cheaper freight helps the retailer, and so cheaper intelligence should help everyone who buys it.

Except it doesn't work that way, and the companies finding this out right now are finding out the hard way.

Source: https://tokencost.app/blog/ai-price-index

Roughly 40 percent of AI startups have shut down within roughly 24 months of launching, and the trajectory points toward something close to 80 percent by the end of this year.

Nobody generally picked a bad model in these startups, but what killed them was believing that access itself was the advantage.

Last week I introduced Framework 01, the Intelligence Supply Chain, and its central law: value migrates toward the constraint, not toward the customer.

As I thought about that framework a bit more, I realized that law tells you where value ends up, but does not tell you how fast it gets there or why some layers of the stack empty out in months while others take years to show even a crack.

That is the missing mechanism, and it is the entire subject of Part 2 today.

Here is the short version, and then I will earn your trust with evidence.

Some layers of the Intelligence Supply Chain commoditize. They get cheaper, more interchangeable, and less defensible every quarter.

Other layers compound.

They get more valuable, more entrenched, and harder to replicate every quarter they operate.

The single highest leverage skill a leader can build right now is learning to tell these two categories apart before signing the next contract, not after reading the shareholder letter that explains why the AI investment didn't pay off.

Table of Contents:

1. The Cheaper It Gets, the More It Costs Someone

2. Two Datasets, Same Eighteen Months, Opposite Directions

3. The Line That Runs Through Every AI Stack

4. Same Model, Different Layer, Completely Different Outcome

5. The Honest Complications, and Why They Matter More Than the Headline

6. The Instrument: Scoring Your Own Stack

7. The Takeaway, and What Comes Next

Two Datasets, Same Eighteen Months, Opposite Directions

You do not have to take my word for any of this.

Watch two numbers move in the same window of time, and the mechanism explains itself.

The first number is the price of intelligence. Cost per million tokens has fallen roughly 300 times since 2023.

Source: https://benchlm.ai/llm-pricing-trends

Anthropic's own pricing tells the story cleanly. Claude Opus 3 launched in early 2024 at $15 per million input and output tokens.

• Claude Sonnet 4.5 now delivers comparable quality at $3, an 80% drop in 18 months for equivalent capability.

• Google cut Gemini 1.5 Flash's price 78% just three months after launch.

• GPT-4o mini landed at a 92% discount to GPT-4's original price, and GPT-4.1 Nano pushed more than 99 percent below it.

• DeepSeek V3 arrived from outside the usual frontier labs at roughly one-hundredth of GPT-4's launch cost.

It is an industry-wide pattern, repeating at every model generation, and the cycle time between generations is itself compressing, from about 18 months in 2023 down to roughly 12 months now.

The second number is the price of everything underneath the model.

Utility interconnection queues in the largest 2026 data center markets, Northern Virginia, Phoenix, Dallas, now run four to seven years.

A campus that joined the Northern Virginia queue this spring cannot expect utility power before 2030. PJM alone is sitting on more than 2,600 gigawatts of pending interconnection requests.

GPU lead times have stretched to 36 to 52 weeks, and the deeper bottleneck isn't even the chip itself; it is TSMC's CoWoS advanced packaging, the process that bonds high bandwidth memory to the die, which is fully booked through at least mid-2027.

Microsoft, Google, Meta, and Amazon have already placed multi-billion-dollar forward orders that consume most of Nvidia's allocation into 2027, which means any organization that didn't pre-commit compute in 2025 is no longer executing a plan.

It is reacting to scarcity.

Put those two numbers side by side, and you are not looking at two unrelated news stories. You are looking at one mechanism, captured twice, in the same eighteen months.

The model layer is collapsing in price because anyone can build it and everyone eventually will. The infrastructure layer is tightening in supply because almost nobody can build it and very few have tried in time.

Convergence and scarcity are not opposite forces. They are the same force, read from two different ends of the same stack.

The Line That Runs Through Every AI Stack

Some layers of your AI stack are racing toward free as we noted earlier.

These are your Commoditizing Layers.

Other layers do the opposite, and the longer you operate them, the harder they become to replicate, and the more value they accumulate, and that’s the Compounding Layers.

Between them sits a boundary that I call the Commoditization Line.

The single most useful thing you can do this quarter is figure out exactly where that line runs through your own organization’s AI stack because every dollar you spend on the wrong side of it is working against you.

Picture the Physical Intelligence stack from Part 1: power and grid at the bottom, then silicon, then data centers, then foundation models, then orchestration and fine-tuning, then applications, with proprietary data and workflow ownership sitting alongside as the layer that can buck the pull in either direction.

The Commoditization Line moves through that stack, and right now, in 2026, it sits roughly between foundation models and everything below them. Power, silicon, and owned infrastructure compound. Model access, prompt engineering, and thin application wrappers commoditize.

But do not mistake this for a single clean cut.

The research is explicit on this point, and so should you be. Within the model layer itself, open-weight models like Llama, Qwen, and GLM are commoditizing faster than frontier proprietary models, because there are more of them competing on the same benchmarks.

Within infrastructure, generic colocation space commoditizes while owned power generation and GPU allocation locked in before 2025 compound, because the owner captured a scarce position before the queue got to seven years.

The line does not run cleanly between named layers but runs inside them. Any executive who treats "foundation models" or "infrastructure" as internally uniform categories is already misreading their own stack.

There is a historical precedent for exactly this pattern, and it is sharper than the usual electricity and oil comparison. When IBM launched the original PC, it did not build the operating system or the microprocessor in house.

It licensed both, from Microsoft and from Intel. Within a handful of years, industry value migrated almost entirely away from the computer assemblers, IBM included, and toward those two upstream suppliers. Through most of the 1990s, the combined profits of Intel and Microsoft alone exceeded the total profit of the entire world PC assembly industry.

Only Dell made real money as an assembler.

Here is the detail that makes the story worth telling twice: Microsoft and Intel were not obviously the best technology available at the time. Zilog and Motorola built chips that many engineers considered superior.

Digital Research had an operating system widely seen as comparable. IBM did not lose the PC business by picking bad partners. IBM lost it by picking partners at the layer that was structurally destined to compound, the layer that set the standard interface everyone else had to build against, and then letting those partners own that interface completely.

The lesson was about knowing which room the value was going to end up in.

Same Model, Different Layer, Completely Different Outcome

If the framework is right, companies with the same commoditized models should end up in very different places because of what they built beneath or beside them, not because of the model they chose. They in fact do, and the contrast is striking.

Jasper AI reached a valuation of roughly 1.5 billion dollars as an AI writing assistant built on top of early GPT access. Its entire value proposition was a friendlier interface to a model anyone could call.

When ChatGPT launched in November 2022, that interface advantage evaporated overnight, because the model vendor itself shipped the same convenience directly to consumers for free.

Jasper cut its internal valuation, pivoted toward enterprise customers, and still watched revenue decline sharply over the following two years, because there was nothing underneath the interface that a competitor, or the model vendor, could not replicate by Tuesday.

Now look at Harvey, a legal AI platform built on the same category of underlying model access. Harvey trains custom agents on each law firm's own proprietary documents, inside that firm's security and compliance boundary, and firms are now running more than 25,000 such agents per company.

Once a firm has built that depth of custom workflow on Harvey's platform, switching to a competitor means rebuilding all of it from scratch. Harvey reached 300 million dollars in annual recurring revenue in May 2026, up from 195 million just five months earlier, and closed a 200 million dollar growth round at an 11 billion dollar valuation.

Or look at Cursor, the AI coding assistant that passed 2 billion dollars in annualized revenue by February 2026 and was bought by SpaceX last month.

Analysts are explicit that the valuation rests on workflow lock-in and developer trust, not on which underlying model Cursor calls, which is precisely why thin AI wrappers are dying while thick ones, the ones with genuine data or workflow depth, are trading at premium multiples well above ordinary SaaS.

All four companies Jasper, Harvey, Cursor, and the model vendors they all built on top of, had access to comparable commoditizing model capability at the time that mattered most. The difference in outcomes was whether each company built on top of it or failed to.

The Honest Complications, and Why They Matter More Than the Headline

Let me challenge my own assumptions and play a devil’s advocate.

The first complication to this model is the capital allocation story. It would be satisfying to say enterprises are recklessly overspending on the commoditizing application layer while starving the compounding infrastructure layer.

The actual 2025 spending data does not fully support that. Enterprise generative AI spend split nearly evenly, $90 billion to applications and $18 billion to infrastructure, and infrastructure already captures something like 45 percent of AI spending broadly defined.

The sharper, better evidenced claim comes from MIT's Project NANDA, which studied 300 public deployments and interviewed 150 enterprise leaders: 95 percent of enterprise generative AI pilots fail to deliver any measurable profit and loss return, and internal builds succeed at roughly 1/3rd the rate of purchasing from a specialized vendor or partner.

The real problem is not a dollar misallocation between layers. It is that most internal builds land one layer too high, competing on ground that commoditizes out from under them, without anyone first asking which layer the build actually sits on.

Your spend goes in the wrong layers.

The second complication is the data moat question, and here the honest answer is domain conditional rather than universal.

One camp argues that because foundation models train on public and increasingly synthetic data, most proprietary datasets are speed bumps rather than castle walls.

A competing camp points out, also correctly, that in specific regulated or high transaction volume domains, payments processors sitting on transaction data no model vendor can replicate, or legal and healthcare firms operating inside a trust and compliance perimeter, genuinely hold a moat no foundation model release will erase.

The honest synthesis is not "data moats are dead."

There you go, I said it.

It is that data moats are domain conditional, and knowing which domain you are in matters more than the general debate.

The third complication is switching cost. The clean version of the model layer is commoditizing; that implies switching vendors is trivial.

For an organization that has already sunk real cost into fine-tuning, retraining, and deep integration with one vendor's specific model behavior, it is not trivial at all, and no dataset currently quantifies exactly how not trivial.

Treat this as real friction, not a footnote.

Finally, a distinction that complicates the model versus infrastructure binary itself: Cursor's advantage is arguably distribution and developer trust as much as it is proprietary data.

Distribution, the depth of platform integration and adoption reach, may be a third kind of compounding asset that sits at neither the model layer nor the classic infrastructure layer. A framework that only looks for compounding value in data and hardware will miss it.

A framework that survives its own best objections is worth more than one that pretends they do not exist.

The Instrument: Scoring Your Own Stack

Here is where this stops being interesting and starts being usable Monday morning.

Take every AI capability your organization currently funds and score it, layer by layer, against three questions.

1. Vendor multiplicity: is this capability available at comparable quality from more than one credible provider today?

2. Price versus value trend: is the price of this capability falling faster than the value it delivers is growing?

3. Structural replication difficulty: does this capability get genuinely harder for a competitor to copy the longer you operate it, or does it stay exactly as copyable next year as it is today?

Rate each dimension one to five and add them up per layer.

A high vendor multiplicity score and a steep price versus value decline signal a commoditizing layer, worth renting cheaply and never overcommitting to.

A high structural replication difficulty score signals a compounding layer, worth the deliberate long term investment.

Run this scoring exercise as a standing gate on every new AI commitment this quarter, because the commoditization cycle is now running at roughly twelve months and getting shorter with every release.

The posture differs by organization. For enterprise companies, this is a capital allocation and vendor contract question that belongs with the CFO and the board, not just the CIO.

Model your multi-year vendor commitments differently depending on which side of the line they fall: favor short, flexible terms at the commoditizing model layer, and favor longer-term ownership or capacity commitments at the compounding layers, the way the hyperscalers already locked in GPU allocation and power contracts years ahead of need.

Ask your Chief Data and AI Officer a harder question than how many seats are activated. Ask which of your internal builds are being defended as proprietary that actually sit one layer too high, competing on the ground a vendor's next release will commoditize for free.

For mid-market, you will never outcapitalise a hyperscaler on the infrastructure layer, and that is not the fight to pick.

Run a two column audit instead. In one column, what you are buying, model access, generic copilots, off the shelf interfaces.

In the other, what you are building or capturing, proprietary workflow data, customer interaction history, documentation of institutional judgment that nobody else has. Stop overpaying for the first column.

A custom AI platform that is really a commoditizing model wrapper with a markup can do real damage to a mid-market P&L over a multi-year contract.

Redirect what you save into the second column, even in small increments, because a mid market firm that correctly identifies its own compounding layer early can build a defensible position before a larger competitor finishes forming its AI governance committee.

The Takeaway, and What Comes Next

Cheaper intelligence doesn't mean cheaper advantage. It means the advantage moved.

If you don't know where it moved to, you are still paying full price for something that is now free.

Part 1 told you value migrates toward the constraint. Part 2 tells you why the migration happens at different speeds across the stack, and gives you a line to draw through your own organization to find out which side you are standing on.

The line is not fixed as it runs inside layers as much as between them; it moved through the PC industry in the 1990s exactly the way it is moving through the Intelligence Supply Chain now, and it will keep moving as fast as the next model release and the next interconnection queue update allow.

Pick your three largest AI line items. For each one, ask honestly which side of the Commoditization Line it sits on, and whether your organization is treating it accordingly, renting the commoditizing ones cheaply and flexibly, and investing deliberately in the compounding ones before the queue for them gets to seven years.

Next week, in Part 3, we go beneath the Commoditization Line itself and into the Infrastructure Inversion, the mechanics of why physical infrastructure now scales ahead of the intelligence it powers, and what that inversion means for how you plan capital, not just capability.

Talk soon,

Sameer Khan

Intelligence Economy

Hey Transformation Leader,

The question your board is most likely answering today without knowing it.

Why Your AI Strategy Is Backwards?

Look at how your organization buys artificial intelligence today. A per-seat license here. A LLM model subscription there. A chatbot pilot that someone in marketing championed. Each line item sits on the software budget, gets approved like software, and gets forgotten like software.

Now ask a harder question. If intelligence is software, why is Microsoft signing a twenty-year contract to restart a nuclear reactor?

That single fact should stop you. Companies do not commit to twenty-year nuclear power purchase agreements to run AI software. They do it to secure a scarce, physical, capital-intensive input that they cannot get any other way, and the moment you understand why Microsoft did that, you understand the entire thesis of this newsletter and the foundation of the next thirteen weeks of writing.

Intelligence is no longer a feature you buy.

It is a manufactured good produced by a physical supply chain. That chain runs from power generation, through the grid, into silicon, through data centers, up into foundation models, through orchestration, and finally out to the application your team actually touches.

Like every industrial supply chain before it, the durable value does not stay where everyone is looking. It migrates, and it migrates in a direction most boards are budgeting against.

This is Framework#01, the Intelligence Supply Chain. It is the spine of everything that follows. Get it right, and every later decision gets clearer. Get it wrong, and you will spend the next two years funding the layer that is commoditizing fastest while your competitor quietly locks up the layer that compounds.

Table of Contents

1. What Changed in 2026

2. Where Value Is Moving

3. Read the Margins, and You Read the Future

4. The Law, and the Word That Makes It Defensible

5. Commoditizing vs. Compounding Layers

6. The Decision: Own, Partner, or Rent, Layer by Layer

7. The Strategic Audit

8. Key Takeaways

What Actually Changed in 2026

For most of the software and cloud era, the smart money sat close to the customer.

The application captured the value. The cloud infrastructure underneath was plumbing, abundant and cheap, and nobody built a strategy around plumbing. That intuition was correct for thirty years.

It is wrong now, and three forces converging in 2026 are the reason.

The first force is convergence. Foundation model capability is collapsing toward a narrow band.

As of mid 2026, the leading models cluster within a few points of each other on standard evaluations. Open weight models now score above 50 on the Artificial Analysis Intelligence Index, up from the low 30s a year earlier, while holding a 6x to 7x times price advantage.

The gap between the closed frontier and the open frontier has narrowed to something like four to eight months. When everyone can call a comparable model from the same API, the differentiation that the application layer once rented evaporates.

The second force is scarcity moving down the stack. The binding constraint on AI is no longer chips alone. It is power and the grid. Interconnection queues now average around five years. High voltage substation lead times run three to five years.

By early 2026, nearly a third of all planned new data center capacity is being designed to bypass the public grid entirely, because the grid cannot deliver in time. When the input becomes the constraint, value migrates toward the input.

That is not my opinion. It is what every industrial economy has done for a century.

The third force is capital.

Hyperscalers are planning somewhere between 600 and 725 billion dollars in capital expenditure for 2026, a jump of 36 to 62 percent over 2025.

Roughly three quarters of that is AI related. This is the financial signature of heavy industry, and it means the competitive structure of the entire AI era is being set right now, in the contracts and buildouts of this year, in ways that will define position for the rest of the decade.

Put those three together, and the conclusion is unavoidable. Intelligence has entered its industrial phase. The question is no longer which vendor you choose.

The question is where on the supply chain you decide to own, to partner, or to rent.

Read the Margins, and You Read the Future

Follow the margins. Margin and scarcity are the same map.

Start at the bottom.

Power has become so strategic that the largest technology companies on earth are buying nuclear plants.

• Microsoft signed a roughly 16 billion dollar, twenty-year agreement to restart Three Mile Island Unit 1, and a separate 10.5 gigawatt renewable deal with Brookfield.

• Amazon put 700 million dollars into small modular reactor developer X Energy, signed a 1.9 gigawatt agreement with Talen’s Susquehanna plant, and committed to a campus worth more than 20 billion dollars.

• Meta lined up TerraPower, Oklo, and Vistra to reach as much as 6.6 gigawatts by 2035, plus a twenty year nuclear agreement with Constellation.

• Google contracted 500 megawatts of Kairos Power reactors.

• SpaceX went public with the largest IPO ever to launch data centers in space.

Committed nuclear capacity for data centers now exceeds 9.8 gigawatts, more capital committed to nuclear in this single cycle than in any prior decade in United States history.

When the most sophisticated capital allocators in the world race to lock up a layer competitors cannot quickly replicate, they are telling you where the value sits.

Move up one layer to silicon.

Nvidia reported 197.3 billion dollars in data center revenue for fiscal 2026, then 75.2 billion in a single quarter, up 92 percent year over year. Gross margins run around 71 to 75 percent at the company level, with chip-level margins on its flagship parts reported in the mid eighties.

Market share in AI accelerators sits somewhere between 75 and 90 percent, higher still in training. That is what value concentration at a scarce, capital and intellectual property bound layer looks like in numbers.

Now go all the way to the top, to the application layer where the old intuition told you the money would be.

The picture inverts.

Generic AI wrapper businesses are living through what analysts have started calling a gross retention apocalypse. Switching costs sit near zero. AI features are compressing classic SaaS gross margins from the comfortable 80 to 90 percent range down to 50 to 60 percent.

API costs alone consume 15 to 30 percent of revenue for even the successful ones.

Only 3-to 5% of AI wrappers clear ten thousand dollars in monthly recurring revenue, and roughly 90 percent of AI startups are projected to fail by the end of 2026.

Read those three layers in sequence.

The scarcest, most capital bound layers carry the fattest, most durable margins. The most crowded layer, the one closest to the customer, carries the thinnest and fastest eroding margins. You can see where value will sit next year by reading where the physical constraint sits today.

The Law, and the Word That Makes It Defensible

Here is the law of Framework 01 (Intelligent Supply Chain), stated precisely. Value migrates toward the constraint.

Not toward the bottom of the stack in the above diagram but toward the constraint.

Bottom of the stack is useful teaching shorthand, and I will keep using it, because in 2026 the constraint genuinely sits at the bottom, in power and silicon.

But the defensible law is the one about scarcity, and that distinction matters more than it looks.

It matters because the lazy version of this argument, the one every venture blog is now publishing, says infrastructure wins and apps lose, end of story.

That version is both wrong and easy to attack. A sharp CIO will dismantle it in one sentence: our proprietary data and workflow are a moat that sits above the model, so value does not live only at the bottom, and the CIO is right.

So let me give that objection its due, because it is the most important nuance in this entire framework.

There is a real exception to the downward pull, and it has a name.

Some layers are commoditized and some layers compound.

A commoditizing layer is one where your advantage is rented from something converging beneath you, the way an application is rented from a model anyone can call.

A compounding layer is one where your advantage accumulates and gets harder to copy over time, the way proprietary data, hard-won workflow, and systems of action do.

The evidence for compounding layers is as real as the evidence for the downward migration. Vertical AI companies that embed proprietary data and deep workflow are sustaining durable margins above the commodity model.

A retailer whose live point of sale data feeds govern forecasting and pricing agents owns something a competitor cannot buy from the same API.

The investors articulating this most clearly, Menlo and HarbourVest among them, are explicit that the moat is the execution layer data and workflow, not the interface.

Value migrates toward scarcity, including the genuine compounding layers that sit above the commodity. Differentiation rented from a converging model layer is not a moat but a lease.

The moment the foundation provider ships your clever feature natively, your lease expires.

That single reframing from the bottom always wins to value moves toward whatever is scarce, whether that is power at the bottom or proprietary data above the model, is what separates a durable framework from a venture blog post.

Hold onto that for now because part 2 next week is built entirely on the commoditizing versus compounding distinction.

Owning the Bottom Is a Guaranteed Cost, Not a Guaranteed Win

Before we turn the framework into a decision, one more piece of honesty, because the consulting world will not give you this one and you need it.

The migration law tells you where value concentrates. It does not tell you where returns are safest. Those are different claims, and conflating them is how executives talk themselves into building power plants they have no business building.

History is blunt here.

Railroads, telecom, and fiber were all capital heavy infrastructure that concentrated enormous value, and all three were over built and commoditized, with the returns often accruing to later operators rather than the original builders.

The same risk is live in AI right now.

Frontier model providers are pricing inferen ce below cost to hold share. One leading lab reportedly spent around 1.35 dollars for every dollar it earned in 2025.

Projected infrastructure debt issuance runs toward 1.5 trillion dollars. As one Forbes analysis put it, foundation model valuations are approaching a trillion dollars, but history says infrastructure builders rarely win.

So owning the constraint is a guaranteed cost, but the point to be noted is that it is not a guaranteed win.

The sophisticated move is rarely to reflexively buy the scarce layer. It is to secure favorable access to it.

The key message is to get close to the constraint. You rarely need to own it outright.

The Decision: Own, Partner, or Rent, Layer by Layer

The framework earns its keep only when it resolves into a decision your leadership team can actually make, so here is the instrument.

Step 1: Take every AI capability your business depends on.

Step 2: Map each one to its layer in the supply chain.

Step 3: Classify each layer as commoditizing or compounding.

Step 4: Set an explicit posture for each: own, partner, or rent.

The discipline is simple to state and uncomfortable to apply, because it will expose places where you are spending your scarcest engineering talent rebuilding something the bottom of the stack will hand you as a commodity next quarter.

The right posture differs sharply by the size and capital position of your organization, so let me separate the two audiences.

If you lead an enterprise (a Fortune 100 or 500):

This is a capital allocation and vertical integration question, and it belongs on the board agenda, not just the CIO’s desk. Margin and strategic control are migrating toward power, compute, and frontier model access.

Your decision is how far down the stack you must reach to secure supply, cost, and differentiation, and where integration stops adding value.

Reach down to secure scarce inputs through multi-year power and compute agreements on favorable terms, the way the hyperscalers are doing.

Warning: Don’t try to outbuild a hyperscaler. It forces your business to combine two incompatible operating models: fast, asset-light software and slow, capital-intensive infrastructure.

Procurement, real estate, finance, and energy purchasing are now on the critical path of your AI strategy.

No CIO led AI program is structured to handle that, and that organizational collision is a topic I will return to later in the series.

If you lead a midsize organization (between roughly 50 and 500 people):

You will never outcapitalise a hyperscaler, and you should stop trying.

Your win is discipline about where you rent versus where you build. Map your dependencies to their layers. Rent the commoditizing layers cheaply, the models and the raw compute, and refuse to lock yourself into a single vendor at any layer that is converging, because that vendor’s competitor will be cheaper and comparable within months.

Then take the scarce engineering effort you just freed up and concentrate all of it on the one layer where your advantage genuinely compounds, which is almost always your proprietary data and your workflow, not the model.

That is the layer a competitor cannot rent from the same API. That is your moat. Everything else is a lease, and you should pay as little as possible for a lease.

The diagnostic question for both audiences:

For each AI capability we depend on, which layer does it actually compete on, and are we investing where our advantage compounds or where it commoditizes?

Most teams have never asked it.

Most teams will not like the answer.

That discomfort is the point, because the gap it exposes is exactly the overinvestment in the wrong layer that you can correct starting next quarter.

There is a name for the mistake this whole framework is built to catch. Budgeting at the wrong layer. It is invisible on a software profit and loss statement, where a model subscription and a data investment look like the same kind of expense.

It becomes obvious eighteen months later, when your model triples in inference cost, a competitor locks in compute and power you can’t match, or your flagship feature becomes a free capability of the foundation model.

The audit is uncomfortable precisely because it works.

Takeaway and Where We Go Next

If I strip this entire piece down to one sentence, here is what remains.

In the Intelligence Economy, value does not migrate toward the customer. It migrates toward the constraint.

That is the inversion of the software era intuition, and it is the foundation of everything in this series. Intelligence is manufactured. It is produced by a physical supply chain with real scarcity at specific layers.

Value pools where the scarcity is, sometimes at the bottom in power and silicon, sometimes above the model in proprietary data and workflow, never reliably at the application layer where the old playbook told you to look.

The strategic question is not how close you are to the customer. It is how close you are to the constraint.

So here is the reflection to take into your next leadership meeting. Pull up your AI budget. For every dollar on it, ask which layer it funds, and whether that layer is one where your advantage compounds or one where it commoditizes beneath you.

If most of your spend is aimed at the commoditizing layer, you are not wrong to be investing in AI. You are investing in the most fragile part of it.

Next week, in Part 2, we go deeper into the single distinction that decides everything: commoditizing layers versus compounding layers.

We will turn that two word vocabulary into a working test you can run on any capability you depend on, so you can tell, before you commit capital, whether you are building a moat or signing a lease.

The supply chain is the map. The constraint is the destination. Now you know which way the value flows.

Talk soon,

Sameer Khan

Creator of Solve with AI.

As I continue my investigation into the future of AI, I came across another interesting insight. There are engineers at Google, Microsoft, and Amazon right now who are not blocked by a lack of GPUs or capital. They have the software teams, the training runs, the roadmaps.

What they don't have is power.

Not enough megawatts or path to grid connection, and no clear timeline for when any of that changes.

That problem, in theory, belongs in a utility conference rather than a boardroom. This is quietly becoming the most important strategic constraint in the AI industry. Last week, we established that AI is no longer a software story. This week, I want to show you what that actually costs.

Gartner put a number on it last November: 40% of AI data centers will be operationally constrained by power shortages by 2027.

As in, the servers are sitting there and cannot run at full capacity because the electricity simply is not available.

In this week 2 of the AI and the Physical Economy series, we are going deep on energy today. By the time you finish reading this, you are going to look at your own AI roadmap very differently.

Table of Content

1. The Bottleneck Nobody Planned For

2. The Physical Intelligence Stack

3. The Transformer Nobody Ordered

4. China’s Substation Advantage

5. The Nuclear Bet

6. What This Means for Your Business

7. The Substation Moat

The Bottleneck Nobody Planned For

For the last three years, the AI conversation has been almost entirely about models, chips, and software.

• GPT-4, then Claude, then Gemini. H100s, then B200s.

• Context windows, reasoning capabilities, multimodal inputs.

The operating assumption across the industry was clear: whoever ships the best model wins.

Here is what that assumption missed.

To run a model at scale, you need a data center. To run a data center, you need enormous amounts of power, somewhere between 50 and 500 megawatts, depending on size, with hyperscale AI campuses now pushing toward gigawatts.

To get that power, you need to connect to the electricity grid, and connecting to the US electricity grid right now is one of the most painfully slow processes in modern infrastructure development.

The US Department of Energy projects that electricity demand from data centers will increase by up to 130% by 2028. The International Energy Agency projects global data center electricity consumption will double by 2030, reaching nearly 945 terawatt-hours annually.

Yet the infrastructure to deliver that electricity is not keeping pace. Projects are stuck in permitting, and companies are waiting years for transformers. Interconnection applications sit in queues measured not in months but in years.

The AI buildout has run headfirst into the physical world, but the physical world does not care about your deployment timeline.

The Physical Intelligence Stack

I have been mapping what it actually takes to deliver AI at scale. Not just on the software side, but on the physical side.

The inputs that rarely appear in an AI strategy deck but determine whether the strategy survives contact with reality. What I found is a six-layer dependency chain I am calling the Physical Intelligence Stack: Land, Grid, Power, Cooling, Compute, Software. Each layer is a precondition for the one above it. Remove any one of them, and the entire stack fails.

Most of the AI conversation lives in the top two layers, i.e., Compute and Software.

That is where the chips live, where the models run, where the prompts get answered. It is the most visible layer because it is the most digital. But the layers below it are what make everything above it possible.

Pull out any layer below, and the whole stack fails regardless of how sophisticated the layer above it is.

Think of it this way. You can spend $10 billion on the world's most powerful AI cluster. You can hire the best machine learning engineers, train the best models, build the best products on top of them, and then you can wait.

• For a transformer to be manufactured

• For an interconnection application to be processed

• For a substation upgrade that will not happen for five years.

The stack collapses at the bottom, and no amount of brilliance at the top rescues it.

This week, we are focused on layers two and three: Grid and Power. What is happening at those layers right now is the most underreported story in the AI economy.

The companies that understand it early are building advantages that cannot be competed away with a model release. The ones that don't are building roadmaps on assumptions that are already breaking down.

2,300 Gigawatts Stuck in Line

As of mid-2026, approximately 2,300 gigawatts of energy generation and storage capacity are sitting in US interconnection queues.

That is more capacity than the entire installed generating capacity of the United States combined.

Every single one of those projects, wind farms, solar arrays, battery storage systems, and gas generation, is waiting for permission to connect to the grid. The average wait has more than doubled over the past fifteen years.

Projects now spend an average of five years in the queue before reaching commercial operation. In some regions, it is longer.

Data centers are caught in this same queue. In ERCOT, which manages the Texas grid, 143.5 gigawatts of data center capacity were seeking to connect as of late 2025.

For context, ERCOT's highest-ever total demand, the peak of the hottest summer on record, was 85.9 gigawatts. The data centers alone are requesting more grid capacity than the entire state of Texas has ever consumed at any point in its history.

In fact, earlier this month, Texas Gov. Greg Abbott sent a letter to state energy regulators, saying they must take steps to prevent the cost of data center infrastructure from being passed on to residents. He asks that data center owners must bring their own power.

I call this the Power Queue Tax.

Every AI investment made today carries a hidden cost that does not show up in your capital budget: the time cost of waiting for the physical infrastructure to catch up.

For a company trying to deploy AI at scale, that tax can add years to a roadmap. In an industry moving this fast, years are dynasties. The full announced pipeline for US data centers in 2026 is approximately 16 gigawatts. Only around 5 gigawatts are actually under construction. The rest is stuck waiting for power.

The Transformer Nobody Ordered

There is a physical object at the heart of this problem that almost nobody outside the energy industry has heard of. It is called a large power transformer.

It is roughly the size of a small house. It steps high-voltage electricity down to levels usable by a data center. You cannot connect a large facility to the grid without one.

The lead time for a large power transformer in the United States right now is somewhere between 128 and 160 weeks. That is two and a half to three years at minimum.

Some high-capacity units are now quoting four years. Prices have surged 400 to 600 percent from pre-2020 levels, driven by competition between utilities, renewable energy projects, and data center developers all scrambling for equipment from a manufacturing base that was never designed to handle this scale of demand.

Demand for generator step-up transformers, the specific type needed for large-scale power delivery, increased by 274% between 2019 and 2025.

The United States does not manufacture most of these domestically. Chinese companies control approximately 60% of global transformer production capacity. The same strategic supply chain dependency that the US is working to reduce in semiconductors exists in power infrastructure, and it is not yet on the front page.

Despite Big Tech committing over $650 billion in AI capital expenditure for 2026, nearly half of the planned US data center projects may be delayed or cancelled.

Because of a shortage of large metal boxes that have been manufactured the same basic way for decades. This is the part of the AI story that should make every business leader deeply uncomfortable.

China's Substation Advantage

In May 2026, Al Jazeera ran a headline that should be required reading for every executive thinking seriously about AI strategy: "China's secret weapon in AI race with US? Lots of cheap energy."

While the US grid is clogged with interconnection applications and permitting delays, China operates from a fundamentally different starting position. China added over 430 gigawatts of new wind and solar capacity in 2025 alone, accounting for more than half of all new renewable capacity added globally that year.

Its data center rack count grew at 30% annually from 2016 to 2023. China has approximately 400 gigawatts of spare grid capacity that can be directed toward AI computing infrastructure on a timeline measured in months, not years.

An Oxford Energy Institute paper from February 2026 spelled out the advantage clearly: China's data center edge is not primarily about models or chips. It is about energy.

Chinese operators run on a reliable baseload grid with sufficient capacity. National policy mandates that new AI computing hubs be located inside pre-designated zones with ready power and connectivity.

The US, by contrast, asks developers to navigate 50 different state regulatory regimes, local zoning boards, and an interconnection queue that moves at the speed of federal bureaucracy.

At least 36 data center projects were blocked or stalled in the US in the 12 months through mid-2025. $64 billion worth of projects have been delayed or cancelled due to community opposition.

71% of Americans now oppose the construction of an AI data center in their local area. That figure is higher than the historical opposition to nuclear plants. The physical foundation of US AI leadership is being contested in town planning meetings, while China constructs at a pace that makes US permitting timelines look like geological time.

The Nuclear Bet

The hyperscalers saw this coming, and they responded the only way a company with essentially unlimited capital can respond: by going around the grid entirely.

Microsoft signed an 835-megawatt deal to bring Three Mile Island Unit 1 back online.

It is a 20-year power purchase agreement worth approximately $16 billion, locking in dedicated nuclear output for AI computing through 2048.

Amazon expanded its nuclear agreement with Talen Energy to 1,920 megawatts through 2042 and is building a $20 billion AI campus adjacent to the Susquehanna power station in Pennsylvania.

Meta went further still, signing agreements with Vistra, Oklo, and TerraPower to secure up to 6.6 gigawatts of power by 2035. As of mid-2026, every major hyperscaler has signed at least one nuclear deal for AI data center operations.

Let that sync in for a moment.

These are software companies. Two years ago, the most pressing infrastructure conversation in Silicon Valley was about GPU allocation and cloud billing rates. Today, it is about reactor restart timelines and multi-decade power purchase agreements.

This is what Infrastructure Sovereignty looks like in practice. When the grid cannot deliver what you need, and the queue is measured in years, companies with enough capital choose to own their energy supply chain directly.

They are not waiting for the permitting process. They are funding the construction of power generation assets and contracting the output for twenty years.

The result is a structural cost and reliability advantage over any competitor who depends on commercial grid power or cloud resale capacity.

The challenge is obvious: not every company can sign a $16 billion nuclear power purchase agreement. In fact, most cannot.

Which means the hyperscalers are not just securing compute capacity for themselves. They are determining the terms on which everyone else will access compute in the years ahead.

What This Means for Your Business

Here is the question you should be asking yourself right now.

If your AI strategy depends on cloud compute, cloud compute depends on data centers, and data centers depend on constrained power infrastructure, what does your AI roadmap actually look like?

Most business leaders I speak with have not connected those dots. Their AI strategy lives entirely in the top two layers of the Physical Intelligence Stack: software and compute.

They are thinking about which models to use, which vendors to evaluate, and which use cases to pilot. They have not thought about the layer five levels below that, where the real constraints are accumulating.

This does not mean you need to buy a nuclear power plant.

But it does mean several things worth acting on right now. Cloud pricing for compute-intensive AI workloads is going to increase, not decrease, as power constraints bite into data center expansion plans. If your AI economics depend on today's cloud pricing holding flat, your business case is built on a fragile assumption.

Build in a 30 to 50 percent cost escalation over the next three years and stress-test your unit economics under that scenario.

Geographic positioning matters more than it has in a decade. Data center capacity is not uniformly distributed. Regions with available power, progressive permitting, and grid capacity will see AI compute costs stay lower and availability stay higher.

Your vendor relationships and infrastructure commitments should reflect that reality. Ask your cloud vendors directly:

1. Where are your next major capacity expansions?

2. What is the power source?

3. What is the go-live date?

The answers will tell you a great deal about whether the compute you are counting on will actually be there when your roadmap demands it.

The Substation Moat

The AI era that everyone assumed would be determined by model quality is being quietly reshaped by something far more fundamental.

The companies that control physical power supply, through owned generation, long-term nuclear power purchase agreements, or positioning in power-rich geographies, are building what I call the Substation Moat.

This moat cannot be crossed by shipping a better model. It cannot be crossed by hiring better engineers or spending more on research and development. It is, by definition, a physical constraint.

The companies that understood this three years ago, the ones signing 20-year nuclear deals and building AI campuses next to power stations, are going to have an operating cost and capacity advantage for the better part of the next decade.

The IEA projects global data center electricity consumption will double by 2030. Demand from AI-optimized infrastructure is expected to quadruple over the same period. The electricity to power that demand does not yet exist in the infrastructure pipeline.

Something has to give.

What gives is the pace of AI deployment for organizations that have not secured their place in the stack.

For example, if two companies are equally brilliant and equally funded. Both are trying to deploy AI at the same scale. One signed a nuclear power purchase agreement in 2024. The other is waiting on a transformer that will not arrive until 2028. The model quality difference between those two companies is irrelevant. The infrastructure difference is everything.

Last week, I told you the AI story is becoming a physical story. This week, I showed you what that costs. Next week, we go one layer up in the Physical Intelligence Stack: cooling and compute density, the engineering problem that makes GPUs the most thermally demanding objects in the history of commercial computing.

The race is on. But it runs through substations, not server farms and the gap is wider than most people realize.

Talk soon,

Sameer Khan

Creator of Intelligene Economy

You’ve been reading Solve with AI for a while now. I appreciate that more than I say, and your engagement keeps me motivated to write and produce useful content.

But I’ve been sitting with something. The name never quite fit what this newsletter actually does. “Solve with AI” sounds like a productivity blog. What I actually write about is bigger than how AI is rewriting the digital and physical world, who controls the infrastructure layer, and what that means for the leaders making decisions inside organizations right now.

So the name is changing. This newsletter is now called Intelligence Economy. Soon it will have its own website.

Note that I am not changing anything (in fact, elevating the experience), and you will continue to experience me as the writer with the same Saturday weekly distribution. I will continue to focus on what is happening inside the companies and industries where AI meets the real world and increasing focus on the actual economics of it.

If you’re a paid subscriber, nothing changes for you. You’re already in.

If you’re on the free list and this resonates, now is a good time to upgrade. The paid tier goes deeper into strategy, frameworks, and what I see inside enterprise AI programs that don’t make it into the public posts.

Get 30 day free trial

See you Saturday.

Talk soon,

Sameer Khan

Every conversation about AI for the past 3.5 years has happened at the software layer.

Models. Code. Interfaces. APIs. Compute efficiency. Token costs.

The assumption stated or not was that AI is fundamentally a digital phenomenon. Something that lives in the cloud and is weightless.

That assumption is breaking.

The next decade of AI will be won by whoever controls the physical supply chain of intelligence: the power, the land, the grid connections, and the permits that make any of the software possible in the first place.

We are not in a software race anymore. We are in an infrastructure race, and most of the people still debating AI strategy have not looked up from their laptops long enough to notice.

Table of Contents

1. The Weight of Intelligence

2. What the Numbers Actually Say

3. The Grid Cannot Keep Up

4. Water, Land, and Political Risk

5. Where the Smart Money Is Actually Going

6. Why This Changes Your Strategy

7. What I Have Seen From Both Sides

8. The Stakes Have Never Been Higher

The Weight of Intelligence

There is a version of the AI story that feels almost frictionless. You open your favorite AI app or SaaS with embedded LLM, type a prompt, and get an answer. The intelligence appears instantly, as if conjured from nothing. That experience is so seamless, so immediate, that it is easy to forget what actually powers it.

Behind every query, every generated document, every AI-assisted decision sits a physical machine drawing electricity, consuming water to stay cool, occupying land inside a building that took years to permit and build.

None of that is abstract or weightless. In fact as the demand for AI capability grows, the physical infrastructure required to support it grows with it at a pace that the world’s power grids, water systems, and regulatory timelines were simply not designed to absorb.

I am not making this up as this is a know fact but unfortunately people only focussed on the visible part of AI.

Think about what it takes to run a large language model at scale. I am not talking about training the model because that is a separate, enormous conversation but just to run it, to serve it to users around the world in real time.

• You need data centers.

• Those data centers need massive, uninterrupted power.

• Need cooling, which means water.

• Need land, which means permits.

• Need grid connections, which means transmission infrastructure.

All of that needs to exist before a single model can answer a single question.

The frictionless experience that users see is built on top of an extraordinarily heavy physical foundation. Once you understand that, the entire AI landscape looks different.

The questions that matter most shift from being purely about model quality and start being about megawatts and permits and water rights and transmission capacity. The conversation today is happening inside a box, and nobody is describing the box.

What the Numbers Actually Say

The International Energy Agency estimated that AI data centers consumed roughly 460 terawatt-hours of electricity globally in 2022.

For context, that is more electricity than many entire countries use in a year. The IEA projects that figure will more than double, reaching 1,000 terawatt-hours by 2026.

Two years from now. The demand curve is not gradual. It is nearly vertical and not sustainable unless something is done about it.

Goldman Sachs has projected $7.6 trillion in AI infrastructure investment over the next decade. That is not investment in models or software or API access where the Wallstreet usually focusses.

That is investment in the physical substrate that makes AI possible at all such as land, buildings, power systems, cooling infrastructure, grid connections, and the energy generation capacity to feed it. 7.6 trillion dollars to keep the lights on for intelligence.

Consider where the capital is actually flowing right now.

In 2024 alone,

• Microsoft announced $80 billion in capital expenditure.

• Google committed $75 billion.

• Amazon allocated $104 billion.

• Meta put forward $65 billion.

Those four companies together are committing something in the range of $725 billion over multiple years, and a substantial portion of that money is flowing directly into physical infrastructure.

Not software development or research teams. Power. Land. Buildings. Grid capacity.

These are the largest capital expenditures in the history of technology. They make the industrial buildouts of the 20th century look measured by comparison. These investments are happening because the people running these companies understand something the broader business conversation has not absorbed: intelligence at scale is a physical problem, and the physical problem is harder than anyone publicly admits.

The Grid Cannot Keep Up

The electrical grid particularly in the United States was not built for this.

It was built for a different era of consumption, with different demand patterns and a different pace of change. Data centers are uniquely demanding in ways that strain grid operators: they want power at high volumes, with minimal interruption and 24/7.

That profile puts enormous pressure on systems that were designed around very different assumptions about who needs power, when they need it, and how much.

Before 2020, the lead time for large electrical transformers the equipment that steps voltage up or down to move power through the grid was around 50 weeks.

By 2024, those lead times had grown to over 160 weeks. That is more than three years just to procure one piece of critical equipment. You cannot accelerate AI deployment faster than you can procure transformers, and right now, you cannot procure transformers on any timeline that matches the speed at which AI companies want to scale.

The consequences are direct and expensive. According to Data Center Watch, by 2024, $64 billion worth of data center projects had been blocked or delayed specifically by grid constraints.

Not by zoning opposition or community resistance.

By the inability to get power to the site. Sixty-four billion dollars in committed capital, stalled because the grid could not keep up. That number will grow.

The permitting and grid interconnection timelines compound everything else. In many US markets, the full process of securing grid interconnection, completing environmental review, navigating local permitting, and getting a large facility online takes five to seven years.

Five to seven years!

Meanwhile, AI model capability is advancing on roughly an annual cycle. The mismatch between those two timelines is not a minor operational detail. It is a structural constraint on how fast AI can actually scale in the real world, and it is one that no amount of model efficiency can close.

Water, Land, and Political Risk

Power is the most visible constraint. It is not the only one.

AI data centers consume water at a scale that is beginning to generate real political friction. Cooling these facilities requires enormous volumes of water i.e. millions of gallons per day at large sites.

As hyperscalers look for land to build on, they have increasingly moved into regions with favorable land costs and permitting environments. Some of those regions also face drought conditions or water scarcity pressures.

The collision of data center demand with municipal water supply is already producing conflicts in multiple US markets, and those conflicts are intensifying.

Mordor Intelligence tracks data center water consumption as a market growing at 12% CAGR faster than almost any other infrastructure segment.

That growth lands somewhere. It lands in communities that were not planning for it.

Cities and counties that once competed aggressively for data center investment are now asking harder questions.

What do we actually get in return?

Data centers employ very few people relative to their footprint and resource consumption. They do not anchor supply chains or generate downstream economic activity at the scale that traditional industrial development does.

The community calculus is shifting, and permits that might have moved through quickly five years ago are now facing longer review cycles, more organized opposition, and more demanding mitigation requirements.

As the most obvious and well-serviced sites in major markets get taken, development is moving into secondary markets, rural regions, and increasingly complex jurisdictional environments.

Every step away from established infrastructure adds regulatory complexity, construction cost, and timeline risk.

The easy sites are gone.

What is left requires more capital, patience, and local political navigation than anything the industry has had to manage before.

The physical economy of AI is not just complicated, but it is getting more complicated every quarter than you and I can imagine.

The companies that treat this as a background problem, as something their cloud vendor handles are accumulating a kind of infrastructure risk that does not show up in any AI adoption dashboard.

Where the Smart Money Is Actually Going

Watch where the hyperscalers put their money when nobody is framing it as an AI story.

Microsoft signed a 20-year power purchase agreement with Constellation Energy for output from the restarted Three Mile Island nuclear reactor.

That is an energy policy, not an IT decision made at a horizon that most technology companies do not even think about.

It reflects a view that the power question for AI is so large, so long-dated, and so structurally difficult that it requires energy sources capable of running at high capacity for decades without interruption.

Nuclear fits that description in a way that solar and wind, without massive storage infrastructure, simply cannot.

The bet Microsoft made is not primarily about carbon targets. It is about reliability, density, and the sheer volume of power required to serve AI at scale without depending on grid conditions that change by season and by year.

If you see the world’s largest technology company making a 20-year commitment to nuclear energy, it should tell you something. They have looked at the power landscape for the next two decades and concluded that conventional grid access alone is not enough.

If you look at what the other hyperscalers are doing, you’ll see the same pattern. They are acquiring land to secure transmission access, building private power infrastructure, investing in energy storage, and pushing for regulatory changes that speed up grid connections.

These are the actions of companies that have internalized the physical reality of what they are building.

They are behaving like industrial companies because, at their scale, that is what they have become.

Why This Changes Your Strategy

If you are a business leader thinking about AI strategy right now, this physical reality has direct implications for how you should approach your own position, even if you will never build a data center yourself.

1. The first implication is about vendors.

The AI providers you depend on, including your model providers, your cloud infrastructure partners are competing in this physical race right now.

The ones who win it will be able to deliver on their product roadmaps and their pricing commitments.

The ones who fall behind will face constraints that slow their development, raise their unit costs, and ultimately limit what they can offer customers. Infrastructure strength is becoming a legitimate axis of vendor evaluation. It was not two years ago.

2. The second implication is concentration risk.

As the physical supply chain of AI consolidates around a handful of hyperscalers with the capital to build at this scale, the dependency risk for everyone else increases.

If your AI strategy runs through one or two cloud providers, and those providers face infrastructure constraints, that constraint cascades to you.

Resilience in AI increasingly means thinking about physical infrastructure diversity, not just model or vendor diversity.

3. The third implication is timing.

The 5-7 year timelines for large data center development mean that the capacity being built today was planned years ago.

The capacity decisions being made right now will determine what is available in 2030 and beyond.

Your window to influence those decisions, through vendor partnerships, policy engagement, or early enterprise commitments that help shape where hyperscalers invest, is much smaller than most business leaders realize.

Infrastructure is not something you react to after the fact. By the time the constraint is obvious, it is already too late to address it.

What I Have Seen From Both Sides

I have stood at both ends of this chain, and that experience is the reason this series exists.

I spent years working in physical industrial operations, where power demand is measured in megawatts, equipment lead times determine whether projects stay on schedule, and constraints are things you can actually see and touch, not just numbers on a dashboard.

That world teaches you a particular discipline. You learn to read the physical limits of a system not as inconveniences to be optimized around but as the actual structure of the problem. You learn that software solutions stop at the edge of physical reality, and that edge is harder and more fixed than people who have only worked in software usually believe.

Then I led an enterprise AI transformation inside one of the most physically demanding industries on earth.

I watched the collision happen in real time: the speed at which AI can theoretically move, hitting the reality of operational timelines, regulatory environments, and infrastructure constraints that do not compress regardless of how sophisticated the software gets.

The gap between what AI can do in a demo and what the physical world allows it to do in deployment is the most underappreciated constraint in enterprise technology today. It shows up everywhere.

It almost never gets talked about honestly.

That dual vantage point is what this series is built on.

The conversation most AI strategists are having about models, about workflows, about use cases and ROI is real and worth having. But it is happening inside a box.

This series is about the box.

The Stakes Have Never Been Higher

The AI conversation of the last decade was about software.

Who had the best models, the fastest APIs, the cleverest applications.

That conversation produced enormous value. It will keep producing value. But it is no longer sufficient as a strategic frame, because it describes only the top layer of a system whose real constraints are buried much deeper in the ground, in the grid, in the permit queue, in the transformer backlog.

The AI conversation of the next decade is about infrastructure. About who controls the physical supply chain of intelligence. About who has secured the power, the land, the water, and the grid access to run AI at the scale that actually matters. About who built when the window was open, and who waited until the window closed.

The companies, leaders, and governments that understand this moment will shape what AI looks like for the next generation.

They will set the terms on pricing, on access, on capability, on resilience.

The ones that do not will spend that generation as consumers of infrastructure built by others, on terms they did not negotiate, at prices they did not influence, with resilience determined by someone else’s planning horizon.

That is the physical economy of AI and that is what this series is about.

This is Week 1 of “AI and the Physical Economy” a 13-week series on how the physical world shapes, constrains, and ultimately determines the trajectory of artificial intelligence.

Six distinct lenses are coming across the next several weeks: from energy and grid dynamics to water and land, from supply chains to geopolitics, from enterprise resilience to the long arc of what all of this means for the people and companies trying to win with AI.

Talk soon,

Sameer Khan

Creator of Solve with AI.

If this was useful, you’ll find more like it every week in Intelligence Economy — AI strategy at the infrastructure layer, for leaders making decisions in the physical AI era. Subscribe free at solvewithai.substack.com

I did a live video about the major constraint emerging beneath the AI boom, and it’s already influencing billion-dollar decisions across the technology industry. The companies that see it early will make very different infrastructure, vendor, and investment decisions than those that don’t.

In this live episode, I unpack a hidden shift that is forcing some of the world’s largest technology companies to rethink their growth plans, rewrite long-term strategies, and pursue solutions that would have seemed unthinkable just a few years ago.

We’ll explore:

• Why some of the biggest assumptions behind enterprise AI roadmaps may no longer be safe

• The overlooked infrastructure challenge quietly reshaping the AI landscape

• The surprising reason technology companies are making moves far outside their traditional business models

• How global competition could be influenced by a factor most AI discussions barely mention

• The strategic questions every CEO, COO, CIO, and transformation leader should be asking right now

If your organization is investing in AI, cloud, automation, or digital transformation, this conversation may change how you think about the next five years.

Talk soon,

Sameer Khan

Creator of Solve with AI.

I came across this statistics and wanted to share with you because it is going to bring the point home of this post. The NANDA initiative studied 300 enterprise deployments, interviewed 150 leaders, and surveyed 350 employees, then published a findings.

95% of enterprise generative AI pilots deliver no measurable impact on the P&L. Not a small return. Zero. The pilots run, the demos dazzle, the slides look great in the steering committee, and the income statement does not move a single basis point.

So let me ask you the question your board is going to ask you. You have spent eighteen months and a real budget on AI.

Where is it on the P&L?

If your honest answer is somewhere between soon and it is complicated, you are not behind on technology. You are stuck inside a mental model that guarantees you stay in the ninety-five. This post is about why, and about exactly what the five percent did differently.

Table of Contents

1. The Productivity Trap

2. The Coordination Tax

3. The Bolt-On Ceiling

4. The Decision Layer

5. The Autonomy Ladder

6. The Agent-Washing Trap

7. The Closing Window

Read more

Over the past year, working across multiple AI engagements, I have audited dozens of opportunities.

Some of them sounded exciting. Two had already quietly burned through the budget before anyone asked what they were for.

So I did what I always do. I ran every single one through the same process, and by the end, something uncomfortable became obvious.

Only three were worth building right now.

The rest got staged, parked, or killed.

The problem was never the technology, the models, or the tools. The problem lived in the business case underneath each idea, and almost none of them had one that held up.

I realize this is not my problem but a general one you may be facing as well: at any given time, we float twenty ideas and have no honest way to choose among them.

That is exactly what this framework fixes. It is called BUILD-$100M, and every idea has to survive two gates before it earns a dollar. First, prove it matters, and then, prove you can actually deliver it.

Let me show you how it works, and what happened when I pointed it at all twenty.

TABLE OF CONTENTS

1. Why Most AI Projects Fail Before They Ship

2. Gate 1: The $100M Filter

3. Gate 2: The BUILD Score

4. What Happened Across All 20

5. The Three Winners

6. The AI Theater Trap

7. What I Would Do If I Were Starting Today

Why most AI projects die before they ship

Walk into almost any company right now, and I will challenge that you will trip over AI ideas.

• Chatbots (many)

• Agents pile

• Copilot experiments that three people are tinkering with on the side.

• Internal assistant nobody has opened in a month.

Everyone has ideas, but almost nobody has measurable value.

The data backs this up in a way that should make every leader stop and breathe. MIT’s 2025 GenAI Divide study examined 300 deployments and found that 95 percent of enterprise generative AI pilots delivered no measurable impact on the bottom line. Only 5 percent created real value.

We are in 2026, and that’s still a big gap.

The researchers were clear that the failures were not caused by weak technology. They were caused by either one or a combination of the following:

• Weak execution

• Nobody owned the workflow

• Nobody planned for adoption

• The data was not ready

• The value was never defined in the first place

• They even found that most budgets were aimed at flashy sales and marketing tools, while the biggest returns were quietly sitting in unglamorous back-office automation.

Gartner saw the same wave coming and put a date on it. They predict more than 40 percent of agentic AI projects will be scrapped by the end of 2027, mostly because of unclear business value and runaway cost.

So the failure does not happen in production. It happens far earlier, before anyone writes a single prompt. It happens at the moment of approval, when a good sounding idea gets a green light it never earned.

Which is exactly why every opportunity has to clear two gates before it goes anywhere.

Gate 1: The $100M Filter

The first gate answers one ruthless question. Does this even matter?

I wrote an entire post on this topic.

Does it move something the business actually cares about? The $100M Filter forces every idea to justify why it deserves to exist, and it measures that across five kinds of value.

Does the idea make money by lifting revenue through something like proposal generation, lead qualification, or sales intelligence?

1. Does it save money by stripping cost out of reporting, invoicing, or the manual admin work that quietly eats payroll?

2. Does it reduce risk by tightening compliance and quality in areas such as contract review and audit prep?

3. Does it increase speed by collapsing a cycle time the customer actually feels, like approvals, onboarding, or quote turnaround?

4. Or does it improve decisions by making forecasting and market intelligence genuinely sharper?

Notice that this is not just my opinion. It is almost word for word how Gartner now tells companies to choose. Their analyst said it plainly: real value comes from applying AI where it drives measurable improvements in cost, speed, quality, or scalability.

The rule at this gate is brutal on purpose. If an idea does not move at least one of those five in a way you can measure, it does not pass. You kill it before it costs you anything more than the meeting you are sitting in.

Gate 2: The BUILD Score

Surviving the first gate only proves an idea is worth wanting.

It does not prove you can pull it off.

This is where most ideas quietly fall apart. The value is real, but the path to capturing it runs straight through messy data, a nervous team, and a workflow nobody actually owns. So the second gate stops asking “should we” and starts asking “can we, realistically.”

The BUILD Score rates each surviving idea across five dimensions, with one to five points each.

• Business impact asks how big the outcome really is.

• Usage frequency asks how often the workflow actually happens, because a brilliant tool used twice a year is a hobby, not an investment.

• Implementation complexity asks how hard this is to ship, and here is the twist: you score it in reverse, so easy earns a five and painful earns a one.

• Likelihood of adoption asks whether your people will genuinely use it once the novelty wears off.

• Data readiness asks whether the information the system needs is already sitting there, clean and available, today.

Add the five together, and you get a score out of 25. Anything from 21 to 25 is a build now. Sixteen to 20 earns a pilot, a small controlled test before you commit real money. Fifteen or below gets parked, no matter how much you happen to like it.

Does it matter, and can we deliver? That is the entire engine.

What happened across all 20

So I lined up all twenty ideas and ran them through both gates, in order, no exceptions.

Here is the honest scorecard.

Now look at what that table is really telling you.

The interesting result was not which ideas scored high. Three of them did, and I will get to those in a second. The interesting result was how many scored low the moment I stopped grading them on vibes and started grading them on value and feasibility.

Generic chatbots. An “AI strategy” assistant. An HR policy bot. A social content generator. On a slide, every one of them looks like progress. Run them through the gates, and they fall apart, because they either fail to move a number anyone cares about, or they have no real shot at adoption once the demo is over.

That is the whole point of a framework. It is allowed to disagree with the room.

The three winners

Three ideas cleared both gates with room to spare. They are worth studying closely because they share a pattern you can copy onto your own list.

The first was proposal automation, and it scored the highest of everything on the list. It is easy to see why, because it hits three kinds of value at once.

It makes money by lifting win rates, it saves money by killing hours of manual drafting, and it increases speed by shrinking turnaround from days to minutes.

The workflow happens constantly, the risk is low, and a human still reviews the final document before it ever reaches a client. You measure it with proposal turnaround time, win rate, and cycle time, and the needle moves fast enough that people believe in it.

The second was a customer onboarding agent. This one wins because it sits directly on top of revenue you have already earned but not yet realized. Every day a new customer waits is a day of value leaking out and a quiet chance for them to wonder if they made a mistake choosing you.

The agent removes the coordination delays, chases the missing information, and keeps the handoffs from falling through the cracks. Track onboarding time, the number of missing information loops, and escalations, and you will watch it earn its keep.

The third one surprised a few people in the room. A project follow-up agent. On the surface, it looks like boring admin. Underneath, it is solving a coordination problem that quietly bleeds every company alive.

Think about it for a second.

• How many decisions get made in a meeting and then simply evaporate?

• How many follow-ups get forgotten?

• How much of your week disappears into chasing status updates that should have updated themselves?

Measure overdue tasks, the percentage of meeting actions actually completed, how long a decision takes to turn into movement, and how the case makes itself.

Notice the common thread. None of these are flashy. All three are frequent, measurable, owned by someone, and built on data the company already has. That is what a winner actually looks like, and it rarely looks like the thing that gets applause in the demo.

The AI theater trap

Now let me warn you about the most expensive mistake in this entire space.

A lot of AI projects get funded not because they create value, but because they look intelligent.

The slick assistant was built mostly to impress the board. I call this AI theater, and it is seductive precisely because it photographs beautifully. It just does not do anything.

These projects almost always score badly at the second gate. Adoption is low because nobody actually needs them. Ownership is fuzzy because they were never tied to a real workflow. Measurement is impossible because there was never a number they were supposed to move in the first place.

Here is the line I want you to carry out of this article.

The most dangerous AI projects are not the ones that fail loudly. They are the ones that look successful while creating no measurable value at all. They survive review after review, soak up budget and attention, and quietly become the reason your real opportunities never get funded.

What I would do if I were starting today

So picture this. You have Microsoft Copilot, Claude, ChatGPT, and Copilot Studio all sitting in front of you.

What would I do? Honestly, it is almost embarrassingly simple.

I would start by writing down twenty opportunities. No filtering, no judging, just everything currently on the table. Then I would run all twenty through the $100M Filter and accept that a good chunk of them die right there.

The survivors earn a BUILD Score, one to five across all five dimensions. From whatever is left standing, I pick the top three. That’s It.

And then comes the part almost nobody has the stomach for. I ignore everything else for ninety days.

That last step is the entire secret.

Remember the MIT finding that focused, well-integrated solutions succeeded far more often than effort spread thin across a dozen half-built experiments?

Focus is not the boring choice here. It is the winning one.

The real bottleneck

Most leaders think their problem is AI adoption.

Their problem is AI prioritization.

The companies that win in the next three years will not be the ones running the most pilots or making the most noise on LinkedIn. They will be the ones who found the handful of workflows where AI creates real, measurable value, and then executed on those few with everything they had.

Every idea you consider deserves the same two questions.

Does it matter?

Can you deliver it?

That is the BUILD-$100M framework in a single breath.

I pointed it at twenty opportunities. It handed me three.

Your list is probably no different. The only real question is whether you have the discipline to say “not yet” to the other seventeen.

Talk soon,

Sameer Khan

Creator of Solve with AI.

Hey Productivity Explorer,

I was catching up with a friend of mine who is a sales executive in a mid-size organization on Labor Day (Monday of this week). I will not disclose the name of the organization for privacy reasons.

When we started talking about work, he mentioned he recently lost a seven figure deal over a sales quote. Not because the price was wrong, but a competitor was better.

It sparked my interest, so I started asking him more questions to dig deeper, so I could understand the root cause.

I found that it took him more than three weeks to send the quote to the customer because of the deal size.

Why?

The quote had to go through various departments in his company. Pricing, legal, finance, supply team, then legal again because something changed.

Five smart people, each doing their job competently. By the time the quote reached the customer, the customer had already signed with someone faster, and my friend lost a big commission check.

Nobody in that chain made a mistake. The deal did not die because there was incompetence, but the space between competent people.

I have been thinking about that deal a lot since Anthropic released Claude Opus 4.8 on May 28. At the surface, it looks like better benchmarks, stronger coding, and another step in a long line of steps. I saw something else entirely.

What I saw is the first real glimpse of a tool built to fix the exact problem that killed my friend’s deal.

Let me explain.

Table of Contents

1. The AI Industry Is Solving the Wrong Problem

2. Every Company Pays a Unique Tax

3. What Anthropic Actually Announced

4. The Eras of Enterprise Productivity (Framework)

5. Why Coordination Beats Individual Productivity

6. Your Next 90 Days (Plan of Action)

7. The Real Story Behind Opus 4.8

Read more

Thank you for being a premium member of Solve with AI.

Since you joined, you’ve had access to frameworks, deep dives, tools, and strategies that most people in this space char…

Read more

If you run a business, big or small, you know that feeling at 11 pm when you’re still at your desk, and the work just keeps coming. The invoice that needs chasing, the campaign that needs building, the books that need closing.

You probably already have some AI in the mix of ChatGPT, Claude, or Gemini, and you’ve used them to write an email, brainstorm an idea, or clean up a report.

But having AI available and having AI actually working inside your business are two completely different things. You’re still the one typing every query, copying the answer into the right place, and switching between tools to get one thing done.

On average, SMBs run 72 different software applications.

That’s pretty significant

A lot of these SaaS come with their own AI, which adds to the complexity.

The manual, repetitive work that eats your day didn’t go anywhere. You just added one more window to your browser.

The tools you use every day, your books, your CRM, your invoicing, your campaigns, they aren’t connected to anything. So even when AI gives you a great answer, you’re still the one who has to do something with it because these tools are disconnected.

That is about to change.

Anthropic launched Claude for Small Business this week, and the difference between this and everything you’ve tried before is exactly that connection.

It lives inside QuickBooks, HubSpot, PayPal, Canva, and DocuSign. It does the work within the tools and waits for your approval before anything moves.

Table of Contents

1. This Is Not Another AI Tool

2. How Does It Impact Your Sales and Marketing Team

3. Finance and Operations Impact

4. How To Get It Running In Five Minutes

5. Your Data Stays Yours

This Is Not Another AI Tool

Claude for Small Business is not a new platform to learn.

I would hate to learn a brand-new platform and spend hours trying to figure things out.

There is no onboarding call, no implementation consultant, or six-week rollout. It is a single toggle inside Claude Cowork that connects directly to the tools you are already running every day.

• QuickBooks

• PayPal

• HubSpot

• Canva

• DocuSign

• Google Workspace

• Microsoft 365.

The software your business already depends on. Claude plugs into all of it.

Fifteen ready-to-run workflows across finance, operations, sales, marketing, HR, and customer service. So you don’t have to tweak and figure out the perfect prompt.

Actual workflows that execute inside your existing tools and wait for your approval before anything moves.

What makes this different from every other AI tool you have tried is exactly that connection. Think about what your day actually looks like.

Think about where your business actually lives right now.

You've got your books in QuickBooks, pipeline in HubSpot, invoices running through PayPal, contracts waiting in DocuSign, and content being built in Canva.

Having Claude directly accessible within these platforms does two things:

1. You don’t need another AI that ships with your SaaS.

2. You can get consistently and personalized AI experience across your apps.

The COO of Purity Coffee said something that stuck with me after reading this announcement. He said Claude did not just solve his problems. It showed him problems he did not know he had. That is the difference between a chat window and AI that is actually inside your business.

How Does It Impact Your Sales and Marketing Team?

Let’s talk about what this actually looks like for the people running your marketing and sales.

For Marketing, Claude pulls your campaign performance data straight from HubSpot, finds the slow stretch in your revenue, drafts a promo strategy around it, and then builds the actual creative assets inside Canva, ready to publish.

That is a full campaign cycle that used to require a team, a briefing meeting, three tool switches, and a week of back and forth. Now it is one connected workflow that runs while you focus on the decisions that actually need you.

For sales, your lead triage is no longer a manual exercise where someone is eyeballing a spreadsheet and making judgment calls about who to follow up with first.

Claude works inside HubSpot, surfaces the right leads, summarises where each one is in the pipeline, and tells you what needs attention today. HubSpot’s own product team said it best when they described it as getting tailored answers, summaries, and pipeline visualisations from your customer data without ever switching tools.

The friction between your data and your next decision is gone.

Finally, for the business owner sitting above all of it, the CEO of Simple Modern put it in a way that is hard to argue with. He said what they used to think were constraints simply are not constraints anymore. Hours of looking at things that did not matter are gone.

That is what it feels like when AI is finally working inside your business instead of sitting next to it.

Finance and Operations Impact

Now let’s talk about the part of running a business that nobody sees, but everybody feels.

The payroll needs and the invoice that has been sitting unpaid for three weeks, and you keep meaning to chase it. This is the work that doesn’t show up on your calendar but somehow fills every gap in your day.

Claude for Small Business has workflows built specifically for this. It pulls your cash position from QuickBooks, checks it against your incoming PayPal settlements, builds a 30-day forecast, and queues the overdue invoice reminders for you to approve and send.

Your month-end close goes from a painful afternoon of reconciling and second guessing to a plain English profit and loss statement that exports as a clean packet you can forward straight to your accountant.

• Tax season gets its own organiser.

• Contracts get reviewed.

• Margins get analysed.

None of this requires you to learn anything new or set anything up. You pick the workflow, Claude does the work, and nothing moves until you say so.

This frees up time that used to be a lot of very tedious clerical work, so his team could focus on higher value tasks instead.

Your Data Stays Yours

At this point, you might be thinking, this sounds good, but what happens to my business data when AI is connected to everything.

It is a fair question. In fact, when Anthropic surveyed small business owners directly, half of them named data security as their single biggest hesitation about AI.

So let’s deal with it straight.

I am not going to take Anthropic's word verbatim, so I had to do some research.

Sensitive data makes up 34.8% of what employees are putting into AI tools like ChatGPT today, up from just 11% in 2023.

Basically, you are putting a third of your business and personal data into AI without knowing where this goes.

Most of the time, it is not intentional. People are just trying to get their work done faster, and I understand that.

A KPMG report found that 69% of business leaders now cite AI data privacy as a concern, a significant jump from 43% just six months earlier. 60% of companies hit by a cyberattack close their doors within six months.

Here is my honest take because any time a new tool asks to connect to your books, your CRM, and your payment processor at the same time, your guard should go up.

How To Get It Running In Five Minutes

You need the Claude desktop app on a Pro, Max, or Team plan. That is the only prerequisite. Once you have that, the whole thing takes about five minutes.

Open Cowork, click Customize in the left sidebar, find Plugins, hit the plus button, and install Small Business.

From there, you have two ways to run it. You can type a forward slash and pick a workflow directly from the list, or you can just describe what you need in plain English, and Claude figures out which workflow fits.

To run a skill:

• Type / in the Cowork chat bar and pick it from the list, or

• Describe the job in plain English and Claude picks the skill that fits.

Something like “which invoices are overdue and who should I chase first” is enough. You do not need to know the exact command. You just need to know what is bothering you most right now.

The first thing I would recommend doing before you run any workflow is typing one prompt into the chat bar. Ask Claude to customise the plugin for your business. It asks you a few questions about what you do, how many people are on your team, and what is hardest right now, and then it rewrites the defaults to match your actual operation. From that point, every skill it runs carries your context. Your industry, your priorities, the way you like things done. It gets more tailored every time you use it.

One thing worth knowing before you start. Some of the design and email staging features may need a higher tier in the connected tool, like Canva or HubSpot. If something is not available, the workflow tells you and offers a workaround rather than just failing silently. Small thing, but it matters.

So I looked at how this actually works under the hood rather than just repeating what the press release says.

What I found is that the architecture is fundamentally different from dropping your data into a public chat window.

How?

You initiate every single workflow, which means nothing is running in the background.

Nothing sends, posts, or pays without your explicit approval. I like this a lot because I highly value humans in the loop.

Your existing permissions carry over exactly as they are, meaning if someone on your team cannot see payroll details in QuickBooks today, they still cannot see them through Claude. The walls do not move just because AI is in the room.

Third, Anthropic does not train on your business data by default on their Team and Enterprise plans. What happens inside your business stays inside your business.

I am not saying trust it blindly. I am saying the architecture is genuinely different from what most people are already doing without thinking twice about it.

So here is what I would suggest. Do not try to boil the ocean.

Go to Claude for Small Business and look at the 15 workflows available.

Pick the one that maps to the single biggest time drain in your week right now. The invoice you keep chasing. The campaign you keep pushing back. The month end close follows you home every single time.

Start there. Run one workflow this week and see what it actually feels like to have that thing done without you doing it.

If you want to go deeper before you dive in, Anthropic partnered with PayPal to build a free AI Fluency for Small Business course taught by real business owners who have already done this.

Actual people running actual businesses show you what worked for them. That is worth an hour of your time before anything else.

The gap between what big companies can do with AI and what small businesses can do just got a lot smaller. The only question now is whether you are the first one in your market to figure that out, or the one watching someone else do it.

Talk soon,

Sameer Khan

Creator of Solve with AI.

Hey Productivity Explorer,

Most companies have no idea what their AI actually costs.

Not even close.

You know what the demo or the pilot costs. You might even know your monthly API bill. But if I asked you what a single customer support workflow costs per resolution, or what your AI SDR spends per booked meeting, you’d probably go quiet.

Maybe pull up a number someone made up during a budget meeting six months ago that nobody has checked since.

This is the AI cost illusion, and it’s about to become one of the most expensive problems in your business.

Here’s what makes it so dangerous. Enterprise AI spend tripled in a single year, jumping from $11.5 billion to $37 billion in 2025. That is the fastest scaling of any software category in history.

Money is moving fast, decisions are being made fast, and somewhere between the boardroom excitement and the production deployment, the question of what this actually costs per outcome got skipped entirely.

It got skipped because AI doesn’t bill like anything you’ve managed before. SaaS is almost a flat line. You pay per seat, per month, done. AI bills by the token, so every prompt, retrieval call, reasoning step, and agent loop runs up a tab that compounds in ways most finance teams were never built to track.

The moment you move from a chatbot to a real agentic workflow, the economics stop looking anything like what you modeled in the pilot. Research shows agentic systems cost 5 to 25 times more per task than simple chat applications. Most people find this out when the production bill arrives.

That’s what this post is about.

Table of Content

1. The AI Cost Illusion

2. Why AI Costs Are Nothing Like SaaS Costs

3. The Hidden Cost Layers Most Teams Never See

4. Model Selection Is Now a Financial Architecture Decision

5. The Metric Nobody Is Tracking Yet

6. Access AI Workflow Cost Calculator (Paid Subscribers Only)

7. The Competitive Moat Is No Longer the Model

Why AI Costs Are Nothing Like SaaS Costs

Think about every software subscription you’ve ever managed. You knew exactly what it cost. X seats, Y dollars per month, done. The bill was predictable because the unit of cost was fixed. A person used the tool, or they didn’t, and either way, the number on the invoice didn’t change.

Read more

Hey Productivity Explorer,

Something made me write this post as I kept having the same conversation with someone exactly like you.

A business owner or leader running a real operation, paying real salaries, making real decisions. At some point in that conversation, you’d mention you’d switched to Claude. Or you were thinking about it. Or someone on your team had started using it, and you weren’t sure what to do next.

It happened enough times that I started paying attention.

When I looked it up, I understood why. In February this year, Claude went from number 131 on the Apple App Store to number one overnight. Paid subscribers more than doubled in four days. The enterprises moved first, as they always do.

Here’s what nobody is telling you, though.

Switching to Claude is not the move. That part is easy because you sign up, open it, type something in, get a great answer, and it feels like an upgrade. Because it is, but then you close the tab and go back to doing everything the same way you always have.

One task at a time.

That is the gap I kept seeing, and that gap is not in the tool but in how you’re using it.

This post is what I wish existed when those conversations started.

A real explanation of how to install Claude into the way your business actually runs, before you upgrade, before you commit, before you scale a single thing.

Table of Contents

1. The shift reaching SMBs right now

2. Prompts vs Skills: why 92% never break the habit

3. What installing Claude actually means

4. Claude architecture: Five layers, one system, built in the right order

5. Step-by-step setup: How to actually install Claude with Skills, Cowork, and scheduled tasks

6. Real business workflows examples

7. Common mistakes

8. What Claude can and cannot do

9. The compounding advantage: Why this decision matters and why the window is closing

The Shift Post People Miss

Lots of business owners are still treating Gen AI and especially Claude like a chatbot.

Read more

In the summer of 1942, the United States government gathered the most brilliant minds in physics under one roof in Los Alamos, New Mexico.

The mission was to build something that had never existed before, and at the center of it all stood J. Robert Oppenheimer, a man who understood, perhaps better than anyone alive, that raw genius deployed everywhere at once is not a strategy. It is a waste.

Oppenheimer did not personally wire every circuit, run every calculation, or supervise every lab bench.

He was the mind behind the mind.

The man the rest of the team escalated to when the problem got hard enough, actually to need him. Every other decision, every routine step, every mechanical execution that happened without him. Because of that structure, one of the most complex scientific projects in human history got done.

Now think about how you are probably deploying AI right now.

Every task, every workflow step, every token, running through the most capable model you have access to.

You never stopped to ask whether they do. That is the quiet inefficiency that is burning through your AI budget right now. The truth is that the majority of the AI product leaders have no idea how much it is actually costing them.

On April 9, 2026, Anthropic published something that changes this completely. They called it the Advisor Strategy. It is an architectural shift in how intelligence gets allocated inside an AI system, and it is the most important design pattern you have not implemented yet.

The core idea is almost embarrassingly simple once you see it.

• A cheaper model runs most of the workflow, including tools, iteration, and output

• When it hits a hard decision, it escalates to a stronger model for guidance (400–700 tokens)

• It resumes execution immediately, applying high-end reasoning only where needed

The way Oppenheimer would have done it.

The benchmark results Anthropic published are not marginal. Sonnet paired with an Opus advisor improved performance on SWE-bench Multilingual by 2.7 percentage points while cutting cost per task by 11.9%.

Haiku with an Opus advisor more than doubled its score on BrowseComp, jumping from 19.7% to 41.2%, at 85% lower cost than running Sonnet solo. You are not trading quality for savings. You are getting both at the same time, because you are finally spending intelligence where it actually matters.

What follows is exactly how it works, why the economics are so compelling, and what you need to do to test it inside your own workflows this week.

Table of Content

1. The Problem: Why You Are Overpaying for AI Right Now

2. What Is the Advisor Strategy

3. Why the Economics Actually Work

4. The Architectural Shift: From Orchestration to Escalation

5. What This Means for Your Business

6. How to Test It This Week: A 3-Day Implementation Plan

7. Risks and Limitations

8. The Only Question That Actually Matters

The Problem: Why Current AI Architectures Are Inefficient

There are two ways you are probably building AI agents right now, and both of them have a cost problem hiding inside them.

Read more

Most people step into a Chief AI Officer role and immediately ask, “What should we build?”

I’d ask a different question.

What do we kill?

Because here’s what nobody tells you before you walk through that door. The AI graveyard isn’t somewhere out there in other companies.

It’s unfortunately inside yours.

It’s sitting in your roadmap, consuming your budget, occupying your best engineers, and wearing the costume of progress while delivering absolutely nothing.

80% of AI projects fail to deliver their intended business value. Not because the technology doesn’t work. Not because the team isn’t smart. Because companies keep pouring resources into the wrong things and calling it innovation.

Day one as a CAIO should not be about vision decks and transformation speeches. It’s about triage.

The most valuable thing I could do for any organization is to have the courage to shut down the ones that were never going to work in the first place.

Table of Content

1. Why the First Question Is Wrong

2. Kill #1: The Chatbot Nobody Asked For

3. Kill #2: The Pilot That Never Graduates

4. Kill #3: The Model Obsession

5. Kill #4: The Internal ChatGPT Nobody Uses

6. Kill #5: The Governance Committee That Governs Nothing

7. This Is Not About Being Destructive

8. What I Would Build Instead

This Is Not What the Boardroom Wants to Hear

Killing a project is harder than launching one. Launching feels like leadership. It gets applause, budget approvals, and LinkedIn announcements.

Killing something requires you to look at the people who built it, the executives who championed it, and the sunk costs that are already on the books, and say, “This isn’t working, and we need to stop.”

That takes a different kind of courage entirely.

But the math doesn’t care about feelings. The average enterprise lost $7.2 million per abandoned AI initiative in 2025. The ones that waited longest to pull the plug lost the most, not just in money but in something far more expensive. Momentum. Trust. The willingness of the organization to try again.

Every zombie AI project that stays alive on life support is quietly taxing the projects that actually deserve a chance.

So before I sketch a single vision, before I hire a single engineer, before I sit in a single strategy session about what AI could do for this company, I need to walk the floor and find out what’s already broken.

Here is exactly what I would shut down on day one.

Kill #1: The Chatbot Nobody Asked For

You’ve seen this one. Every company has one.

It lives on the customer service page. It has a friendly name, maybe even a little avatar. It was announced internally with great excitement about “transforming the customer experience.”

Right now, it is handling about 4% of incoming tickets, frustrating everyone it touches, and quietly destroying customer trust one bad interaction at a time.

Nobody built this because customers asked for it. Nobody built it because the data showed a gap that AI could fill. They built it because a competitor launched one, someone in the C-suite saw it at a conference, and suddenly it became the thing we needed to do to prove we were serious about AI.

I would not call that a strategy. That is fear dressed up as innovation.

Air Canada learned this the hard way when their chatbot gave customers misleading information and ended up in court over it.

That is what happens when you deploy AI for optics instead of outcomes.

Vanity AI is the most expensive kind because it burns two things at once. Budget and trust. When your employees and customers have a bad AI experience early, you spend twice as much convincing them to give the next thing a chance.

On day one, this one is gone.

Kill #2: The Pilot That Never Graduates

You know this project too.

It has been “in testing” for fourteen months. There is a dedicated Slack channel. There are weekly update emails with color-coded status reports. Someone gave it a codename that people say with genuine affection. And if you ask anyone when it goes live, they smile and say, “We’re getting close.”

Close to what, exactly?

This is what I call Pilot Purgatory.

The proof of concept that works beautifully in a sandbox impresses everyone in the demo room, and somehow never makes it into the real world.

Not because the technology failed, but because nobody defined what graduation looked like before the project started.

If there was no clear path to production written into the original brief, there never will be. What exists instead is an expensive science experiment that makes people feel like progress is happening without any of the accountability that real progress demands.

The average organization scrapped 46% of AI proof of concepts before they ever reached production in 2025. Nearly half gone before they ever touched a real user, a real workflow, or a real business outcome.

I have sat in rooms where a pilot has consumed two engineers, a product manager, and 12 months of runway, and the team is still describing it as “early stage.” That is not an early stage. That is a project without a destination.

On day one, if it cannot tell me its production date, it cannot tell me anything useful at all.

Kill #3: The Model Obsession

This one is sneaky because it looks like real work.

The team is heads down. The engineers are brilliant. The Notion docs are immaculate, and if you ask them how the project is going, they will pull up a dashboard and show you accuracy metrics, F1 scores, precision, and recall numbers that would make any data scientist nod with genuine respect.

Then you ask the one question that matters.

What business problem does this solve?

…And the room gets quiet in a very specific way.

This is the model obsession trap. A team so focused on making the AI technically impressive that they forgot to stay connected to the reason it was built in the first place.

No business stakeholder owns the outcome. No success metric tied to revenue, retention, or cost. Just a model getting incrementally smarter in a vacuum while the rest of the organization waits for something useful to arrive.

Engineering teams spending quarters optimizing model performance while integration work sits untouched in the backlog is one of the most consistent patterns separating AI projects that ship from AI projects that die quietly in a shared drive folder.

The brutal truth is that a model with 94% accuracy that nobody uses is worth less than a simple rule based system that solves the actual problem every single day.

I am not paying for impressive. I am paying for impact.

On day one, if the success metric lives inside the model and not inside the business, the project does not survive the morning.

Kill #4: The Internal ChatGPT Nobody Uses

This one has the biggest gap between how it sounds in the boardroom and what it actually does in the real world.

The pitch is always compelling: We need our own secure internal AI assistant.

We cannot trust third party tools with our data. We will build something custom, something that knows our company, something our people will actually want to use every single day.

Two million dollars later, the system summarizes emails and answers basic questions.
It’s over-engineered, slow, and disconnected from real workflows.

Every query runs through layers of orchestration, retrieval, and approvals, turning a simple task into a prolonged wait.
Not because the models are weak, but because the system was designed without a clear business outcome.

I have seen this project in more organizations than I can count, and it follows the same arc every single time. Months of infrastructure work, a lengthy security review, and a big internal launch with an impressive demo.

Six weeks later, adoption flatlines because it only does what employees could already do with a $20/month tool they were likely using anyway.

The justification is always data security and governance.

I do agree that sometimes that concern is legitimate. But more often, it is a cover story for something less comfortable to admit. The organization wanted to feel like it was building something, not just buying something. There is a pride tax buried inside this project, and the company is paying it every single month.

MIT’s research is unambiguous on this. Companies that force generative AI into existing processes with minimal workflow redesign see no measurable business impact. A private wrapper around a foundation model with no workflow integration is the purest version of that mistake.

On day one, the budget it was consuming goes somewhere that will actually move the needle.

Kill #5: The Governance Committee That Governs Nothing

This is the kill that will make people uncomfortable. That’s the intent.

Because not every zombie project is a technical one. Sometimes, the most dangerous AI initiative in a company is not a model or a chatbot or a pilot stuck in purgatory.

Sometimes it is a 12 person committee that meets on the third Thursday of every month, produces beautifully formatted PDF reports, and has never once made a decision that changed how anything actually works.

I call this Governance Theater.

It exists because someone senior got nervous about AI risk, which is a completely legitimate feeling, and the response was to form a committee, which feels like a solution but is actually just a way of distributing the discomfort across more calendars.

The committee discusses and documents risk.

It creates frameworks for thinking about risk. Meanwhile, the teams doing real AI work either wait months for approvals that mean nothing or quietly route around the committee entirely because they cannot afford to wait.

Real governance is not a meeting. It is not a PDF. It is not a working group with a charter that nobody has read since the kickoff call.

Real governance is embedded directly into the workflow. It is the guardrail that exists inside the system, not the report that describes what the guardrail should probably look like someday.

Traditional governance that locks everything down and documents everything creates the exact opposite of what AI development actually needs. You end up with bureaucracy wearing the costume of responsibility, while the people doing the real work find ways around it just to get anything done.

On day one, the committee is dissolved.

The two people in that room who were actually driving anything useful get seats at the real table. The rest of the meetings get cancelled, and everyone quietly gets three hours of their week back.

This Is Not About Being Destructive

If you have read this far and you think this is the perspective of someone who does not believe in AI, you have missed the point entirely.

That’s not what I am sharing.

Everything I just described killing represents millions of dollars, months of engineering time, and organizational energy that could have been pointed at something real.

That is what this is about. Not cynicism. Surgery.

Because here is what I have learned watching companies navigate this. The ones actually winning with AI right now are not the ones who launched the most initiatives.

They are the ones who said no fast, killed ruthlessly, and protected their best people’s attention long enough to go deep on the things that actually mattered.

McKinsey’s research confirms it. Organizations reporting significant financial returns from AI are twice as likely to have redesigned their workflows before they ever selected a model or wrote a single line of code. They started with an unambiguous business problem and worked backwards. Every single time.

That discipline is the whole game.

The companies stuck in the AI graveyard did the opposite. They started with the technology and went looking for a problem to attach it to. They optimized for looking innovative rather than being effective. And they are now sitting on hundreds of millions in sunk costs, with boards asking increasingly uncomfortable questions about where the returns are.

Killing bad projects is not the end of an AI strategy. It is the beginning of one worth having.

The moment you stop protecting things that do not deserve protection, you create space for something that does.

What I Would Build Instead

Once you clear the deck, something interesting happens.

The engineers buried in pilot purgatory are suddenly available. The budget funding a chatbot nobody asked for is back on the table.

The two smart people writing governance PDFs nobody reads are now in the room where real decisions get made. For the first time in a long time, the organization has one resource no amount of AI investment can manufacture.

Focus.

But focus without direction is just a cleaner kind of confusion. And this is where most companies make their next mistake. They kill the bad projects, free up the budget, and then immediately start filling the space with new initiatives before they have done the one thing that actually determines whether any of it works.

Choosing the right problem.

I wrote about this in depth in my last piece.

The companies generating real AI returns are not the ones doing more. BCG found that leaders prioritize an average of 3.5 use cases, compared to 6.1 for everyone else, and generate twice the ROI.

Because they are more disciplined about where AI actually creates business value versus where it just looks impressive in a demo.

If you want a practical framework for finding where that value actually lives inside your business, and how to quantify it before you spend a single dollar, that is exactly what I broke down in how to find $100M in AI value without building more models. Start there before you start building anything.

Because the question was never what AI can do.

It was always about which decisions in your business are broken, what it costs you every day they stay broken, and whether AI is genuinely the right tool to fix them.

That clarity is what separates the 5% that win from the 95% still searching for their first real return.

So here is my question for you.

What is the zombie project sitting inside your organization right now that everyone knows is not working, but nobody has dared to kill yet?

Drop it in the comments. Because I have a feeling you already know exactly which one it is.

Talk soon,

Sameer Khan

Creator of Solve with AI.

Here is something most AI consultants will never tell you.

The biggest barrier to AI value in most companies is not the technology. It is not the data. It is not even the talent. It is a room full of smart people who have quietly decided that AI is not in their interest.

I watched this play out firsthand. The company had a budget. It had data. It had access to the tools. What it did not have was a leadership support to put its weight behind it, and in that vacuum, something predictable happened.

People defaulted to protecting what they already owned. Manual processes became deliberate choices. Headcount justification became the real agenda.

A small team of believers was left spending more time proving the case for AI than actually building it, fighting for every inch of progress against an organisation that had silently decided the old way was safer.

Does any of that sound familiar?

Maybe you are the one in that room right now, carrying the case for AI almost entirely alone. Or maybe you are leading the charge from the top and still wondering why the results are not showing up where they should.

That is a prioritization problem dressed up as a technology problem, and it is far more common than anyone in this space wants to admit.

Because here is the uncomfortable truth. McKinsey’s 2025 State of AI survey found that 88% of organizations are using AI in at least one business function, and yet only about one in three have begun scaling their programs.

Your company is probably in that 88%. You have the tools. You have the pilots. You have the slide decks, and yet only 39% of organizations attribute any EBIT impact to AI, and among those, most report that less than 5% of their EBIT is attributable to AI. Broad adoption, thin results.

The tools are everywhere. The outcomes are not.

So what is actually going wrong?

It is the absence of disciplined thinking about which decisions AI should actually be improving, where those decisions sit in your business, and what they are genuinely worth when you do the math honestly.

You are probably running dozens of AI experiments right now and calling it a strategy. You are measuring activity instead of impact. And every week that passes without a clear answer to that question is a week your competitors are quietly pulling ahead.

The companies generating real AI returns are not doing more than you.

They are doing less, but choosing better. BCG’s 2025 research found that leading companies prioritize an average of 3.5 use cases, compared with 6.1 for others, and anticipate generating 2.1 times greater ROI on their AI initiatives.

Twice the return. From focus, not from more experimentation.

This is exactly what the next 30 days can fix.

Table of Contents

1. The Real Barrier to AI Value

2. AI Value Lives in Decisions, Not Models

3. What a $100M Opportunity Actually Means

4. Week 1: Map the Decisions That Move Your P&L

5. Week 2: Find Where Those Decisions Are Broken

6. Week 3: Quantify the Prize

7. Week 4: Narrow to Two or Three Bets and Fund Them

AI Value Lives in Decisions, Not Models

Before I show you the method, I need to shift how you are thinking about this. Because if you walk into this exercise with the wrong mental model, you will end up in the same place you started.

Read more

When the Claude Code leak broke earlier this week, I watched the takes roll in almost in real time. “Anthropic got hacked.” “AI model exposed.” “Major security breach.” People were sharing it like the sky was falling.

You likely assumed one of three things:

• Claude got hacked.

• The model was exposed.

• This is another AI security failure.

Reality check: None of that is true.

Anthropic has said this was a release packaging mistake, not a breach. A public npm release accidentally included a source map that exposed a large part of the Claude Code application layer, reportedly over 500,000 lines of code.

No model weights were leaked.

No customer data or credentials were exposed, based on current disclosures.

This wasn’t a model failure.

It was a basic software release failure that exposed the system around the model, and that distinction is what most people are missing.

Table of Content

1. The leak is being misunderstood

   What actually happened vs what people think

2. What was really exposed

   The agent and system layer, not the model

3. Where AI value is actually shifting

   From models to orchestration, memory, and workflows

4. Why the “wrapper” narrative breaks

   The system layer is becoming the product

5. AI risk is now full-stack

   From model risk to release, supply chain, and permissions

6. What leaders should do differently

   Where to invest and what to fix now

What was really exposed

When you hear source code leaked, your brain probably goes straight to embarrassment and PR damage. Contained. Uncomfortable but not serious.

That framing misses everything that matters.

What got exposed was the orchestration layer. The part of Claude Code that makes it actually useful inside a real business.

1. Tool routing.

2. Permission architecture.

3. Memory that persists across sessions.

4. Hooks that fire automatically during the agent lifecycle.

5. Subagents that spin up, execute, and report back.

6. MCP integrations that connect everything into your existing stack.

Think about that for a second.

The model itself, the thing everyone obsesses over in benchmark comparisons, was never touched. What leaked was everything built around it to make it work in the real world.

That distinction matters more than most people realize. Because it tells you where the real engineering effort actually goes inside a frontier AI lab. It is not just training a smarter model. It is building the control system around the model that makes it reliable, governable, and deployable inside complex environments.

Where AI value is actually shifting

The leak did not just expose Anthropic’s code. It exposed something the market has been slow to admit.

The model is not where the value lives anymore.

Think about how most conversations around AI tools still go inside your organization. Someone asks which model is best. Teams run benchmark comparisons. Leaders debate GPT versus Claude versus Gemini as if it were a horse race.

In fact, underneath all of that is a silent assumption that the model is the product.

This leak quietly dismantles that assumption.

What Anthropic built around Claude is not a thin convenience layer. It is an entire system. Agent loops that manage how tasks get broken down and executed. Context handling that decides what the model remembers and what it forgets.

Permission architecture that controls what the agent can and cannot touch.

Workflow logic that handles the messy, unpredictable edges of real work. That is the product's operating system.

Here is why that matters for you specifically.

Two organizations can run the exact same model and get completely different outcomes.

• One gets reliable, governable, repeatable output.

• The other gets chaos dressed up as capability.

The difference is rarely the model. It is the system built around it.

That is where enterprise value is concentrating right now. Not in who has access to the best model, because access is becoming a commodity. But in terms of who has built the better orchestration layer around it.

The leaders still debating model selection as their primary AI decision are looking at the wrong layer entirely. The real competition has already moved one level up.

Why the “wrapper” narrative breaks

You have probably heard it, and maybe you have even said it. Most AI products are just wrappers.

The implication being that if you strip away the branding, there is nothing underneath worth paying for. The model is the whole product, and everyone else is just reselling access to it with a nicer interface.

This leak makes that position very hard to defend.

Here is what engineers actually found inside when they dug into the code. Only about 1.6% of the leaked code directly involves the AI model itself. The rest is engineering: context management, multi-agent coordination, memory systems, tool orchestration, and permission handling.

Read that again. 1.6% is the model. Everything else is the system built around it.

The scale of that system is not trivial. The tool system uses a plugin architecture in which each capability, from file reading to Bash execution, is a discrete, permission-gated module.

Its base definition alone accounts for 29,000 lines of TypeScript. The largest single module is the query engine at 46,000 lines, handling all LLM calls, streaming, caching, and orchestration.

The leak revealed that Claude Code is not just a wrapper around a Large Language Model, but a complex, multi-threaded operating system for software engineering.

When something that sophisticated gets exposed, competitors pay attention. Not because they wanted to see Anthropic embarrassed, but because what was inside was genuinely valuable and genuinely hard to build.

Here is the uncomfortable truth for anyone still using the wrapper label. The products being built on top of frontier models are getting thicker, not thinner. The orchestration layer is expanding.

The permission and memory systems are becoming more sophisticated with every release cycle.

Five context compaction strategies, 14 cache-break vectors, 23 security checks on bash commands, and three subagent execution models. That kind of engineering depth does not happen by accident, and it does not get replicated overnight.

Enterprise buyers are not paying for raw model access. They are paying for trustworthy execution, reliable output, auditable decisions, and controllable autonomy. None of that lives in the model. It lives in everything built around it.

The wrapper was never just a wrapper. It was always the product.

AI risk is now full-stack

Within hours of the leak going public, something predictable happened. Threat actors moved fast.

Zscaler’s ThreatLabz researchers found a malicious GitHub repository disguised as a leaked TypeScript source code for Claude Code, claiming to offer unlocked enterprise features and no usage restrictions.

The README file replicated the exact language researchers were using to describe the real leak. It looked legitimate. People downloaded it.

Inside was something far worse than leaked code. The malicious archive contained a dropper that silently installed two pieces of malware on your machine.

• One stole your sensitive data.

• The other hijacked your device entirely, turning it into a tool criminals could route their own attacks through without you ever knowing.

That happened within hours. Not days. Hours.

GitHub immediately disabled over 8,100 repositories within hours. But the code had already spread. Every fork was a potential delivery vehicle. Every curious developer downloading an unofficial copy was a potential target.

But here is the part most commentary skipped entirely.

A concurrent, separate supply chain attack on the axios npm package occurred hours before the leak itself.

Anyone who installed or updated Claude Code via npm on March 31 between 00:21 and 03:29 UTC may have pulled in a malicious version containing a Remote Access Trojan. Two completely separate incidents, colliding on the same morning, on the same platform, targeting the same users.

Think about what that morning looked like inside any organization running Claude Code. Your developers are updating a routine package. Two different attack vectors are live simultaneously. One is a trojanized dependency hiding inside a legitimate update. The other is a flood of fake repositories that look credible enough to fool people who know what they are looking for.

This is what modern AI risk actually looks like.

It is not a rogue model hallucinating in a board presentation. It is your release pipeline. Your dependency chain. Your developers are downloading what looks like an official repository because it appeared at the top of a Google search.

The real lesson is not to panic. It is that you cannot keep a lax runtime posture just because today’s headline was about source disclosure rather than direct compromise.

The risk surface of AI products is expanding fast.

You are no longer just governing what the model says.

You are governing everything around it, the packages it depends on, the connectors it integrates with, the repositories your team pulls from when something interesting leaks on a Tuesday morning.

What leaders should do differently?

The wrong response is to panic about frontier models or spend the next quarter debating which AI vendor to trust.

The right response is to look honestly at where your priorities are today and ask whether they match what this incident just revealed.

The leak showed you where the real engineering depth lives. It is not in the model. It is in everything around it. Which means if your AI strategy stops at model selection, you are building on a foundation that is missing most of the structure.

Here is what that means practically for you right now.

Step 1. Audit your orchestration layer like a product, not a side project.

Map out every workflow, permission control, memory system, and tool integration your AI agents touch today.

If you cannot describe how your agents make decisions and what guardrails exist around those decisions, that is your first gap.

This is where your competitive advantage either compounds or collapses, and most organizations have not even started designing it deliberately.

Step 2. Treat your release pipeline and dependency chain as critical infrastructure.

A single misconfigured field in package.json can expose an entire proprietary codebase to the public. If that is true for Anthropic with a reported $19 billion annualized revenue run rate, it is true for your team.

Audit your .npmignore files, your package configurations, and your build pipeline controls. Assign clear ownership. Make it someone’s job to catch this before it ships.

Step 3. Lock down your connector and permission architecture before your agent footprint grows.

As AI tools gain deeper access to codebases, filesystems, and external APIs, the consequences of downtime, rate limits, model deprecations, or unexpected incidents become more severe.

Every new integration your agent touches is a new surface area. Define what your agents can and cannot access, and design those boundaries now. Redesigning them after something goes wrong is significantly more expensive.

Step 4. Stop letting model selection dominate your AI investment conversations.

The next time your team sits down to evaluate AI tools, shift the first question from “which model is best” to “what is the system we are building around it.”

Benchmark comparisons are useful, but they are looking at 1.6% of where the value actually lives. The other 98.4% is the system your team designs, governs, and iterates on.

Step 5. Build your incident response muscle before you need it.

The Claude Code situation went from a routine npm release to active malware campaigns in hours. That speed is not going away. The organizations that handled it well already had clear ownership, clear processes, and clear communication lines before anything happened.

Run a tabletop exercise.

Decide today who owns AI-related incidents in your organization, what the escalation path looks like, and how fast you can respond.

Talk soon,

Sameer Khan

Creator of Solve with AI.