There Is Not Enough Claude
AI will prove to be a General Purpose Technology, and demand for compute will be beyond anything we can imagine today.
The internet is the most groundbreaking invention of modern times.
In all times, it’s only second to inventions like writing, the wheel, or discoveries like fire that made humans endure and thrive over the course of centuries. If you leave these aside, nothing comes close to the internet's impact.
What the internet did differently from all other groundbreaking human inventions that changed the world, like the steam engine, was that it made the transfer and spread of knowledge infinitely faster. That led to shorter cycles of groundbreaking discoveries, accelerating human development.
The historical development of internet traffic illustrates this.
Tim Werner Lee invented the World Wide Web in 1989 while he was working at CERN. That was the ground zero. The internet traffic was essentially nonexistent.
CERN opened the World Wide Web to the public for free in 1993. Later that year, the global internet traffic reached 8,715 gigabytes per month. By 1997, this number reached 5 million, more than a 500x increase. What do you think it currently is? It’s around 733 exabytes. 1 exabyte is 1 billion gigabytes. This is how far we have come.
In many aspects, 2026 is to AI what 1997 was to the Internet.
This is even though AI has been adopted pretty rapidly compared to other technologies so far, including the internet.
I think this very rapid adoption led many people to underestimate AI’s capabilities. We suddenly forgot that this technology is still in its infancy and started to assess its impact, drawing on its relatively high adoption rate.
The simple assumption was as follows: If the curve goes like this, it’ll hit the internet level adoption in a few years, and it’s unlikely to lead to an internet level transformation in this period, based on what we are currently seeing and what we can imagine for the foreseeable future.
This led many of us to question the data center buildout, thinking that we’ll have a massive amount of excess capacity if the AI doesn’t fulfill its promises.
What most people miss is that demand is a function of three things:
Adoption
Capabilities
Spillover effects
Most people are focused on the adoption and are completely overlooking capabilities and spillover while they assess AI’s potential and future demand.
This is what we are going to lean into today. I’ll try to explain why I believe the demand will far exceed anything we can build in the foreseeable future by discussing:
How adoption, capabilities, and spillover determine demand.
How AI’s capability improvements are driving demand.
How large is the spillover area for AI?
Finally, I’ll briefly discuss our exposure to compute demand and how I am looking to increase this exposure over time.
Let’s get started.
Demand = Adoption X Capabilities X Spillover
Let’s take a closer look at the adoption curves I have just shared above.
What do you think is striking here?
To me, the most striking thing is that television actually has a steeper adoption curve than the internet. Within almost 10 years after the introduction, television reached 80% of the American population. For reference, this was just around 45% for the internet.
Yet, despite exponential early demand and adoption, we know that the demand for television had been pretty flat even before streaming became mainstream.
Why? Because, despite its contribution to the human experience as a standalone invention, it’s not what is called a “General Purpose Technology” or GPT.
You may ask what the hell is a General Purpose Technology?
The term was coined by two economists, Timothy Bresnahan and Manuel Trajtenberg, in their seminal 1992 paper called “General Purpose Technologies, Engines of Growth?”
Their seminal insight was that most of the industrial and economic growth in human history stems from a handful of technological developments/inventions, while the rest of the gadgets contributed only marginally to our development.
When they dug deeper into what makes a GPT, they found three properties:
Pervasiveness: This is adoption. GPTs are adopted by the majority of humanity.
Improving Capability Frontier: Capabilities of GPTs don’t stay static; they improve over time.
Complementarities and Innovation Spillover: GPTs spawn other innovations as the productivity of R&D in downstream sectors increases as a consequence of innovation in the GPT.
Adoption is straightforward; for something to become a GPT, it should reach wide adoption. But let’s dig a bit deeper into capability frontiers and innovation spillover.
I am using the term “improving capability frontier” to describe the expanding usage of technologies; this isn’t the original term used by the writers. The original term Bresnahan and Trejtanberg used was “technological dynamism.”
They found that:
“continuous innovational efforts about GPTs, as well as learning, increase over time the efficiency with which a GPT performs its generic function. This may show up as reductions in the price/performance ratio of the products, systems, or components in which the GPT is embodied, or as multidimensional qualitative improvements in them.”
General Purpose Technologies, 1992, p.5
Their initial focus was on the cost of performance by a GPT. They observed that, unlike other technological advancements, the cost of performance by a GPT plummets over time as wide adoption leads to higher innovation efforts on GPTs, which inevitably leads to price reduction over time.
So, what they understood from “dynamism” was increasing efficiency, not capability expansion. They didn’t dwell on the idea that a GPT could, over time, become able to perform a function that it initially wasn’t able to. The transistor was their example.
They argued that the original semiconductor was performing binary logic. Developed chips essentially do the same thing. They do it faster, at a larger scale, which leads to significant cost reduction, but they essentially do the same thing.
This is essentially true, as the very reason they are faster can scale better is that it’s become cheaper and cheaper to produce transistors at smaller sizes, so we can pack more of them in a chip, which delivers better performance at a lower price. This is a direct result of widespread innovation efforts on transistors, as the writers suggest.
However, Bresnahan himself revised this idea in 2010 in his solo chapter for the Handbook of the Economics of Innovation and wrote that “a GPT may have attributes other than cost, and that the transition from one GPT to another or the creation of an important new version of a GPT may involve changes in attributes other than cost.”
He pointed to the possibility of capability expansion by admitting that a new version of the same GPT may have different non-cost attributes than the original.
In short, based on where the literature has reached today, the ability to improve both in terms of costs and capabilities is an integral property of a GPT.
The second distinguishing factor of GPTs that requires closer attention is Complementarities and Innovation Spillover. Again, this is the term I picked for better clarity. The original term Bresnahan and Trejtanberg used was Innovational Complementarities, or simply IC.
Their powerful insight was as follows:
“GPT's are characterized by the existence of innovational complementarities with the application sectors, in the sense that technical advances in the GPT make it more profitable for its users to innovate, and vice versa.”
General Purpose Technologies, 1992, p.5
In a more straightforward language, GPTs are critical inputs to downstream R&D processes. Thus, the cheaper a GPT becomes, the cheaper all the downstream R&D processes become, leading to a higher volume of innovation that can reach more people. Those innovations can also be inputs to further downstream R&D processes.
They again illustrated this by using semiconductors as an example:
Their insight was that these characteristics of GPTs create a virtuous cycle that leads to ever-increasing demand for a GPT:
Step 1: A GPT improves both in cost efficiency and capability.
Step 2: An improved GPT makes R&D more viable and profitable for downstream sectors.
Step 3: Downstream sectors innovate, developing new and better products. This increases their demand for the GPT input.
Step 4: The increased demand and the expanding set of application sectors make it more profitable for the GPT provider to invest in further R&D.
Step 5: Return to step 1. GPT improves again, triggering a new cycle.
The natural consequence that Bresnahan and Trejtanberg arrive at is that the demand curve for GPTs reminds us of increasing returns to scale, where a unit of input turns into more than a unit of output.
This is because when we are looking at a GPT, demand doesn’t closely follow adoption as it does with other gadgets, but it’s a function of adoption, improvement in the capability frontier, and innovation spillovers.
This is why television has a very different demand curve than the internet, despite its faster historical adoption:
Television is highly limited when it comes to capabilities and spillover. Its capabilities haven’t expanded much since its invention, and it’s not an input to many downstream R&D sectors. It’s not a GPT. Thus, its demand curve closely follows adoption. When adoption flatlines, so does the demand.
The internet, on the other hand, is essentially a GPT that is input to many other sectors, its capabilities improve by the day, and it leads to other innovations. Just last year, the internet traffic grew by 15% according to Cloudflare. If you look at the adoption curve, you’ll see a maturing technology; if you look at demand, you’ll see sustained growth as it’s a GPT.
This is why the seemingly “already high” adoption of AI doesn’t mean much for the future demand, as it’ll be a function of capability improvements and innovation spillover, because it squarely fits the definition of GPT. This is what most people miss.
Imagine how the demand for microprocessors has increased since Bob Noyce invented them in 1959. Then go beyond it for AI because it’s one of the rare GPTs that can experience a capability expansion, not just cost reduction. And remember, one of the parameters of the demand for a GPT is capabilities.
Improved Capabilities = Higher Demand
This is the core feature of GPTs.
Capability expansion is one of the main drivers of demand for GPTs, together with adoption and innovation spillovers. Because of this, the demand for a GPT can grow for decades or even centuries after the widespread adoption is achieved for generic uses.
Bresnahan and Trejtanberg only observed this in their 1992 paper, but didn’t provide an empirical background. They later did this by tracking how an upgrade from an early steam engine to what’s known as the Corliss Steam Engine expanded the use cases of the steam engine and increased demand.
Some basic context about steam engines first.
In the simplest terms, heat turns water to steam, which expands with enormous force. Channel that force against a piston, connect the piston to a wheel, and you have continuous rotary motion, which is basically what a steam engine is:
The problem with the early steam engine was that it didn’t have a mechanism to control the power delivery. Steam flooded the cylinder at full blast regardless of what the engine was actually doing.
Imagine you have 20 looms connected to that engine via belts and shafts. Someone starts up 5 more looms. Then someone shuts off 10 looms for maintenance. The resistance drops, and the engine speeds up. The engine has no way to automatically adjust how much steam it's getting; it just takes whatever comes.
This worked fine for work where jerky, uneven power delivery didn't matter much. Sawmills, for instance. You're just pushing a blade through wood, and if the speed fluctuates a bit, who cares? Pumping water out of mines, just straight up-and-down motion, no precision needed. Basic grain milling, steamboats, and locomotives, etc.
But it wasn’t helpful for things that required precision, stability, and even power delivery. Delicate textiles, for instance. Cotton thread snaps easily, and an engine that lurches when the load shifts means constant breakage. So textile manufacturers were stuck producing only coarse, low-grade fabrics with steam, but not fine textiles.
This is where a new version of a GPT, the Corliss Steam Engine, stepped in.
It changed the valve design. Previous engines used a single slide valve that opened and closed in a fixed pattern. Steam always entered for the same duration regardless of conditions. Corliss replaced that with separate intake and exhaust valves that could be opened and closed independently, and connected them to the governor through a trip mechanism that could release the valve at any point in the stroke.
That meant the governor could decide, mid-stroke, that the steam was enough and snap the valve shut. Instead of a fixed pattern that delivered the same amount of steam regardless of what the engine was doing, the Corliss engine could sense its own speed and adjust its fuel intake cycle by cycle, in real time.
This had two important implications:
30% higher fuel efficiency than conventional fixed-cutoff steam engines
The steam engine became suitable for textile mills, factory lines, and electricity generation.
Because of this new capability, the use cases expanded substantially, and demand for steam horsepower (hp) kept increasing.
As you see below, the steam hp grew pretty fast between 1838 and 1850. The Corliss engine was commercialized in the early 1950s. Then, from the 1850s to the 1890s, number of steam engines increased by another 8x:
We know that Corliss contributed substantially to this growth, as a rapidly growing industry, textiles accounted for almost 50% of all Corliss engines. Before Corliss, steam hp wasn’t that widespread in textiles due to the limitations explained above.
Thus, if it weren’t for Corliss, steam hp wouldn’t be that common in textiles. This is a concrete example of a new version of a GPT that has improved capabilities, creating demand that didn’t exist previously.
This is exactly the case for AI, and we are at the very early innings of it.
Take a look at this:
In July 2024, Claude had around 15 million monthly active users. This number doubled to ~30 million by July 2025. Yet, token usage increased by 50x.
This was because the demand is a function of adoption and capability expansion. Every user was now consuming, on average, 25x more tokens because Sonnet 4.0 was able to do many things that Sonnet 3.5 wasn’t able to.
Sonnet 3.5 was working, but it was annoying. It was impossible to code a decent app or software end-to-end. Sonnet 4.0 was almost there. We now have 4.6 Opus, and it one-shots basic-to-medium complexity apps.
Open Router data further validates this.
Even though active user counts didn’t move much from July to December 2025 (it, of course, increased, but it certainly didn’t triple), token usage tripled:
It’s a direct result of expanding capabilities. If an LLM can’t vibe-code an entire app, you will use it as a sidekick; if it can, you’ll make it your employee. In the latter, you consume multiples of what you would in the former.
This is easily observable even without data. Think about it. When ChatGPT was launched, we didn’t have an image model or a video model. Now people are generating whole advertisements and short animations/movies using these models. This is a serious capability expansion that drives demand.
We should acknowledge that every new capability expansion will likely require more compute, at least initially. We have had two clear examples of this over the last two weeks.
Anthropic announced computer use in Claude Code and the ability to control it remotely through the Claude app on your mobile:
The moment I saw these, I thought token usage would skyrocket. People using Claude Code have already caused a significant jump in token usage. Now we are talking about AI itself using AI tools through your computer, which gives it the ability to use other internal or external AI applications to achieve a goal.
What do you think this will do to the usage?
Now model this to every function:
AI uses AI+computer to create end-to-end marketing campaigns.
AI uses AI+computer to do academic-level research and write reports.
AI uses AI+computer for end-to-end customer management.
The list goes on and one and this is just one step. This is just the step where AI can use a local computer + other AI tools. There will be a step where AI will be able to do it in every edge device, even in your Apple Watch.
In this sense, it’ll be much like the internet. It’ll infiltrate everything, but it’ll be even more explosive than the internet because token usage can self-perpetuate. AI using AI in a loop. This wasn’t the case for the internet.
And even then, the internet traffic grew 1,000x from 2002 to 2022:
We aren’t even in 2002 for AI. 2026 to AI is probably what 1995-1996 was to the internet. We have seen that every inch of capabilities we unlock drives insane growth in demand, and we are just starting to scratch the surface of it. Demand that’ll be driven just by capability expansion looks beyond comprehension.
Innovation Spillovers Perpetuate The Demand
In their original article, Bresnahan and Trajtenberg avoided listing GPTs, as they were trying to lay down the theoretical basis. Instead, they provided only three examples, thinking that it would be clear that they were not exhaustive: Steam engine, electric engine, and semiconductors.
It was only in 2005 that a broader attempt to list GPTs came. Richard G. Lipsey, Kenneth I. Carlaw, and Clifford T. Bekar listed 25 GPTs in their 2005 book, Economic Transformations: General Purpose Technologies and Long-Term Economic Growth.
This was their full list:
When I saw their list, the wheel instantly caught my attention. It was invented between BC 4000-3000. So, it’s been around for 5 to 6 thousand years, yet the spillover it created keeps expanding, which keeps increasing the demand for wheels.
Here is a very interesting example.
By the 1970s, we used wheels in daily life, in simple and complex machines, in industrial lines, etc. Without wheels, there wouldn’t be a modern society. At the time, it was probably hard for anybody to imagine a new use case that would substantially increase the overall demand for wheels. Yet, we still didn’t have bags with wheels.
It never stops surprising me that we put a man on the moon before we put wheels below our luggages.
It was in 1972 that we put two wheels under a luggage. It had only two wheels on the back. You had to grab the handle at the front, lift it a little bit, so that it could slide on its rear wheels. Isn’t that insane? It took another 20 years for the modern Rollaboard with 2-4 wheels and a retractable handle to be introduced.
Imagine how this little innovation affected the demand for wheels? Every traveller in the world created a demand for 2-4 wheels, a demand that didn’t exist before.
This is a clear example of how innovation spillovers from GPTs can be sustained for a very long time, even for thousands of years, as innovations create new use cases.
We could easily see such a development for the internet as well.
When it was first introduced, streaming didn’t exist; now it’s the norm. How do you think this affected the internet traffic?
This is one of the crucial points the current skeptics are missing when modeling the demand for AI and thus for compute. They look at today and predict the near future, and can’t see factors that’ll substantially increase the demand. Yet, innovation spillovers from GPTs can easily span decades, even centuries.
It takes time to see substantial productivity gains from a GPT as we discover the capabilities and enhance them only slowly. If we could go back in time, we would likely see pretty stable demand for the wheel in its first millennia or two. Use cases would be largely the same.
Indeed, productivity gains from new technologies may take time to strongly manifest themselves:
Thus, when people look at the immediate impact of AI and don’t see any material contribution to growth and reach the conclusion that AI may be a bubble, or we may be overbuilding, they are completely ignoring the very long period of capability expansion for new GPTs and innovation spillovers.
The data below falls into the same mistake. It shows that real GDP growth in the US in the first half of 2025 was essentially 0.1%, excluding investment in processing equipment and software. It implies that AI itself didn’t contribute to the growth:
It’s true, but we are still at the very early innings of capability discovery/expansion and, naturally, spillover is very limited. At this stage, we can’t see the gains at scale. We need to look for the early signals.
And we are seeing many of these early signals. The most obvious one is coding:
It’s a very big deal that the IOS app releases, which were stable from 2022 to 2025, skyrocketed after Claude Code became mainstream and foundational models became capable enough.
I think many people underestimate the importance of this and tend to see coding as an isolated area where AI has a disproportionate effect. They still think the effects outside this will be pretty limited.
This is false. We are seeing the early signals of spillovers.
An amazing story is that an Australian entrepreneur who had no experience in biotech was able to develop an mRNA cancer vaccine for his dog. Yes, he took help from professionals at later stages. Yes, the vaccine didn’t completely kill the tumor and was only able to shrink it.
Qualifications exist. But the main question is whether this was even possible before?
It was nowhere near possible.
Another example is Matthew Gallagher, who basically created the first one-person billion-dollar company. He literally vibe-coded an entire telehealth company, used AI to create image and video ads, and sold through Facebook ads.
Again, the question is simple—was it even possible before?
It wasn’t. Things that weren’t possible before are becoming possible. This is the definition of innovation spillovers from a GPT. We are seeing anecdotes today; we’ll see them at scale in the future.
You can understand how broad the spillover area for AI is by thinking in just two steps:
Almost everything that works on electricity can be made smarter.
Then all those things will be used in related/unrelated fields to create new things.
Imagine how this will affect the demand.
We are just scratching the surface now. At this stage, we can’t see the impact at scale. What we can see are signals, and we are seeing very powerful signals.
How To Position For The Coming Demand?
We won’t have enough Claude anytime soon, as we are just scratching the surface of an emerging GPT. Thus, we won’t have enough compute.
Cloud providers are the obvious bets, but there are caveats.
First, demand doesn’t mean execution is granted. Firms still need to execute to capture as much demand as they can without risking their survival. If there is a mismatch between existing demand and their capabilities, the broader demand projections won’t save them.
Second, many hyperscalers have other businesses that can be affected by AI. Thus, valuations should account for other parts as well. If the survival of the software business was guaranteed, Microsoft would be a no-brainer here. But we aren’t sure how the productivity software business will evolve at the age of AI.
Third, Bitcoin miners pivoting to AI infrastructure to capitalize on the power they have already secured are exposed to higher execution risks. Despite this, the market has richly valued them, largely based on their connected power potential. Thus, most of them are out of scope for me.
This leaves me two players: Amazon and Nebius.
We own both of them in the portfolio.
Amazon is attractive to me because its other businesses aren’t threatened by AI. E-commerce and advertising will live, and may even accelerate. On top of these, you are getting the dominant cloud business.
AWS growth lagged other hyperscalers like Google and Microsoft for a long-time due to both its bigger base and lower exposure to OpenAI demand. This is changing now as Claude is growing way faster than ChatGPT. Thus, AWS growth accelerated as it is the preferred cloud provider of Anthropic.
SemiAnalysis predicted AWS growth to reach 24-25% by late 2025 as Anthropic’s gigawatt-scale AWS clusters were scheduled to come online:
We saw this played out as AWS growth reached 24% in Q4 2025. It’s set to further reaccelerate this year as Anthropic keeps growing fast and OpenAI is committed to expanding its AWS consumption by $100 billion.
These will certainly result in acceleration for AWS this year. UBS says AWS growth this year may reach 38%:
I think 38% is a long shot, but it’s possible that it can reach ~30% growth. Despite accelerating AWS growth, Amazon is trading at its lowest valuation at 27x forward earnings. Both the company and the stock are very well positioned now, so we’ll keep holding and increasing exposure over time.
Nebius is my neo-cloud pick.
I don’t like Coreweave’s customer concentration and its debt structures. Nebius is more straightforward; it’s starting from a smaller base, and valuation is cheaper.
They are targeting $7-$9 billion ARR this year with around 800 MW-1 GW connected power. They raised their contracted power guidance for the year from 2.5 GW to 3 GW.
They recently signed a $27 billion deal with Meta and raised $2 billion from Nvidia. According to the terms of the Nvidia investment, they’ll increase their Nvidia capacity to 5GW by 2030.
If they can deliver on this, they’ll be doing ~$50 billion ARR in 2030, based on the current GPU/hour rates. The whole company is now valued at $27 billion. The risk-reward here is still asymmetric.
We already own these two. What I am also considering recently is adding Oracle as well.
There has been a lot of discussion about Oracle’s buildout, and it’s undeniable that they are taking risks. The fact that most of their commitments came from OpenAI made it riskier, as OpenAI isn’t expected to become profitable before 2029.
However, OpenAI recently closed a $122 billion financing round, and we know that it’s eyeing an IPO later this year or early 2027. So, it’ll raise a substantial amount there as well. Thus, Oracle’s plans are now significantly derisked compared to what it was.
I am not buying any Oracle right now. I’ll look deeper at it, evaluate the risk and reward potential in more detail before I make any decisions. But it’s clearly more interesting now, and I am intrigued to dive deeper as it’s at just 19x forward earnings.
Last Words
There is not enough Claude.
This is what I feel whenever I use Claude Code or use it as my research assistant. And this is only the beginning, as AI looks very much like a General Purpose Technology in its infancy.
This means that we are looking at decades of capability enhancements, each unlocking new use cases, leading to innovation spillover and further increase in demand. Yes, adoption may look already decent, but we are just in the early innings of capability expansion and spillover. Demand will exceed anything we can imagine now.
Think about the transistor, another GPT. Today, we have more transistors in a single chip than anything its inventors could imagine in terms of total demand.
We have to see this broader picture, and don’t fall into believing that the capacity we are planning now may exceed demand in the long-term. It won’t.
We should position accordingly, but this doesn’t mean buy everything. Demand doesn’t guarantee execution. The key is getting your hands on the companies that are positioned to benefit from the demand without much execution overhang.
Amazon and Nebius are the two companies that fit this criteria for me, while Oracle is getting increasingly interesting as its biggest customer polished its financial position.
It’s important to pull yourself out of the daily discussions about AI’s contribution to growth, productivity, or efficiency. It’s too early to see developments at scale, but we are seeing the signals and anecdotes. Some things that weren’t possible before are now possible. This is what counts.
We are looking at the development of a new GPT. This is rare. And demand for every other GPT we have had so far has only increased over time.
It’ll be the same for AI.
There won’t be enough Claude, there won’t be enough cloud, there won’t be enough compute. We have to keep this in mind and position ourselves accordingly.
(If you have been thinking whether the “GPT” of “ChatGPT” was a reference to General Purpose Technology, the answer is no. It’s “General Pre-trained Transformer.”)
That’s all friends!
Thanks for reading Capitalist-Letters!
Please share your thoughts in the comments below.
👋🏽👋🏽See you in the next issue!