Compute is all
that matters

One prompt away

I quit my job as a medical doctor in the middle of 2023. Seeing how fast AI was moving, I realised this is going to be the most important technology of our lifetime. I hadn't realised how big it was going to be.

The first thing I built was a relatively complicated AI product at the time. It took three months of talking to ChatGPT, copy-pasting code and trying to understand things, and I had to work with a couple of engineers.

I built my own startup, I became an engineer, and now I'm a founding engineer at an AI startup. I'm now certain that there is no task that isn't one prompt away, as long as you have enough compute.

For each pull request, the smaller unit of work an engineer produces, I now get several AI models working on the same thing from code implementation to generating on demand videos about it. This was unthinkable two years ago.

While I contribute as much as 10 engineers at work, my old startup runs on agents, very profitably, with not much human involvement. I also made my own mobile game in a couple of weeks of free time; according to Claude, it would have taken a small games studio a couple of years and a couple of million dollars.

01

Why it's not a bubble

Right now there's been a big sell-off in the Nasdaq over the last week through the earnings reports. People believe capital expenditure may be limited, and that the companies might face the realities of the economy. Interesting contrast is looking at companies spending and what they are betting on. They are reducing the free cash flow, laying off people and trying to get any money they can to spend more on compute.

Contrary to the bearish belief, markets are not pricing a lower demand. What we see is a temporary cash problem, mostly because companies spend more while revenues actually grow. This volatility is a good period to spend more on compute.

This turbulence shows the stark reality of the world. Big companies are willing to take a big hit, as we've seen with Google, because they know that spending on compute is the only thing that matters in the long term. Memory and compute are going to get even more expensive from now on. The winners of the future are going to be the people who own these chips, and the people who are able to do work on them. That's the pure reality.

Sticky inflation and owning the hardware

The world has been experiencing what inflation means since COVID. Everyone hoped that inflation would go down. It's not going down. The part of the market that is not related to AI and compute is bearish, while AI growth seems to be the only thing that matters. The investment thesis has been very strong with AI compute and inference, and it is going to be even stronger regardless of how the market turns.

In this scenario, you want to own hardware. This can be through physical ownership or long term commitment but owning the hardware is owning the means of work. In this environment, investing in AI compute infrastructure is the only thing that will make money. And being in this chain will be even more important as what models are able to do gets increasingly better.

This asset class is becoming the only thing that holds, and will hold, value for many people.

On the personal level I'm often cautious about committing money and stayed away from any speculative investments even when they made sense. Now the only thing I'm trying to do is get credit and buy more compute. The hardware you buy appreciates because demand is incredibly high. You can turn the equipment you buy into money by running AI models on it while the underlying asset appreciates in value.

02

There is no cap on the demand

In 2023, I was spending twenty dollars on ChatGPT subscription, and one subscription was enough. There wasn't much of a limit to how much you could do. Not just because OpenAI was generous, but also the value you could get from a model was not that much. You were on one chat, trying to do one thing, in slow mode.

Then I started using Cursor, which had usage limits, and I moved to the forty-dollar plan. When Codex came out I was paying a hundred dollars, around summer 2025, and that was magic. Then the two-hundred-dollar Claude Code plan seemed very attractive, and at the start of this year I was using two Claude subscriptions, around four hundred dollars.

Right now, between my job and personal projects, I spend $3,500 a month on AI subscriptions. If you include the GPUs I use from Modal and other cloud services, it is well beyond $4,000. I spend a couple of billion tokens a day across multiple Codex and Claude accounts.

When I did the jump from 40 dollars to 100 dollars, it was a big decision. But 400 to 4000 was not. The models are now so good, and the value that they bring is so high.

The price of intelligence has been reduced, but the money we put on models increased, because what models can achieve increased much more significantly.

Models being much better means they spend more tokens, and even if they are more token efficient, they will spend more time on the task. People's spending is going up.

There is no world where we will not consume the endless compute. It is just the natural way things are going to progress.

It's becoming a norm to pay for higher tiers as an individual.

The demand that is still locked

AI token spending is mostly limited to computer tasks so far. If we take this example and apply it to other sectors, it's very easy to see how much demand is still locked.

Computer and maths jobs are 3.4% of US workers and 5.8% of the $10.8tn US wage bill. Management, healthcare, business and finance, office work and sales each pay more in wages. Yet computer and maths tasks are 36% of Claude conversations and a little less than half of its API traffic (Anthropic Economic Index, September 2025). Across all models on OpenRouter, programming went from about 11% of tokens in early 2025 to more than half by the end of the year (OpenRouter, State of AI). If every kind of work used AI as heavily, for each dollar of wages, as computer work does now, usage would be about six times what it is today, before coding grows any further.

02

Prohibitively expensive

As a species, we've been resource constrained all of our lives. Maybe we look back at how people lived a hundred or two hundred years ago and think, how did they do it? But what we did was create more resources to make the world bigger.

Going through medical school the whole training was really built on resource constraint thinking. When you think about a patient, you always consider doing simpler things, not always because that's the best for the patient, but because you're resource constrained. You can't get ten different consultants to examine a patient, ten different diagnostic methods to collect data about them, or continuous monitoring by people who are top of their field.

Even the referral system: you go to your GP, you wait for months for a specialist, and you're on this cycle. God forbid you have a chronic condition. Then you will see how limited modern medicine actually is.

Modern medicine is quite good at making sure you don't die from simple things like microbes, because we have antibiotics, or from very traumatic accidents, because we have surgery. But most of the chronic conditions, we just manage them, quote unquote, and often very poorly.

Some things are still financially expensive, so we are not doing them. Very fast inference, for one. Because of this constrained world of compute, providers have to batch requests across thousands of users, so you don't get an instant response but a middle speed, whatever is best for the provider. OpenAI started offering fast and then ultra-fast serving, and Anthropic offers fast serving, but for many people it's prohibitively expensive.

When you look at how OpenAI went after the Navier–Stokes equation, they used swarms of agents. It wasn't easy. Once compute is abundant, and there is a clear story around spending that compute resulting in a very positive result, you would be happy to pay for that.

And then there's robotics. It's very early days, but you can very easily imagine billions of robots, a couple of robots per person, doing industrial work and household work, and how much compute they will require to operate.

Right now there are lots of very expensive services people buy: a very expensive surgeon, a very expensive personal service, transport, whatever makes their life better and more convenient. In the future all those tasks will be done by AI, physical or software, or AI will accelerate them, and you will just pay more money.

Dario Amodei expects that when AI makes a change to a molecule worth tens of millions of dollars, it will be paid for the value of the task, not only by the token. If it was a human that created that drug and licensed it to a pharma company, the human or the team would get the licence money. So getting the compute is so important. It's like grabbing land, and it's not just about becoming another inference company, but about the possibilities of having the compute.

03

Human attention is not the limiter anymore

We are moving away from one API call, one Responses API request, to agents. Karpathy called this the decade of agents. We had the year of agents, but the way agents worked was still a human prompting the agent, with some going back and forth.

Then we had the goal feature from Codex, where you define the end goal and agents work toward it, maybe together with other agents. What's happening right now is that you don't want a human to slow down the agents. So you have managed agent runtimes, or super-agents, as the DeepSeek Harness team calls them, and you just run agent swarms.

We have cracked the point where human intelligence, human attention, is not the limiter anymore for agents to do work. Right now it's all about how much compute you can get. OpenAI put 10,000 agents on one proposed proof. Ten thousand today, hundreds of thousands tomorrow.

04

When the scaling is going to stop?

I often wondered where the scaling is going to stop, and where the demand is going to stop. Models were helpful but the economics around serving models were not always very clear.

What we are seeing is models that are able to improve themselves and the next generation of models, while making inference much more economic. What about demand?

Up to two months ago, I wasn't this sure how much this can scale. Then I was able to experience one Codex thread managing tens of other threads, and it became much more clear: once you have a good system in place, what models can achieve is significantly different.

05

There is literally not enough compute

Try to get a cluster of nodes right now: you will see how painful that is, and how expensive that is. There is literally not enough.

Modal, one of the best inference providers, just announced it is increasing its GPU prices. In the last three to six months I've been constantly increasing my spend on Modal, but it's coming to a point where it's going to be more expensive, and more prohibitive for some.

Demand exceeds supply, in the companies' own words

• "A compute-constrained environment where customer demand exceeds available compute."

  Anthropic, amended S-1, 2 Nov 2026

• "Customer demand continues to exceed available supply."

  Amy Hood, Microsoft CFO, 28 Oct 2026

• Demand for TSMC's chips still exceeds supply, and the gap is still large.

  C.C. Wei, TSMC chief executive, 15 Oct 2026 (paraphrase)

As the big labs grow, they find that the compute required by training is also increasing rapidly. There is always a big fight over whether to use the chips to serve customers or for research. If there was 10× more compute at OpenAI, the researchers would be more than happy to take it.

There's the aspect of getting new compute, the chips, and then the other aspect of getting those chips online, and there is a big limiter around that.

Compute islands

Being able to get the compute is very important, and the moment you're able to sort out the infra, the demand is there. Distributed training is possible, and everyone is trying to scrape whatever compute they can find.

It's not just GPU

And it's not just GPU. Multi-agent workflows expose the limitations we have with CPU and memory.

Something very interesting has happened over the last two months: because of how much AI I'm using, my computer has started to be insufficient, and it's a MacBook with an M4 Pro and 24GB of RAM. It ran out of memory. If the number of agent tokens I use goes up tenfold, I need a machine ten times as big.

It's not just NVIDIA

Apple shows how to connect multiple Mac Studios and serve state-of-the-art open-source models across them, which is what I'm going to do. I've ordered a £24,000 Mac Studio M5 Ultra with 512GB of memory, because of how much RAM I'm going to need if my agent usage keeps increasing at this speed.

Sovereign compute

Right now, most of this build is happening in the US. But everyone is realising that building the compute, and having that compute sovereign, is way more important. When a US government directive on foreign access made Anthropic switch off its Fable model for a while, it told everyone that as much as there is a build in the US, that build is going to happen anywhere and everywhere else.

For now the chips still go to the US: it took 78% of new AI chips in the first half of 2026 (Epoch AI).

06

There is no moat

We talk about moat, and where the moat lies at the moment when looking at the market. We've seen in the last couple of weeks how closed-source software is reverse engineered, and how games are getting mashed and merged with AI. Years of "human" effort turned into one prompt.

If you have spent all the years gaining a stronghold, you might think that you have distribution. But if someone comes in with unlimited compute to create the 10× better version of your product, and creates the distribution, the go-to-market engine, with models, what's left?

I've gone through this. It was not that easy to create good software with AI. You were proud of what you've created and people would appreciate the craft. Right now, everything is generated by AI without too much of a bug, and once you have a good system in place, it recursively improves itself.

There is no moat. No moat for model providers, no moat for the application layer, no moat in the middle. The only reality is compute and how efficiently you're able to use that compute.

Anthropic says much the same in its own filing. As AI models are increasingly used to develop other AI models, it warns, the know-how to build frontier models may spread, including through distillation. Competition "could increasingly depend on access to large-scale compute", and compute on economical terms could become "the primary constraint" (amended S-1).

Open-weight models already trail the frontier by about four months (Epoch AI, May 2026). On OpenRouter, closed models' share of spending fell from 89% to 73% between January 2025 and September 2026, and spending on Moonshot, Z.ai and DeepSeek models each grew at least tenfold between January and August 2026 (OpenRouter). Revolut serves about 2,000 engineers with the open GLM 5.3 model (07).

07

In both cases, spending goes to compute

And for enterprises: right now, someone is running an application and you are paying money to use part of that application. Then you realise you should actually run it yourself, because you have enough users in-house, you have sensitive data you can't risk, and open-source models are good enough for you. It's like deciding to run your application in the cloud: you use a cloud service to make life easier rather than going on-prem, and that's what's going to happen. You pay Modal, you pay Nebius, and that means the GPU pricing is never going to go down.

The way people's spending is going up is not going to hold for very long. AI product companies are going to have an awakening, both in their serving and unit economics and in how lean they can stay with smaller engineering teams. Then they will have two choices: set up their own infrastructure to serve local models, which Revolut seems to be doing, or spend more. In both cases, spending goes to compute.

08

There are very different ways of being part of this

I worked at a medical device company that was creating hardware, and I currently work at an AI company that is selling agents. What I'm realising is that AI agents are going to be the future of workers. But selling agents is going to be a sales business. As we've seen with Harvey, the legal AI company, and with Revolut, they're going to utilise the compute, and everyone is going to utilise the compute. Similar to how you have the IT team in a hospital, you are going to have the inference and AI team in the hospital, but they're not necessarily going to have their chips in the hospital. They're going to use chips somewhere, whether that's OpenAI, Anthropic or a smaller provider.

I think the engineering effort is getting less and less important in building agents and all these AI product companies. It's going to be sales. And when you look at OpenAI and Anthropic, although they are great and they are training better models, the only moat they have is compute, and compute seems to be the only thing that matters.

When the music stops

When the music stops. When RSI becomes possible, when the agent swarms are doing a thousand years of human tasks in seconds, where do you want to find yourself?

When you look at London house prices in the last ten years the value of the property significantly reduced once you adjust for inflation. You're putting millions into a property that goes down in value, while you probably spend lots of money to maintain it.

Do you want to be the owner of chips, the worker of the future? Or do you want to hold on to assets that are not moving, and that are not going to mean anything when AI agents, a physical swarm of robots, are able to build a house in days?

At the end of the day, compute is going to be all that matters, because there is no moat.