Author: Dwarkesh Patel

Translation: Shenchao TechFlow

Introduction by Shenchao: As AI model capabilities catch up with top software engineers, a reasonable rental price for an H100 will exceed $250,000 per year—15 times today's spot price. This is not a distant prediction, but a real trend driven by the structural tension between Anthropic's tenfold revenue growth and threefold growth in computing power. For investors, whoever first grabs computing power and trains the most efficient models will decide the next round of competition.

A Minimalist Experiment: A 2-Hour Power Calculation

I want to try a minimalist writing approach: set myself a 2-hour time limit to quickly write a blog post. Many important sub-issues I won't have time to delve into, but if I don't do this, many topics I'm interested in will remain shelved, with no progress.

Today, I want to discuss the computing power situation in AI laboratories over the next few years.

Anthropic Revenue 10 Times, Computing Power Only Increases 3 Times

Anthropic's revenue grows 10 times each year. Following this trend, by the end of this year, Anthropic's revenue will likely fall between $100 billion and $150 billion. If this trend continues, they will need to achieve $1 trillion in revenue by the end of next year. Of course, there is no hard reason to say this trend will definitely persist—it ultimately comes down to the capabilities of AI. But assuming it does sustain, what conditions must hold true in this world?

Laboratory computing power increases 3 times each year. If a laboratory wants to achieve revenue growth of 10 times while only seeing computing power grow 3 times, at least a few of the following three things must happen simultaneously: first, laboratory profit margins improve; second, the price of computing power increases; third, laboratories allocate a larger proportion of computing power for inference.

My understanding is that all three of these are basically happening: first, Anthropic’s profit margin has risen from 40% in 2025 to possibly over 80% in this year’s Fable inference business (though the marginal profit margin for computing power may differ, which I will address later); second, the spot price for computing power has risen by more than 40% since its low in February this year, while the actual price laboratories pay may have risen even more (which I will elaborate on later); third, based on Epoch's data, OpenAI's spending on computing power for inference was about a quarter in 2024, and this ratio is now much closer to 50%, if not higher.

Laboratories Reluctant to Take the Inference Route

Laboratories are actually reluctant to take the third route (allocating more computing power to inference). There’s a saying that captures this well: the significance of inference revenue is to convince investors to give you more money to buy more computing power and train larger models. If you allocate most of your computing power to inference, you’re in some sense declaring that AI advancement has stagnated—because continuing to invest in training has become unprofitable, your business fundamentally becomes a cloud service provider. Major laboratories do not see it this way—they believe that the models they are deploying now will look very outdated in a year, and they are currently in this business to accumulate commercial cases to support further training of smarter models.

Price Increases Are the Most Likely Answer

So that leaves two routes: first, laboratory profit margins increase; second, computing power becomes more expensive. If one or two leading laboratories are far ahead of their competitors, the first route will be more noticeable. In market competition, profit margins depend on how much better you are than the next best option. For the first effect to dominate, profit margins must exceed 90% by the end of next year. This seems quite absurd to me. But I do believe that it’s possible for AI laboratories' revenues to continue growing at an astonishing rate.

Thus, explaining how "revenue skyrockets to $1 trillion by the end of next year" will come true, there's one final effect left: a significant increase in computing prices. As I mentioned, this has already begun to happen. If we look at the computing power that laboratories genuinely require, the price increase will be even more staggering—they clearly cannot rely on spot instances; they need to ensure the safety of model weights and user data while also needing sufficient scale to achieve good utilization and flexibility. Let’s take a look at how crazy this market is: the price for computing power rented by Google and Anthropic from SpaceX. Reports indicate that Google spends $900 million per month to rent 110,000 GPUs, a mix of GB200 and GB300. This equates to roughly twice the spot price of these GPUs. And the current spot price itself is already 40% higher than in February.

I want to emphasize a core conclusion: as AI models become stronger, the value generated by the same amount of computing power also increases. If a truly human-level software engineer could run on H100-class devices, given the current market salary levels for software engineers, the annual rental cost of this H100 should exceed $250,000—15 times today’s spot price.

What Happens When Computing Power Becomes More Expensive?

Of course, if suddenly there are an additional 10 million software engineers, the marginal value of engineers will naturally decline, so that H100 might not truly generate 15 times its current revenue. But I’m actually uncertain if this logic holds. If we apply this argument to humans rather than AI, it represents the classic "labor force total fallacy." Economists generally believe that high-skilled immigration will not lower wages in the long term because specialization and innovation will increase the value of labor. Perhaps this time the shock in labor supply is too large and the speed too swift, leading to the failure of this empirical rule. But if we trust standard economic judgments on labor, then the marginal value of labor (and the marginal price of computing power) should remain at astonishingly high levels.

What will happen in this world?

Top models are increasingly able to extract value from computing power, making it harder to catch up. If by 2028 software engineering is automated and computing power prices are 15 times higher than now, then competing for computing power with top laboratories without any revenue will become extremely difficult.

If you can train the best and most efficient model, the profit margin you can charge will be far higher than today. This is the Alchian-Allen effect: when a fixed cost is layered onto two different quality goods, the higher quality one becomes relatively more cost-effective, thus shifting demand toward it. When the hourly rate for H100 rises to $20, using a weaker, less efficient model becomes extremely expensive and foolish—because you need to burn more tokens and run more expensive computing power to achieve the same results. Therefore, whoever can train a top model that uses less computing power can charge a significant premium. Since the underlying computing power already costs so much, why not spend a little more to use the most efficient model running on the same level of computing power?

Many currently popular AI applications will be priced out of the market. AI is currently relatively cheap, at least cheaper than human labor, partly because it still cannot perform many tasks that top humans can. One day, this will no longer be the case. By then, using GPUs for producing short video trash content will become prohibitively expensive.

Lessons from History: Predictions of Scarcity Can Also Be Wrong

However, these types of predictions share similarities with historical misjudgments regarding scarcity. I am reminded of the Simon-Ehrlich bet: Paul Ehrlich bet that the price of a basket of commodities would rise rather than fall over the decade leading up to 1990. This bet is quite famous in popular economics discussions because it is said to have disproven Ehrlich's Malthusian worldview—underestimating the role of market signals and human ingenuity in conserving scarce resources (though some analyses suggest that if this bet were made over a different decade, Ehrlich would have won).

How Low is the Elasticity of Computing Power Supply?

I suspect that the commodity reference frame is not suitable for computing power, because the elasticity of computing power supply is far lower than that of metal extraction and far weaker in absorbing large-scale demand shocks or finding alternatives. The 3 times annual growth in computing power is derived from the multiplication of the following numbers: 1.4 times from Moore's Law, 1.2 times from new wafer fab construction (limited by EUV equipment supply at least until 2030), and 1.8 times from AI capturing advanced process wafer quotas from other devices (this will hit a wall around 2027, when AI's share of N3 wafers will rise from 60% to 86%). Any of these three multipliers is difficult to increase further, and the last one may even peak within a year or two as wafer quotas become saturated.

The Endgame: Computing Power Will Eventually Become Cheap Again, But Not Now

I want to illustrate that at some point in the future, computing power will again become cheap. One day, robots will be able to directly turn beach silica and underground copper mines into computers, and by then computing power prices should return to levels close to the costs of raw materials and tools. What I am discussing now is just the current stage—AI computing power is only increasing 3 times per year, and this pace cannot keep up with the price-boosting effects of AI becoming increasingly valuable each year.

By the way, Anthropic’s revenue is growing 10 times per year, while computing power only increases 3 times, which indicates that the model business has strong economies of scale. Logically, this makes sense—when training a model, you incur a one-time learning cost that can then be spread across all users (unlike human labor, each instance has to be trained from scratch). I wouldn’t want to live in a world where intelligence has such strong economies of scale (because I worry about the concentration of power). But reality seems to be just that.