AI compute costs may rise 15-fold as demand outpaces supply growth

Author: Dwarkesh Patel

Compiled by: Deep潮 TechFlow

DeepChain Overview: When AI model capabilities catch up to top software engineers, the reasonable rental price for a single 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 10x revenue growth and its 3x compute growth. For investors, whoever secures compute first and trains the most efficient models will determine the next competitive landscape.

I want to try a minimalist approach: set aside two hours and quickly finish a blog post. I won’t have time to deeply explore many important subtopics, but if I don’t do this, many topics I’m interested in will remain stagnant and make no progress.

Today, I’d like to discuss the computing power landscape for AI labs over the coming years.

Anthropic's revenue grows tenfold annually. At this rate, Anthropic's revenue by the end of this year is likely to fall between $100 billion and $150 billion. If this trend continues, they would need to achieve $1 trillion in revenue by the end of next year. Of course, there is no hard guarantee that this trend will persist—it ultimately depends on AI capabilities. But assuming it does continue, what conditions must be true in such a world?

Laboratory computing power grows threefold annually. For a laboratory to achieve tenfold revenue growth while computing power increases by only 3 times, at least two of the following three conditions must occur simultaneously: one, improved laboratory profit margins; two, rising computing power prices; three, the laboratory allocates a larger proportion of its computing power to inference.

My understanding is that all three of these things are essentially happening: first, Anthropic’s profit margin has risen from 40% in 2025 to potentially over 80% for its Fable inference business this year (though marginal compute profit margins may differ, as I’ll discuss later); second, spot compute prices have increased by more than 40% since their February low this year, and the actual price increase faced by labs may be even higher (details to follow); third, according to Epoch data, OpenAI spent about a quarter of its compute budget on inference in 2024, and that proportion is now certainly closer to 50% or even higher.

Labs are actually reluctant to take the third path (allocating increasing amounts of compute to inference). There’s a joke that captures this perfectly: the value of inference revenue is to convince investors to give you more money so you can buy more compute and train larger models. If you devote most of your compute to inference, you’re essentially declaring that AI progress has stalled—because further investment in training is no longer worthwhile, and your business essentially becomes a cloud service provider. But the labs don’t see it that way—they believe the models they deploy today will look extremely outdated in a year, and they’re in this business now to build commercial use cases that support continued training of smarter models.

So there are two remaining paths: one, improving laboratory profit margins; two, increasing the cost of computing power. If the top one or two laboratories significantly outpace their competitors, the first path will be more pronounced. In market competition, profit margins depend on how much better you are than the next best alternative. For the first effect to dominate, profit margins must reach over 90% by the end of next year—which seems absurd to me. But I do believe it’s possible for AI laboratories’ revenues to continue growing at an astonishing rate.

So, explaining how the world could become real where revenue surges to $1 trillion by the end of next year, there’s one final effect left: a dramatic spike in compute pricing. As I mentioned, this has already begun. If we look at the tier of compute that labs truly need, the price increases are even more staggering—they clearly can’t rely on spot instances; they need to ensure the security of model weights and user data, and require sufficient scale to achieve good utilization and flexibility. Take a look at how wild this market has become: Google and Anthropic are renting compute from SpaceX. Reports indicate that Google spends $900 million per month to rent 110,000 GPUs, a mix of GB200s and GB300s. That works out to roughly twice the spot price of these GPUs. And even the current spot prices are already 40% higher than they were in February.

I want to emphasize a key conclusion: as AI models become more powerful, the value generated by the same amount of computing power increases significantly. If a software engineer with truly human-level capabilities could run on an H100-class device, based on current market salaries for software engineers, the annual rental cost of that H100 should exceed $250,000—15 times today’s spot price.

Of course, if 10 million software engineers suddenly appeared, the marginal value of engineers would naturally decline, so that H100 might not truly generate 15 times the current revenue. But I’m not sure this logic holds. If we apply this argument to humans rather than AI, it becomes the classic "lump of labor fallacy." Economists generally agree that high-skilled immigration does not suppress wages in the long run, because specialization and innovation increase the value of labor. Perhaps this time, the scale and speed of the labor supply shock are so extreme that this empirical rule no longer applies. But if we accept standard economic views on labor, the marginal value of labor (and the marginal price of computing power) should remain remarkably high.

What will happen in this world?

Leading models are increasingly extracting more value from compute, making it harder to catch up. If software engineering is fully automated by 2028 and compute costs are 15 times higher than today, it will be extremely difficult to compete with leading labs for compute resources without any revenue.

If you can train the best and most efficient model, your profit margin will far exceed today’s levels. This is the Alchian-Allen effect: when a fixed cost is added to two goods of different qualities, the higher-quality one becomes relatively more attractive, shifting demand toward it. When the H100’s hourly price rises to $20, using a weaker, less efficient model becomes extremely expensive and foolish—you’d need to burn more tokens and consume far more costly compute to achieve the same result. Therefore, anyone who trains a top-tier model that uses far less compute can charge a substantial premium. Since the underlying compute cost is already this high, why not spend a little more to use the most efficient model that delivers the same performance?

Many currently popular AI applications will be priced out of existence. AI is relatively cheap now—at least cheaper than human labor—partly because it still can’t do many things that top humans can. That will no longer be true one day. At that point, using GPUs to produce low-quality short-form content will become too costly to be worthwhile.

However, such predictions resemble historical misjudgments about scarcity. I’m thinking of the Simon-Ehrlich wager: Paul Ehrlich bet that, over the decade before 1990, a basket of commodities would increase in price rather than decrease. This bet became famous in popular economic discussions because it was said to disprove Ehrlich’s Malthusian worldview—demonstrating that he underestimated the role of market signals and human ingenuity in conserving scarce resources (though some analyses suggest Ehrlich would have won if the bet had been made over a different decade).

I suspect that commodities are not a suitable reference for computing power, as the supply elasticity of computing power is far lower than that of metal mining, and its ability to absorb large demand shocks or find alternatives is much weaker. The annual 3x growth in computing power is the product of three factors: a 1.4x increase from Moore’s Law, a 1.2x increase from new fabs being built (limited until at least 2030 by EUV equipment supply), and a 1.8x increase from AI capturing advanced-node wafer allocations from other devices (this factor will hit a wall around 2027, when AI’s share of N3 wafers rises from 60% to 86%). None of these three multipliers can be easily pushed higher, and the last one will peak within one or two years as wafer allocations become saturated.

What I’m saying is that at some point in the future, computing power will become cheaper again. One day, robots may be able to directly turn beach sand and underground copper mines into computers, at which point the price of computing power should fall back to levels close to the cost of raw materials and tools. What I’m discussing right now is just this current phase—where AI computing power grows only threefold per year, a rate that simply cannot keep up with the price inflation driven by AI’s rapidly increasing value.

By the way, Anthropic’s revenue is growing 10 times annually, while its compute resources are only growing threefold—this suggests that the model business exhibits extremely strong economies of scale. Logically, this makes sense—you pay a one-time cost to train a model, and then that cost can be amortized across all users (unlike human labor, where each instance requires full retraining). I don’t want to live in a world where intelligence has such strong economies of scale (because I fear concentration of power). But reality seems to be heading that way.