
What this covers
This is a solo video adaptation of a written essay by Dwarkesh Patel, built around what he treats as an accounting identity. Anthropic has grown revenue roughly 10x a year for three years, ending last year near nine billion dollars, and he expects 100 to 150 billion by the end of this year. He walks through the three ways the compute-revenue gap can close and checks each against current evidence. Anthropic's reported inference gross margins have risen from about 40 percent to over 80. OpenAI's inference share of compute has reportedly gone from a quarter in 2024 toward half. Spot prices are more than 40 percent above February's trough, and Google reportedly pays about 900 million dollars a month for 110,000 GPUs, at twice the spot rate.
The middle of the piece covers supply. He splits 3x annual compute growth into roughly 1.4x from Moore's Law, 1.2x from new fabs, and 1.8x from AI absorbing wafer allocation once used for phones and PCs. He argues none of these can speed up: AI's share of TSMC's leading-edge N3 capacity is already heading from 60 to 86 percent, and EUV machines from ASML limit fabs through 2030. On the demand side, a human-level software engineer running on one H100 would justify a rental price above 250,000 dollars a year, over 15 times today's spot price.
He then tests his own case. The lump of labor fallacy and high-skill immigration both suggest the marginal value of labor and compute stays high, and the Simon-Ehrlich bet warns against betting on scarcity. He judges the analogy probably wrong for compute, whose supply is less elastic. His consequences: consumer AI uses get priced out, and smaller rivals struggle to outbid the best labs.
Patel argues that if AI lab revenue keeps growing about 10x per year while compute grows only about 3x, the gap must be closed by higher lab margins or higher compute prices, and he expects compute prices to rise because AI's value per unit of compute is increasing.
- Lab revenue growing 10x per year while compute grows 3x per year forces margins or compute prices up, or a shift toward inference.
- Compute supply is inelastic because Moore's Law, fab construction, and EUV capacity are all constrained.
- Economic logic (the lump of labor fallacy, Alchian-Allen effect) implies the marginal value of compute stays high as AI improves.
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
If ten million extra software engineers appeared, their marginal value would fall, so the H100's revenue would not stay 15 times higher; but Patel is not sure this holds, because if the argument applies to people it would be the lump of labor fallacy, and standard economics suggests the marginal value of labor and compute should stay high.
“If we apply this argument to people instead of AIs, then this would be the classic lump of labor fallacy.”
Patel worries his analysis pattern-matches to past mistakes about scarcity, such as the Simon-Ehrlich bet, where Paul Ehrlich bet that commodity prices would rise over the decade before 1990 and lost, illustrating how market signals and ingenuity economize scarce inputs.
“I'm a bit worried that this kind of analysis honestly pattern matches a lot onto the ways that people in the past have been wrong about scarcity. I'm thinking, for example, of the famous Simon-Ehrlich bet.”
Fab construction will be bottlenecked through 2030 and possibly beyond by the production of ASML EUV lithography machines, a constraint Dylan Patel discussed in detail on a previous podcast.
“This process is ultimately gonna be bottlenecked up to 2030 and potentially even beyond by just building new ASML EUV machines. Dylan, when he was on the podcast a few months ago, talked about this in great detail.”
If a lab trains the most efficient model, it can charge much higher margins, because a weaker model burns more tokens on expensive compute to produce the same result, so a model that economizes scarce compute effectively creates more compute and raises its value.
“if you can train the best, most efficient model, then you'll be able to charge much higher margins than you can today. This is the Alchian-Allen effect in economics”
Compute capacity growth of about 3x per year is decomposed into roughly 1.4x from Moore's Law, 1.2x from building new fabs, and 1.8x from AI absorbing wafer allocation previously going to smartphones and PCs, and Patel believes none of these three factors can be much accelerated.
“1.4x of that is coming from Moore's Law. Far from increasing it, I think it'll be a miracle if we can just keep it going for a few more years. 1.2x is coming from building new fabs.”
High-skill immigration does not decrease wages in the long run because innovation and specialization increase the value of labor, but the labor supply shock from AI might be so large and fast that this general heuristic no longer applies.
“economists generally believe that high-skill immigration does not decrease wages in the long run because of how innovation and specialization increase the value of labor. Maybe this labor supply shock will be so big and so fast that we can't count on this general heuristic anymore.”
Moore's Law contributes about 1.4x of the 3x annual compute growth, and Patel thinks it will be a miracle if it can even be sustained for a few more years rather than accelerated.
“1.4x of that is coming from Moore's Law. Far from increasing it, I think it'll be a miracle if we can just keep it going for a few more years.”
The Simon-Ehrlich analogy is probably wrong for compute: other analysis suggests Ehrlich might have won had the bet been made in a different decade, and the supply of compute is much less elastic, less able to absorb demand shocks, and less substitutable than metal extraction.
“the supply of compute is much less elastic, much less capable of absorbing large demand shocks, and much less capable of being accommodated by using different substitutes than the extraction of different metals is”
Anthropic's revenue growing about 10x per year while compute grows only 3x per year illustrates how strong economies of scale are in the model business, because training is a one-time cost whose learned skills are shared across all users, unlike human labor that must be retrained for each instance.
“When you train a model, you just have to spend this one-time cost to learn all these different skills that then get to be shared across all your users. This is very unlike human labor”
Google is paying about 900 million dollars per month for 110,000 GPUs (a blend of GB200s and GB300s), a price that is 2x the spot price per hour for those GPUs, and that spot price is itself more than 40 percent higher than it was in February.
“Google, for example, is paying nine hundred million dollars a month for a hundred and ten thousand GPUs that are a blend of GB200s and GB300s. The price that Google is paying here is 2x the spot price per hour”
Spot prices for compute are more than 40 percent higher than they were in the February trough earlier this year, so the price of compute is already rising, which is the second channel for closing the revenue-compute gap.
“the spot prices for compute are more than forty percent higher than they were in the February trough that we had earlier this year”
Lab compute is growing only about 3x year over year, so a lab that keeps growing revenue 10x per year while compute grows 3x must resolve the gap through one of three channels: higher lab margins, higher compute prices, or a larger share of compute devoted to inference rather than training.
“For a lab to keep 10x'ing revenue year over year while compute only 3x's, one of the following three things needs to happen, or some combination of the three”
Labs prefer not to shift compute toward inference because heavy inference spending signals that AI progress has stalled and that the lab is becoming a cloud provider, which is a less compelling business than building AGI, so labs want to keep most compute on training the next model.
“if you're spending most of your compute on inference, then you're basically declaring that AI progress has stalled and you're just now in the business of being a cloud provider”
Lab gross margins could exceed 90 percent only if the leading model is so far ahead of competitors that users cannot replace it; Patel finds it hard to believe margins on intelligence would stay above 90 percent without being competed away.
“But it's just really wild for me to consider that the margins for something like intelligence will be greater than 90% and they don't get competed away at that level.”
As AI models get smarter, they will be able to monetize the same amount of compute more profitably; if a true human-level software engineer could run on an H100-equivalent, the H100 should rent for over 250,000 dollars a year at current software engineer prices, more than 15 times the current spot price.
“as AI models get smarter, they will be better able to monetize the same amount of compute. If a true human-level software engineer could run on an H100 equivalent, then at today's prices for software engineers, that H100 should rent for over 250K a year.”
AI's share of leading-edge N3 wafer capacity at TSMC is rising from 60 percent to 86 percent by the end of next year, so the wafer-reallocation channel will hit a wall because there is only so much leading-edge capacity to absorb.
“This is probably gonna hit a wall by the end of next year, when at the leading edge N3 nodes at TSMC, AI will have gone from 60% to 86%.”
Compute will eventually get cheap again once robots can convert silica sand and copper mines into chips, but in the current pre-singularity regime compute grows only 3x per year, which is not enough to offset how much more valuable AI is becoming.
“at some point in the future, compute will get cheap again. At some point, we'll just have robots that can convert shores of silica sand and mines of copper into new computer chips”
Frontier labs cannot simply buy spot instances for compute; they need enough scale for efficiency and flexibility, and compute that supports the security required for protecting their model weights and customer data, which makes the relevant compute tranche more expensive than spot prices.
“the tranche of compute that the frontier labs actually need to accumulate, because they can't just go out and buy a spot instance. They need to make sure that they get enough scale... and also that they have the kind of compute that lends itself to the security they need for their own weights”
According to Epoch, OpenAI spent just a quarter of its compute on inference in 2024, and that share is likely closer to 50% or higher now, meaning the inference-share channel is also operating.
“in 2024, according to Epoch, OpenAI was spending just a quarter of its compute on inference, and that number is likely closer to fifty percent, if not higher, now.”
Anthropic's inference gross margins reportedly rose from about 40 percent in the middle of last year to upwards of 80 percent now for its model, and this margin expansion is one of the ways the revenue-compute gap is already being closed.
“Anthropic's inference margins reportedly went from forty percent in the middle of last year to upwards of eighty percent now”
Many current popular AI applications will probably be priced out once AI can do tasks top humans do, because labs will pay more for tokens to automate AI research than consumers will pay for low-value AI content.
“a lot of current popular applications of AI will probably get priced out. The reason AI is relatively cheap right now is that AI just can't do a lot of things that top humans can do.”
As top labs get better at monetizing compute and the cost of compute rises, it becomes harder for other players to compete, because they must bid for the same resource against a buyer who can use it more productively.
“as the top labs get better and better at monetizing compute, and the cost of compute increases, it becomes harder for anybody else to compete against them”
Patel does not see how compute can keep scaling 3x year over year for the next few years, let alone beyond that, given that each of its three drivers faces a hard limit.
“So I don't know how we even continue to do 3x compute scaling year over year for the next few years, much less go beyond that.”
Anthropic's revenue has grown roughly 10x year over year for the last three consecutive years, ending last year at about nine billion dollars, and is likely to grow about 10x again, reaching somewhere between 100 and 150 billion dollars by the end of this year.
“For the last three consecutive years, Anthropic's revenue has 10x'd year over year, and it's likely to do so again this year. They ended last year with nine billion in revenue.”
For Anthropic's revenue to keep growing 10x per year, it would need to reach about one trillion dollars in revenue by the end of next year, which is a very wild conclusion that depends on whether AI capabilities become useful enough.
“Anthropic would need to make one trillion dollars in revenue by the end of next year. Of course, there's no deep reason why this has to be true.”
Patel wishes the world did not have such strong economies of scale for intelligence because he is worried about power concentration, but he thinks the economics show that these economies of scale exist.
“I wish we didn't live in a world with such strong economies of scale for intelligence, because I'm worried about power concentration, but it seems we do.”
The Mercury banking platform's AI feature, Command, automatically categorizes transactions and reviews vendor and memo context for bookkeeping.
“I have Command, Mercury's built-in AI, take a stab at all of them at once.”