## Background
Published by Dwarkesh Patel on RSS / web on August 3, 2026, this audio recording presents a read-through of his previously published post. The description highlights that the episode is sponsored by Mercury and its AI platform, Command, which automates bookkeeping and categorizes transactions.
## TL;DR
The episode explores the widening gap between exponential AI revenue growth (10x annually) and the physical limitations of frontier compute expansion (3x annually). To reconcile this imbalance, compute prices are projected to surge dramatically, particularly as models achieve human-level software engineering capabilities that unlock immense economic value. Driven by hard physical constraints in chip fabrication and power allocation, high-value AI labor will price out casual consumer applications during this 'pre-singularity' transition. Ultimately, this dynamic creates unprecedented economies of scale that could concentrate immense power among a few elite AI labs.
## Atmosphere & style
The audio features an analytical, dense, and intellectually rigorous delivery typical of a solo post-reading. Dwarkesh presents his arguments with structured, systematic clarity, bridging economic theory with hard hardware engineering limits. The vibe feels pragmatic yet monumental—contemplating the massive economic disruptions and power concentrations of the pre-singularity regime in a grounded, matter-of-fact tone.
## Key takeaways
- `00:00` Anthropic's exponential revenue growth ($10x YoY) is colliding with a much slower growth rate in frontier compute supply ($3x YoY).
- `00:00` To bridge the gap between revenue and compute growth, labs must raise profit margins, increase compute prices, or shift compute from training to inference.
- `00:00` Labs are resistant to shifting most compute to inference because training next-generation models is essential to their identity as AGI builders.
- `05:00` Surging demand for secure, dedicated compute is already driving prices higher, as seen in premium rents paid for SpaceX-hosted GPUs.
- `05:00` As AI models reach human-level software engineering abilities, the economic value of GPUs will surge, potentially reaching over $250,000 per H100 annually.
- `05:00` High compute costs will create entry barriers, pricing out smaller competitors and low-value consumer applications in favor of high-value tasks.
- `05:00` The Alchian-Allen effect will make running weaker models economically inefficient when compute inputs are expensive.
- `05:00` Physical bottlenecks in Moore's Law, fab construction, and finite silicon reallocation strictly cap annual compute supply growth at roughly $3x.
- `10:00` In the pre-singularity regime, hardware constraints keep compute expensive despite the high scalability and zero marginal replication cost of trained AI models.
Over the past three consecutive years, Anthropic’s revenue has grown $10\times$ year-over-year. Ending last year at \$9 billion in revenue, they are on track to hit somewhere between \$100 billion and \$150 billion by the end of this year. For this exponential trend to continue, Anthropic would need to reach an astounding \$1 trillion in revenue by the end of next year. While there is no fundamental rule dictating that this must happen, it remains a question of AI capabilities—specifically, whether AI becomes useful enough fast enough. However, if this revenue trend holds true, it runs directly into a major constraint: frontier lab compute is only growing $3\times$ year-over-year.
To bridge the gap between $10\times$ revenue growth and $3\times$ compute growth, one (or a combination) of three levers must yield:
1. Lab profit margins must increase.
2. The price of compute must increase.
3. The percentage of compute dedicated to inference rather than training must increase.
Currently, all three dynamics are already occurring to some extent. According to reports, Anthropic's inference margins rose from 40% in mid-2023 to upwards of 80%. Spot prices for compute have jumped over 40% from their February trough. Meanwhile, Epoch data shows that OpenAI allocated only 25% of its compute to inference in 2024, a figure that has likely climbed to 50% or higher today.
However, labs are fundamentally reluctant to divert the majority of their compute toward inference. From their perspective, inference revenue exists primarily to convince investors to fund the next, larger training run. Shift too heavily into serving existing models, and a lab is essentially declaring that AI progress has stalled—effectively reducing itself to a standard cloud provider rather than an AGI builder. Because labs firmly believe that next year's models will make today's look "extremely shitty," they must continue reserving most of their compute for training and experimentation.
- **Anthropic Revenue Trajectory:** From \$9 billion at the end of last year to an estimated \$100 billion–\$150 billion this year, targeting \$1 trillion next year if trends hold.
- **Compute Growth Rate:** Frontier lab compute scales at only $3\times$ year-over-year.
- **Inference Shift:** OpenAI's inference allocation moved from 25% in 2024 to 50%+ today.
---
`03:00` Escalating Compute Prices and the Value of AI Labor
If labs refuse to give up their training compute to serve inference, only two primary escape valves remain: lab margins must increase, or the price of compute must rise. While lab margins pushing past 90% is possible, standard economic logic dictates that margins that high require a model to be so dramatically superior that customers cannot substitute it. It seems far more plausible that the primary adjustment mechanism will be a surge in the price of compute itself.
This price pressure is most acute for the dedicated, high-security compute tranches frontier labs require. Spot market instances are insufficient for labs that need massive unified scale for efficiency alongside tight security guarantees for proprietary weights and user data. A prime example is the compute Google and Anthropic rent from SpaceX: Google pays \$2 million a month for 110 GPUs (a mix of Nvidia GB200s and GB300s). That equates to double the hourly spot price, which itself sits 40% higher than its February baseline.
> "As AI models get smarter, they'll be better able to monetize the same amount of compute."
As AI models reach true human-level software engineering capabilities, their ability to extract economic value from hardware will skyrocket:
* An $H100$ GPU capable of running a full software engineer would be worth over **\$250,000 a year** based on human software engineering compensation.
* This represents more than a **$15\times$ increase** over the current annual spot rental price of an $H100$.
* This calculation doesn't even factor in the reality that AI software engineers can work nights, weekends, and holidays without stopping.
Standard economic theory (avoiding the "lump of labor fallacy") suggests that introducing millions of AI workers will not necessarily collapse the marginal value of labor or compute. Historical immigration patterns show that labor supply shocks often increase overall productivity through specialization and innovation, keeping the marginal value of compute astonishingly high.
---
`05:00` Economic Ripples: Alchian-Allen Effect, Pricing Out Apps, and Physical Bottlenecks
As top labs maximize compute efficiency, several major market shifts will occur:
1. **Higher Barriers to Entry:** Entrenched labs that monetize compute effectively will bid up hardware costs, pricing out smaller competitors who cannot squeeze as much value from the same silicon.
2. **The Alchian-Allen Effect:** If renting an $H100$ costs \$20 an hour, utilizing a weaker, less efficient model becomes economic folly because it burns expensive compute tokens to produce inferior results. Premium models that economize scarce hardware inputs will command massive price premiums.
3. **Consumer Application Displacement:** Casual AI applications will be priced out. Labs will prioritize selling compute tokens to automate high-value tasks like internal AI research over lower-value consumer uses like "AI slop talk."
Skeptics might view this narrative as a repeat of Malthusian scarcity traps—like Paul Ehrlich's famous commodity price bet against Julian Simon—where critics failed to predict human innovation expanding supply. However, compute supply is far less elastic than raw metal extraction and cannot easily absorb massive demand shocks. The $3\times$ annual compute growth ceiling is bound by hard physical constraints across its three drivers:
* **Moore's Law (1.4$\times$):** Keeping pace even at this rate for a few more years will be an engineering miracle.
* **Fab Construction (1.2$\times$):** Global chip fabrication capacity is bottlenecked through 2030 and beyond by the production limits of ASML EUV lithography machines.
* **Silicon Allocation Shift (1.8$\times$):** AI has aggressively absorbed wafer allocations previously dedicated to smartphones and PCs. On TSMC's $N3$ leading-edge nodes, AI wafer share is expanding from 60% to 86%, leaving almost no remaining consumer capacity to cannibalize.
In a brief operational tangent, the host walks through his end-of-month financial workflow, illustrating how Mercury's AI platform, Command, automates bookkeeping tasks. Managing payments across diverse contractors—such as tutors, researchers, and videographers—typically creates hours of administrative work. Command evaluates incoming transactions, assigns categories, and provides plain-text rationales based on contextual clues like notes, memos, and the team member making the purchase. The host reviews and approves these categorization calls before Mercury automatically syncs the data directly into QuickBooks.
---
`10:00` The Pre-Singularity Regime and Scale Economies
In a post-singularity future, compute will eventually become cheap again once autonomous robotic systems can directly mine copper and refine silica sand into silicon chips, reducing costs down to raw materials and equipment processing. However, during the current "pre-singularity regime," hardware growth remains capped at $3\times$ annually while model capabilities scale far faster.
This structural divide highlights the massive economies of scale inherent to the frontier model business. Unlike human employees—who must each be trained individually from scratch—an AI model represents a one-time training investment. Once a skill set is acquired, it can be instantaneously duplicated and served across millions of end users at scale. While this dynamic creates immense commercial leverage, it raises serious concerns regarding the extreme concentration of power among a handful of leading AI labs.
## Notable quotes
> `03:00` As AI models get smarter, they'll be better able to monetize the same amount of compute.
Dwarkesh Patel argues that as AI revenue and capabilities expand exponentially, physical supply constraints on chip fabrication will fail to keep pace, driving compute rental prices up by an order of magnitude as smarter models extract far higher economic value per chip.
The Widening Gap Between Revenue and Compute (0:00)
Anthropic's revenue has grown approximately 10x year-over-year for three consecutive years, rising from $9 billion at the end of last year toward an estimated $100 billion to $150 billion this year, and potentially $1 trillion next year if the trend holds. In contrast, frontier lab compute capacity is expanding at only 3x annually. Patel identifies three mechanisms that can bridge this gap: higher profit margins, rising compute prices, or shifting hardware allocation away from model training toward inference.
While Anthropic's inference margins reportedly rose from 40% in mid-2023 to over 80%, and Epoch data shows OpenAI expanded inference compute from 25% in 2024 to over 50%, labs remain deeply reluctant to abandon training. In Patel's view, frontier labs view inference revenue primarily as a tool to finance subsequent training runs; shifting too heavily to serving current models would signal that AI progress has stalled and reduce frontier labs to standard cloud service providers.
AI Labor Valuation and Compute Price Surges (3:00)
Because labs refuse to sacrifice training runs, Patel expects compute price surges to serve as the primary economic relief valve. Dedicated, high-security compute already commands high premiums; Google pays SpaceX $2 million monthly for 110 Nvidia GB200 and GB300 GPUs, roughly double the spot price, which itself sits 40% above its February trough.
Patel argues that as models reach human-level software engineering proficiency, their ability to monetize silicon will increase dramatically. An Nvidia H100 GPU capable of replacing a full-time human software engineer would be worth more than $250,000 annually—over 15 times its current annual spot rental rate—even before accounting for uninterrupted continuous operation. Patel rejects the lump of labor fallacy, suggesting that historical labor supply shocks demonstrate that massive infusions of automated capability expand overall economic productivity while keeping the marginal value of compute high.
Market Repercussions and Application Displacement (5:00)
Elevated compute costs are expected to trigger the Alchian-Allen effect across the AI sector. If renting an H100 climbs to rates such as $20 per hour, operating weaker or less efficient models becomes economically irrational because they consume expensive hardware inputs to produce inferior results. Premium, highly capable models that maximize output value per token will consequently command steep pricing power, raising entry barriers against smaller competitors.
Simultaneously, rising hardware costs will crowd out low-margin consumer applications. Patel anticipates that frontier labs will increasingly ration compute away from casual conversational use or novelty consumer software to prioritize high-return industrial workflows and internal automated AI research.
Physical Bottlenecks in the Pre-Singularity Regime (10:00)
Patel contends that compute supply cannot easily scale to match demand due to hard physical limits across its three growth drivers: Moore's Law progression (1.4x), fab expansion constrained through 2030 by ASML EUV lithography tool availability (1.2x), and wafer reallocation (1.8x). AI's share of TSMC's N3 leading-edge nodes is already expanding from 60% to 86%, exhausting consumer hardware capacity that can be cannibalized.
While Patel notes that a hypothetical post-singularity economy with fully automated robotic manufacturing could eventually drop compute costs to the price of raw materials, the current pre-singularity regime remains capped at 3x annual compute growth. Because frontier models require massive upfront training costs but can be duplicated instantaneously once trained, these bottlenecks reinforce immense economies of scale and concentrate market power among a few leading labs.
Patel's thesis relies on the premise that AI capabilities will continue to advance rapidly enough to sustain human-labor parity pricing, while physical fab constraints strictly cap hardware supply. If model capability gains stall, or if architectural breakthroughs sharply reduce compute requirements, the projected upward pressure on compute rental markets could fail to materialize.
[3:00] As AI models get smarter, they'll be better able to monetize the same amount of compute.
Original transcript
[0:00] Today, I want to talk about what the compute situation for the labs will look like over the next few years. For the last three consecutive years, Anthropic's revenue has 10x year over year, and it's likely to do so again this year. So they ended last year with $9 billion in revenue. I think they'll probably end this year with somewhere between $100 billion to $150 billion in revenue. Now, for this trend to continue, Anthropic would need to make $1 trillion in revenue by the end of next year. Of course, there's no deep reason why this has to be true. It's a very wild conclusion, and it's ultimately a question of AI capabilities. Does AI get that useful by the end of next year? But suppose the trend does continue. Well, I want to think through what happens in that world. Now, the other big trend in AI is that lab compute only 3x's year over year. 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 needs to happen. One, lab margins have to increase. Two, the price of compute has to increase. Or three, the percentage of compute that labs spent on inference rather than training has to increase. My understanding is that basically all three of these things are already happening. With regards to the margins Anthropics inference margins reportedly went from 40 in the middle of last year to upwards of 80 now available With regards to compute the spot prices for compute are more than 40 higher than they were in the February trough that we had earlier this year. And with regards to the share of compute that goes to trading versus inference, in 2024, according to EPOC, OpenAI was spending just a quarter of its compute on inference, and that number is likely closer to 50%, if not higher now. Now, labs have deferred not to do this final thing of increasing the share of compute they spend on inference. The way the labs see the world, the whole point of inference revenue is to help convince investors to give you more money in order to train the next bigger, better model. And if you're spending most of your compute on inference, you're basically declaring that AI progress has stalled and you're just now in the business of being a cloud provider. Now, this is a less compelling business than building AGI. And so the labs do not want to be in this business, nor do they think they're in this world. They think that within a year, they'll have built models that make the current ones look extremely shitty. But they need to invest a lot of their compute, the majority of their compute, into doing the training and experiments that are necessary to build the next model. So that leaves only two options for how you can get out of this gap between the fact that lab compute only increases 3x year over year but revenue increases 10x Either the lab margins have to increase so that they get this surplus or the price of compute has to increase so that everybody in the stack below the lab gets a surplus. It's not clear to me which world we end up in. Do we end up in a world where we go from 80% margins for some of the top models to greater than 90% margins if the lab margin effect dominates? Well, that would require the leading model to be so far ahead of the competition because the nature of margins, why they exist in a market economy, is that the thing you are serving is so much better than what somebody else could go get and replace you on the market. 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 competitive at that level. So that leaves only one other possibility of this escape valve between these two trends, which is that the price of compute has to increase. As I mentioned, this is already starting to happen. And the effect is even stronger when you look at 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 to get really good efficiency and flexibility and also that they have the kind of compute that lends itself to the security they need for their own weights and for their customers information So I think a relevant case study here is to look at the compute that Google and Anthropic are renting from SpaceX Google for example is paying million a month for 110 GPUs that are a blend of GB200s and GB300s. The price that Google is paying here is 2x the spot price per hour for those GPUs. And that spot price itself is more than 40% higher than it would have been in February. I want to emphasize the key conclusion here, that as AI models get smarter, they'll 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. That's over 15x the current spot price for an H100. And this is not even accounting for the fact that your AI can work nights and weekends. Of course, you might expect that if we had 10 million extra software engineers suddenly appear in the economy, the marginal value of a software engineer would decrease and thus the revenue that that H100 would be able to generate would not be 15x higher than it is right now. But I actually don't know if this is true. If we apply this argument to people instead of AIs, then this would be the classic lump of labor fallacy. For example, economists generally believe that high school immigration does not decrease
[5:00] 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. But if you believe what standard economics says, then the marginal value of labor and thus the marginal value of compute should stay astonishingly high. So let's think about what changes in such a world. One of the things that would happen is that 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 because they have to bid for this resource against somebody who is basically able to make better use of it. Another thing that will happen, and I think this is actually the most interesting implication of this whole thought exercise, is that 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 Alkin-Allen effect in economics, and what it's basically saying is that if it costs $20 an hour to rent an H-100, then it would be extremely stupid to use a weaker, less efficient model, because it's gonna burn more tokens on your expensive compute to get the exact same result. So labs will be able to charge a much larger premium if they can train a model that better economizes this scarce input Basically if you have a model that can get the same result by using less compute then you in some sense created more compute and the value of compute is going to increase Another thing that will happen is that 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. But this, at some point, will no longer be the case. And at that point, And Google or Anthropic or OpenAI will be willing to pay more for the tokens to automate AI research than you or I will be willing to pay to make more AI slop talk. 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, for example, thinking of the famous Simon Ehrlich bet. Paul Ehrlich was this famous doomer about population growth, and he made this bet that a basket of commodities would increase in price rather than decrease in the decade preceding 1990. And this is a very famous bet because it's supposed to illustrate how Ehrlich's Malthusian worldview was wrong and how he did not anticipate the way in which market signals and human ingenuity can find better ways to economize scarce inputs. I guessing that the analogy to this bet is probably wrong Other analysis has shown that if that bet had been made in a different decade Ehrlich might well have won But more generally I think the supply of compute is much less elastic and 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. To illustrate why I think this 3x in compute capacity euro over euro is hard to budge or potentially even sustain, is that I don't see how any of the three elements that constitute that 3x can be much accelerated. So 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. This process is ultimately going to 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. And 1.8X comes from the fact that AI is absorbing a lot of wafer allocation that was previously going to smartphones and PCs. This is probably going to hit a wall by the end of next year when at the leading edge N3 nodes at TF2C, AI will have gone from 60 to 86 At some point you have just absorbed all leading edge wafer capacity for AI and you can keep increasing this number So I don know how we get even to continue to do 3x compute scaling year over year for the next few years, much less go beyond that. At the end of the month, I go through the time-honored tradition of closing my books. I start by opening Mercury, which is my banking platform, to make sure that all my transactions are properly categorized. Auto-categorization rules handle the predictable stuff pretty well. But I'm constantly working with new contractors, you know, tutors and researchers and videographers, and I'm also trying new tools. Manually categorizing all of these transactions would add a couple of hours of overhead every single month. So instead of going through them one by one, I have Command, which is Mercury's built-in AI, take a stab at all of them at once. Command proposes a category for each transaction and provides its rationale. I just review, I fix anything that's off, and I approve. And once all this work is done in Mercury, it syncs everything with QuickBooks. And Command's judgment calls are genuinely good. It does the obvious things like looking at the vendor, but it also investigates who on my team made the purchase and looks at notes and memos to build up as much context as possible. This is just one of the ways you can use Command to automate the backend of your business. To learn more, go to mercury.com slash command. Mercury is a fintech company, not an FDIC insured bank. Banking services provided through Choice Financial Group,
[10:00] Now, I want to clarify that 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. And then the price of compute is basically the raw inputs and the tools required to do this processing. I'm just talking about this current precincularity regime where AI compute merely 3x is year over year, which is not enough to offset how much more valuable AI is becoming over time. By the way, the fact that anthropic revenue has been 10x in year over year, whereas their compute has only been 3x in year over year, I think illustrates how strong the economies of scale are in the model business. And logically, this makes sense. 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, where each instance has to be retrained from scratch. 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. Okay, this was a narration of a blog post that I also released on my website at dhwarkesh.com. Check it out for other posts or to be notified when I release a post in the future. Otherwise, I'll see you for the next full episode.
