If a human-level software engineer that could run on an H100 equivalent, at current market rates for software engineers, that H100 should rent for over $250k a year. That’s 15x today’s spot price.
Really enjoyed this. I have a different view, but rather than pushing back, I'd like to add a dimension to the discussion.
Software engineering is a special "light cone" domain — no atoms, no human-scale latency in the loop. That's exactly where the $250k/H100 mechanism is strongest, because intelligence really is the bottleneck, so improving it monetizes the whole labor loop.
But once the loop touches the physical world, the bottleneck moves off intelligence — and stops tracking labor price. AI can collapse drug design toward compute time, but the first dose in a human is still a trial: real-world inference, not AI inference. So compute only monetizes at displaced-labor rates where the loop is fully simulable. The $250k might be real as the first human→AI conversion rate, but it's elastic — once the coding agent takes over, it can fall to $50k.
The supply-wall point stands either way — a scarcity story independent of capability, and probably the more robust half for the next few years.
> A lot of current popular applications of AI get priced out. The reason AI is relatively cheap right now, at least in comparison to human labor, is partly that it can’t do a lot of things that top humans can do. At some point that will no longer be the case. And so using GPUs to make short-form video slop will just get priced out.
Token costs are not compute costs. You don't need a drop-in-SWE-level model (and associate cost per token) to serve AI slop. I'd be very surprised if, given continual efficiency gains, it costs more to serve a 2026-level AI slop video or chat response in the future. The proportion of compute serving them likely goes down as AI moves into more and more advanced use cases like drop-in workers.
Lots of interesting stuff but the claim that the proprietary frontier models are the most efficient use of compute is an important sub-question that requires more scrutiny. The non-frontier OS and Chinese models tend to win on intelligence/price (a proxy for compute and also tend to be smaller and more lightweight).
In fact, if we assume there is some other constraint or diminishing returns on the utility of more intelligence systems it might actually be the case that running 10 instances of GLM-5.5 is a better use of computational resources than 1 instance of GPT-6.
I wouldn't automatically assume that the newest models are the most efficient in a world where the OS models are so hot on the heels of the proprietary ones.
There are many tasks where you can run 1000 instances of GLM-5.5 and not get a solution, but where you can run 10 Fable-5-max instances and solve your task.
This heavily underestimates “compute” supply growth and efficiency gains, specifically from software-hardware co-design. For example GPU manufacturers optimising for LLM usage, and AI labs optimising kernels and model structures for training. Vera rubin GPUs as an example are estimated to be 4x more efficient than blackwell for inference, which was 2.5x more efficient than Hopper.
The question comes down to whether algorithmic/hardware improvements increase effective compute supply to a point where demand is satiated with surplus compute, or whether it just further boosts the monetization of the compute (eg. 2 SWEs per H100e instead of 1) and the economy can just absorb the increased "labor" productively
Really interesting thx for sharing. I wrote something adjacent to this last night but from a much more simplistic energy angle. If abundant intelligence/ advanced AI should improve algorithms, chips and efficiency, why are we extrapolating today’s energy curve so far out?
I landed on a similar tension. Less compute per unit of intelligence can still mean far more total compute if usage explodes. Your point on the value and price of scarce compute during the transition adds a key layer.
Why are you fixated on LLMs? Progress more likely to come from world models, reasoning, local learning and shared memory, which are likely to depend much more on algorithmic improvements than pure compute. The niche that is programming tools and bash scripts moving files around is likely to be a small subset of future applications.
Better chips do not guarantee cheaper AI when demand is growing faster than supply. If compute becomes 10x more expensive, will efficiency save smaller players, or will early access to capacity become the real competitive moat?
Very interesting line of thinking. I could see the value/price of compute rising a good bit once we get to the drop-in worker milestone. At the same time, the lump of labor fallacy doesn't repeal supply and demand, nor the response of markets to prices.
If an H100 could generate $250k per year, people would find nearly infinite ways to invest in providing compute. Consider as one example that (according to my quick search, could be wrong), an iPhone has 1/1000 the compute of an H100. That implies an iPhone could generate $2.5k per year, and so both consumers and companies like Apple would very quickly monetize this as well and expand the relevant supply of compute.
I also think the H100 / SWE equivalence assumption is already too low and will only become more so as inference gets more efficient. According to ChatGPT, an H100 can probably support ~10 5.6 Sol sessions concurrently. I would also guess a session will be able to do as much work as 10 human SWEs once we get to drop-in worker. So we would actually be looking at up to 1,000 M new supply of SWE-equivalents, not just 10M. It seems implausible that the clearing price for SWE work would not go down in this scenario... Unless we also have lots of agents generating demand for (profitable) work, in which case we're post-AGI, economic growth is off the charts, and predictions get very hard indeed.
if compute gets 15x more expensive, doesn't that price out a lot of customers? and wouldn't competition between labs/open source push prices back down toward cost rather than toward the value of the labor being replaced?
Really enjoyed this. I have a different view, but rather than pushing back, I'd like to add a dimension to the discussion.
Software engineering is a special "light cone" domain — no atoms, no human-scale latency in the loop. That's exactly where the $250k/H100 mechanism is strongest, because intelligence really is the bottleneck, so improving it monetizes the whole labor loop.
But once the loop touches the physical world, the bottleneck moves off intelligence — and stops tracking labor price. AI can collapse drug design toward compute time, but the first dose in a human is still a trial: real-world inference, not AI inference. So compute only monetizes at displaced-labor rates where the loop is fully simulable. The $250k might be real as the first human→AI conversion rate, but it's elastic — once the coding agent takes over, it can fall to $50k.
The supply-wall point stands either way — a scarcity story independent of capability, and probably the more robust half for the next few years.
> A lot of current popular applications of AI get priced out. The reason AI is relatively cheap right now, at least in comparison to human labor, is partly that it can’t do a lot of things that top humans can do. At some point that will no longer be the case. And so using GPUs to make short-form video slop will just get priced out.
Token costs are not compute costs. You don't need a drop-in-SWE-level model (and associate cost per token) to serve AI slop. I'd be very surprised if, given continual efficiency gains, it costs more to serve a 2026-level AI slop video or chat response in the future. The proportion of compute serving them likely goes down as AI moves into more and more advanced use cases like drop-in workers.
Lots of interesting stuff but the claim that the proprietary frontier models are the most efficient use of compute is an important sub-question that requires more scrutiny. The non-frontier OS and Chinese models tend to win on intelligence/price (a proxy for compute and also tend to be smaller and more lightweight).
https://artificialanalysis.ai/#price-and-cost
In fact, if we assume there is some other constraint or diminishing returns on the utility of more intelligence systems it might actually be the case that running 10 instances of GLM-5.5 is a better use of computational resources than 1 instance of GPT-6.
I wouldn't automatically assume that the newest models are the most efficient in a world where the OS models are so hot on the heels of the proprietary ones.
There are many tasks where you can run 1000 instances of GLM-5.5 and not get a solution, but where you can run 10 Fable-5-max instances and solve your task.
Naive $/intelligence measurements miss that part.
This heavily underestimates “compute” supply growth and efficiency gains, specifically from software-hardware co-design. For example GPU manufacturers optimising for LLM usage, and AI labs optimising kernels and model structures for training. Vera rubin GPUs as an example are estimated to be 4x more efficient than blackwell for inference, which was 2.5x more efficient than Hopper.
The question comes down to whether algorithmic/hardware improvements increase effective compute supply to a point where demand is satiated with surplus compute, or whether it just further boosts the monetization of the compute (eg. 2 SWEs per H100e instead of 1) and the economy can just absorb the increased "labor" productively
I wonder how massive efficiency gains bring a lot more existing compute onboard (eg macbooks) balancing this out.
Really interesting thx for sharing. I wrote something adjacent to this last night but from a much more simplistic energy angle. If abundant intelligence/ advanced AI should improve algorithms, chips and efficiency, why are we extrapolating today’s energy curve so far out?
I landed on a similar tension. Less compute per unit of intelligence can still mean far more total compute if usage explodes. Your point on the value and price of scarce compute during the transition adds a key layer.
https://substack.com/@vrash/note/p-208862969?r=776ej
Why are you fixated on LLMs? Progress more likely to come from world models, reasoning, local learning and shared memory, which are likely to depend much more on algorithmic improvements than pure compute. The niche that is programming tools and bash scripts moving files around is likely to be a small subset of future applications.
Better chips do not guarantee cheaper AI when demand is growing faster than supply. If compute becomes 10x more expensive, will efficiency save smaller players, or will early access to capacity become the real competitive moat?
Curious how did you estimate 1.4x from Moore's law?
We are not bullish enough!!!
I've created a market based off the title's 10x hypothesis: https://manifold.markets/nsokolsky/how-high-will-the-monthly-average-h
Very interesting line of thinking. I could see the value/price of compute rising a good bit once we get to the drop-in worker milestone. At the same time, the lump of labor fallacy doesn't repeal supply and demand, nor the response of markets to prices.
If an H100 could generate $250k per year, people would find nearly infinite ways to invest in providing compute. Consider as one example that (according to my quick search, could be wrong), an iPhone has 1/1000 the compute of an H100. That implies an iPhone could generate $2.5k per year, and so both consumers and companies like Apple would very quickly monetize this as well and expand the relevant supply of compute.
I also think the H100 / SWE equivalence assumption is already too low and will only become more so as inference gets more efficient. According to ChatGPT, an H100 can probably support ~10 5.6 Sol sessions concurrently. I would also guess a session will be able to do as much work as 10 human SWEs once we get to drop-in worker. So we would actually be looking at up to 1,000 M new supply of SWE-equivalents, not just 10M. It seems implausible that the clearing price for SWE work would not go down in this scenario... Unless we also have lots of agents generating demand for (profitable) work, in which case we're post-AGI, economic growth is off the charts, and predictions get very hard indeed.
dwarkesh is becoming a bull! what did sholto revealed? continual learning is coming?
if compute gets 15x more expensive, doesn't that price out a lot of customers? and wouldn't competition between labs/open source push prices back down toward cost rather than toward the value of the labor being replaced?