ExpertsFMWhy smarter AI models could drive up compute prices 10x
Dwarkesh Podcast · August 3, 2026
5 Selected Lessons:
reasoning_pattern · 3 framework · 2
Revenue-compute gap decomposition
Source
Dwarkesh Podcast — Why smarter AI models could drive up compute prices 10x (August 3, 2026) · at 0:00 ▸
When this applies
when a business's output/revenue is scaling faster than its core scarce input, and you need to explain where the gap is being absorbed
The takeaway
When growth in revenue outpaces growth in a scarce input, decompose the gap into margin expansion, input price inflation, and reallocation of the input's use — at least one must move.
Why it works
If output revenue grows faster than the input that produces it, the gap must be absorbed by some combination of higher margins, higher input prices, or reallocating input away from other uses.
One level up
Any apparent violation of a supply-demand identity signals that some adjustment variable is absorbing the imbalance; find it.
Where it stops holding
This stops working when the input itself is not actually a binding constraint, or when revenue growth is illusory/unsustainable and will simply fall back to match input growth.
Detail
procedure:
- Identify the growth rate mismatch (e.g., revenue 10x/yr vs compute 3x/yr)
- Enumerate the only possible absorption mechanisms: margin increase, input price increase, reallocation of input share
- Check actual data for evidence of each mechanism already occurring
- Determine which mechanism(s) must dominate going forward
benchmarks: {"anthropic_margin_increase":"40% to 80%","compute_price_increase":"40%+ above Feb trough","inference_share_increase":"25% to ~50%"}
rationale: An accounting identity forces these three levers; all three were observed to be moving simultaneously.
In their words
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.
Decomposing compute capacity growth into its structural drivers
Source
Dwarkesh Podcast — Why smarter AI models could drive up compute prices 10x (August 3, 2026) · at 7:49 ▸
When this applies
when evaluating whether a historical growth rate in any capacity metric can continue, break it into its independent structural components and check each for headroom
The takeaway
To judge whether a compounding growth trend can continue, decompose it into its independent multiplicative drivers and check each one against its own ceiling.
Why it works
A composite growth rate is the product of multiple independent structural factors; if each factor individually has hard physical, engineering, or allocation limits, the composite rate cannot be sustained once those limits bind.
One level up
Aggregate growth rates are only as sustainable as their most constrained underlying component.
Where it stops holding
This stops working when new structural drivers (not yet identified) can emerge to replace saturating ones, or when the components aren't actually independent/multiplicative.
Detail
procedure:
- Decompose the aggregate growth rate into multiplicative components
- Assess each component's headroom against known physical/engineering limits
- Flag components nearing saturation (e.g., wafer allocation to AI going from 60% to 86%)
- Conclude on likely future trajectory of the aggregate rate
benchmarks: {"moores_law":"1.4x/yr, likely topping out","fab_buildout":"1.2x/yr, bottlenecked by ASML EUV machine supply through 2030+","wafer_reallocation_to_AI":"1.8x/yr, N3 node AI share going from 60% to 86%, nearing saturation by end of next year"}
rationale: Each of the three multiplicative drivers of compute growth faces a concrete, near-term ceiling, making sustained 3x/year growth implausible.
In their words
So 1.4x of that is coming from Moore's Law. Far from increasing it, think it'll be a miracle if you 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.
Lump-of-labor fallacy applied to AI/compute monetization
Source
Dwarkesh Podcast — Why smarter AI models could drive up compute prices 10x (August 3, 2026) · at 2:19 ▸
When this applies
when forecasting whether a huge new supply of a productive capability (labor, compute, any input) will collapse its marginal value
The takeaway
Don't assume a supply shock in a productive input will proportionally crash its marginal value; consider whether specialization/innovation effects could expand demand to match.
Why it works
Standard economics holds that new labor supply doesn't necessarily depress wages long-term because innovation and specialization expand the total value created, rather than just being divided among more supply — the total pie grows.
One level up
Apply the same economic logic to non-human analogues of labor/capital that you'd apply to humans, unless there's a specific reason the analogy breaks (e.g., speed or scale of the shock).
Where it stops holding
This stops working when the supply shock is so large or fast that normal adaptive/specialization mechanisms can't keep pace — the speaker explicitly flags this as a possible exception for AI.
Detail
approach: Apply the standard economic rebuttal to lump-of-labor fears (used for immigration) to the AI compute case
signals:
- claims that abundant new capable labor/compute will crash its own value
- comparison to historical immigration wage fears
rationale: Economists generally find immigration doesn't decrease wages long-run due to innovation/specialization effects; the speaker extends this cautiously to AI/compute but flags the analogy could break.
In their words
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 wages in the long run because of how innovation and specialization increase the value of labor.
Alchian-Allen effect applied to AI model efficiency and compute scarcity
Source
Dwarkesh Podcast — Why smarter AI models could drive up compute prices 10x (August 3, 2026) · at 5:18 ▸
When this applies
when a fixed/overhead cost rises, the relative penalty for using a lower-quality or less efficient option on top of that fixed cost increases, favoring the premium option
The takeaway
When the cost of a scarce input rises, the premium commanded by whoever uses that input most efficiently increases disproportionately (the Alchian-Allen effect), not just linearly.
Why it works
As the expensive base cost (compute) rises, the opportunity cost of wasting it on an inefficient model (that burns more of it for the same output) grows, so efficient/premium producers can charge proportionally more.
One level up
Rising costs for a shared scarce input shift value toward efficiency and quality, not away from it — scarcity amplifies the value of efficiency.
Where it stops holding
This stops working when the quality/efficiency difference between options is negligible, or when the input in question isn't actually expensive relative to other costs.
Detail
approach: Reason from the base economic principle (Alchian-Allen) to a prediction about AI lab pricing power
signals:
- rising cost of the fixed/base input (compute)
- existence of variable-efficiency competitors (models) using that input
rationale: If compute is $20/hr, using an inefficient model that burns more tokens for the same result becomes relatively more costly, so efficient model providers can charge a larger premium — effectively 'creating more compute' through efficiency.
In their words
This is the Alkin Allen effect in economics and what it's basically saying is that it would cost $20 an hour to rent an H100. That 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.
Scarce input reallocates to highest-value use as capability rises
Source
Dwarkesh Podcast — Why smarter AI models could drive up compute prices 10x (August 3, 2026) · at 5:18 ▸
When this applies
when a previously abundant resource becomes scarce, expect it to flow toward whichever use generates the highest marginal value, displacing lower-value uses
The takeaway
Rising scarcity of an input causes allocation to concentrate on its highest-value applications, pricing out marginal/low-value uses.
Why it works
As an input's price rises due to scarcity, usages with lower willingness-to-pay get priced out while usages with higher economic value (e.g., automating valuable R&D) continue to be funded.
One level up
Scarcity functions as a sorting mechanism, reallocating resources toward the economically most productive use cases.
Where it stops holding
This stops working when low-value uses are subsidized, bundled, or otherwise insulated from true market pricing of the scarce input.
Detail
approach: Project current economic allocation logic (ability-to-pay sorting) forward as compute costs rise
signals:
- rising compute cost
- divergent willingness-to-pay between institutional and consumer uses
rationale: Google/Anthropic/OpenAI will be willing to pay more for tokens that automate AI research than ordinary users will pay for low-value content generation, pricing out many current cheap AI applications.
In their words
At that point, 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.