Aug 3, 2026, 1:36 PM ETOther

Why compute prices might 10x as AI gets smarter

Anthropic's revenue grew 10x year-over-year for the last three consecutive years, ending the prior fiscal year at $9 billion.
Projected revenue for the current year ranges between $100 billion and $150 billion, contingent on AI capability utility.
For the 10x revenue trend to continue, Anthropic would theoretically need to reach $1 trillion in revenue by the end of next year.
Conversely, AI lab compute capacity is currently scaling at approximately 3x year-over-year.
A mathematical divergence exists where revenue growth (10x) vastly outpaces compute supply growth (3x), necessitating structural shifts in the industry economics.

To sustain 10x revenue growth with 3x compute growth, one or a combination of three economic factors must occur:
- Margin Expansion: Lab inference margins are reportedly rising from 40% (mid-last year) to over 80% currently.
- Compute Price Increases: Spot prices for compute are more than 40% higher than the February trough of this year.
- Inference Share Shift: The percentage of compute allocated to inference rather than training is rising; OpenAI's inference spend grew from ~25% in 2024 to an estimated 50% or higher.
Labs deliberately delay shifting the majority of compute to inference because they fear signaling that AI progress has stalled, which would weaken their narrative to investors focused on AGI rather than cloud services.

A key case study involves Google renting compute from SpaceX:
- Google pays $900 million monthly for 110,000 GPUs (a blend of GB200s and GB300s).
- This rate is 2x the standard spot price, itself already >40% higher than February lows.
If an AI model achieves human-level software engineering capability on current hardware (e.g., H100 equivalent), its marginal revenue could theoretically support rental rates exceeding $250,000/year, or 15x the current spot price for an H200.
Alkin-Allen Effect: Labs can charge significant premiums for models that economize scarce compute; a more efficient model effectively "creates" more compute by requiring fewer tokens for the same output.
High marginal value of compute creates a moat for frontier labs, making it difficult for competitors to bid for resources against entities that can monetize the same hardware more effectively.

The current 3x annual compute growth rate is unlikely to be sustained due to three hard constraints:
- Moore's Law: Contributing 1.4x; the speaker notes it is a "miracle" if this pace can be maintained for even a few more years.
- Fab Construction: Contributing 1.2x; bottlenecked by the supply of new ASML EUV machines until 2030 or later.
- Wafer Allocation: Contributing 1.8x; AI is absorbing lead-edge TSMC capacity (N3 nodes), projected to rise from 60% to 86% by the end of next year.
Once AI consumes nearly all leading-edge wafer capacity, the supply of compute becomes inelastic and unable to absorb further demand shocks or substitute for other industries (unlike metal extraction).

Power Concentration: The strong economies of scale in model training (one-time cost vs. shared user access) are expected to concentrate intelligence power, a trend the speaker views with concern.
Commoditization vs. Scarcity: The speaker argues that current compute scarcity differs from historical resource bets (e.g., Simon Ehrlich's commodity bet) because compute supply is less elastic and harder to substitute than raw materials.
Long-term Deflation: The speaker anticipates a future "post-singularity" regime where automated manufacturing converts silica and copper into chips, eventually reducing compute costs to raw input levels.
Current Regime: We remain in a "pre-singularity" phase where the 3x compute growth is insufficient to offset the exponential increase in AI value, driving prices higher.