Overview
The argument begins with a striking mismatch: Anthropic's revenue has reportedly grown tenfold annually, while frontier-lab compute expands only about threefold. If that divergence persists, the economic value generated by each unit of compute must rise dramatically, with the gains captured through higher model margins, higher compute prices, or a greater allocation of hardware to inference. The speaker says all three are already visible, citing improving inference margins, rising spot prices, and inference's growing share of OpenAI's compute. Yet labs resist shifting most capacity away from training because their strategic value depends on producing substantially better models, not becoming ordinary cloud providers. This tension could make secure, large-scale GPU capacity much more expensive and strengthen incumbent labs that can monetize it most efficiently. Better models would then command premiums because they achieve the same outcome with fewer costly tokens, while low-value uses could be displaced by applications such as software engineering and automated AI research. The thesis is explicitly bounded: compute scarcity may eventually disappear under advanced robotic production, but semiconductor constraints make rapid supply expansion difficult in the current pre-singularity period. The likely near-term result is higher prices, stronger scale economies, tougher entry barriers, and greater concentration of economic power.
Sections
Forecasts
Conditional projections about revenue, compute economics, applications, and long-run supply.
- Anthropic may finish the current year with $100 billion to $150 billion in revenue.
- If Anthropic's tenfold annual growth continues, it would need roughly $1 trillion in revenue by the end of next year, although the speaker calls this a wild and capability-dependent conclusion.
- AI could consume 86% of TSMC's leading-edge N3 wafer allocation, up from 60%, limiting further gains from reallocating capacity away from smartphones and PCs.
- Many currently popular, relatively low-value AI applications will probably be priced out when higher-value automation can bid more for compute.
- Compute will eventually become cheap again if advanced robots can manufacture chips largely from raw materials and automated tools.
Economic and Infrastructure Data
Specific figures and mechanisms used to support the compute-scarcity thesis.
- Anthropic reportedly ended last year with $9 billion in revenue after three consecutive years of approximately tenfold annual growth.
- Anthropic's inference margins reportedly increased from about 40% in the middle of last year to more than 80% for Fable.
- Compute spot prices were said to be more than 40% above their February trough.
- Epoch estimated that OpenAI used about 25% of its compute for inference in 2024; the speaker believes the current share may be around 50% or higher.
- Google reportedly rents 110,000 GB200 and GB300 GPUs from SpaceX for $900 million per month, at roughly twice the prevailing spot hourly price.
- The claimed threefold annual compute increase is decomposed into about 1.4x from Moore's Law, 1.2x from new fabs, and 1.8x from shifting wafer allocation toward AI.
Strategic Implications
Higher-level conclusions derived from the interaction between capability growth, compute scarcity, and model economics.
- Compute scarcity and model efficiency may reinforce each other: expensive compute increases demand for efficient models, while efficient models raise the economic value and affordable bid price of compute.
- The strongest labs could gain a self-reinforcing advantage because superior monetization lets them outbid rivals for scarce hardware, which then supports further training and capability gains.
- Inference revenue is not merely a profit center for frontier labs; it functions as evidence used to attract capital for continued training, creating tension between present monetization and future capability development.
- AI's replication economics differ fundamentally from human labor: a single training investment produces reusable capabilities across many users, making intelligence unusually scalable and potentially concentrated.
Risks and Uncertainties
Weak points, adverse outcomes, and limitations acknowledged by the speaker.
- The revenue extrapolation could fail if AI capabilities do not become sufficiently useful to sustain another tenfold increase.
- Margins above 90% may attract competition and be difficult to sustain unless the leading model remains substantially better than available substitutes.
- A massive, rapid increase in AI labor supply could reduce the marginal value of software engineering despite standard economic arguments about innovation and specialization.
- The scarcity thesis may repeat historical Malthusian errors by underestimating substitution, efficiency improvements, and supply responses.
- Strong economies of scale in model training may concentrate economic and political power among a small number of frontier labs.
- Compute growth may fall below the assumed threefold annual rate because Moore's Law, fab construction, EUV-machine supply, and wafer reallocation all face physical limits.