Other
Why compute prices might 10x as AI gets smarter
- Anthropic is projected to achieve tenfold year-over-year revenue growth again this year, potentially concluding with revenue between $100 billion and $150 billion, though maintaining this trend would require $1 trillion in revenue by the end of the next year.
- Compute spot prices have increased by more than 40% compared to the February trough, and OpenAI's inference compute allocation is estimated at 50% or higher relative to a quarter in 2024, with inference margins for top models potentially rising from over 80% to greater than 90%.
- Industry labs anticipate building models within a year that drastically outperform current iterations, intending to allocate the majority of compute toward training and experiments for the next generation of models.
- Economic viability for leading models depends on the Alkin-Allen effect allowing higher margins for the most efficient training, requiring models to be sufficiently ahead of competition to prevent margin erosion, while smarter models are expected to monetize compute more effectively.
- Compute scaling faces significant supply constraints, including a 1.4x growth rate from Moore's law becoming difficult to sustain, a 3x year-over-year capacity scaling limit, and bottlenecks in ASML EUV machine production extending until 2030 or later.
- Wafer allocation trends indicate that AI consumption of leading-edge N3 nodes at TSMC will shift from 60% to 86% allocation by the end of next year, potentially reaching a limit where AI absorbs supply from smartphones and PCs.
- Labor economics regarding AI are uncertain, with potential for the marginal value of software engineers to decrease if supply increases tenfold, though standard economics suggests high marginal value for labor and compute should persist if the market does not break established heuristics.
- Scarcity of compute resources is highlighted by an inelastic supply compared to metal extraction, leading to expectations that labs will pay more for tokens to automate research than individual users pay for low-quality generation, and eventually rendering current applications too expensive as capabilities improve.
- Future market dynamics may see a true human-level software engineer running on an H100 equivalent command a rental price exceeding $250,000 annually, and the industry awaits a point where robotics can cheapen compute by converting raw materials into chips.
- Risks include the difficulty of sustaining current scaling rates, the potential error in applying scarcity analysis from past economic debates, and the uncertainty of whether the marginal value of labor will withstand a massive labor supply shock.