Interview, Fireside Chat
Why AI Demand Is Outrunning Compute Supply
- The AI sector is projected to be defined by two major themes: a fundamental societal transformation led by key industry figures and the emergence of a "compute inequality" within two years, where supply shortages restrict access primarily to large corporations and wealthy individuals.
- Acceleration is expected to continue through August with significant gains in open-source models and Grok, though growth cannot be sustained indefinitely; conversely, Anthropic is anticipated to have slowed relative to the broader market but is preparing for a re-acceleration following internal testing and upcoming IPO activities.
- Market expectations include continued volatility and drawdowns for public AI stocks, alongside a shift in investor sentiment from immediate ROI concerns to acceptance of long-term research prioritization and operating cash flow generation rather than free cash flow in the near term.
- Financial modeling suggests that shifting compute allocation from inference to training could reduce annual revenue from $480 billion to $120 billion, yet the potential for $60 billion per gig of power in revenue remains attractive if the eight gig allocation is maintained for inference.
- Several entities, including NVIDIA, the "five lab companies," and clouds, are forecast to perform well, driven by aggressive reinvestment of incremental profits into training and the belief that no single model will dominate, necessitating an ensemble approach and vertical integration for competitive advantage.
- Supply-side economics are expected to tighten significantly, with predictions that token prices could rise 10x due to massive demand and limited supply, while internal token consumption has already surged 100x from March to August and may increase 10-20x monthly with enterprise bot deployments.
- Strategic financial instruments are expected to evolve, with companies utilizing Residual Value Guarantees to finance $50 billion data centers with $15 billion equity checks, while chip companies like NVIDIA are positioned to benefit from $15 billion equity checks and TSMC allocation constraints.
- The physical infrastructure outlook includes the re-industrialization of America driven by data centers, which could increase tax revenues 10-fold in small towns due to low natural gas costs, alongside a transition to orbital compute facilitated by Starship's target of under $1 billion launch costs and a future fleet of 5,000 annual launches.
- Long-term horizon predictions extend to 2027 with the launch of the Rubin rack, asteroid mining viability, and the establishment of heavy industry in outer space, alongside the deployment of fleets landing on Mars within eight years and the societal shift of Earth becoming primarily residential.
- Competitive dynamics in the application and model layer will involve companies using base models for RL and supervised fine-tuning to create proprietary intelligence, with specific predictions that American open source will closely rival frontier tiers and that Microsoft will face competition from Databricks and Palantir for the abstraction layer.
- Risks include the potential for market bubbles caused by overvaluation leading to overbuild and debt-funded construction demanding immediate returns, as well as the possibility that real rates will rise due to heavy investment and that stock volatility will severely impact employee morale and retention strategies.