Dylan Patel – Two labs will soon control most of the world's workforce

Compute Economics and Market Concentration

• CapEx Trajectory: Global AI infrastructure CapEx is projected to exceed $1 trillion in 2024, rising to over $2 trillion in 2025, with potential scaling to $7–$10 trillion annually by the end of the decade when including data center and energy infrastructure.
• Frontier Lab Dominance: OpenAI and Anthropic currently represent ~30% of marginal compute added in 2024, with projections indicating they will absorb 40–50% of incremental compute in 2025 and potentially 70–80% by 2028.
• Compute Ownership: By late 2027 or 2028, it is projected that OpenAI and Anthropic will control a majority of the world's usable compute capacity, driven by their ability to outbid all other entities for resources.
• Revenue vs. Cost Efficiency: Revenue per megawatt for frontier labs has skyrocketed from negative gross margins in 2023 to approximately $50 million per megawatt for Anthropic, compared to base infrastructure costs of $10–$15 million per megawatt.
• Pricing Dynamics: To secure the massive compute required for future growth, labs will likely drive compute prices up to $25–$50 million per megawatt, a price point currently only justified by their superior monetization capabilities.
• Supply Chain Bottlenecks: Production of ASML EUV tools is projected to reach 100 units annually by 2030; this physical constraint creates a bottleneck that delays the scaling of fab capacity despite massive capital availability.
• Value Capture Shift: Value creation has shifted from the hardware supply chain (which captured value in 2023) to the model layer, where OpenAI and Anthropic are now capturing the majority of gross margins.

Investment Flows and Capital Constraints

• Funding Gap: Total AI CapEx is estimated at $11 trillion between 2024–2029; while cash flows will fund roughly $6 trillion, approximately $5 trillion will require new debt issuance.
• Interest Rate Pressure: The massive demand for capital to fund AI infrastructure is expected to push interest rates higher (e.g., from ~5% to 8%+ for hyperscalers), creating a "crowding out" effect that squeezes other sectors of the economy.
• Sovereign Debt Risk: Rising interest rates and debt servicing costs may trigger a "second Volcker shock," potentially causing default crises in high-debt developing nations (e.g., Pakistan, Nigeria) and reducing the value of non-AI equities with long-duration cash flows.
• Hyperscaler Role: Hyperscalers (Meta, Microsoft, Amazon) are transitioning from cash-flow generators to massive borrowers, utilizing balance sheets to hoard compute and lease it to frontier labs at premium rates.
• Internal vs. External Allocation: Frontier labs are increasingly allocating a larger fraction of their compute budget to training and R&D (up to 70%) rather than inference to accelerate AGI development, prioritizing long-term exponential returns over immediate short-term inference profits.

Geopolitics and Regional Disparities

• US vs. China Compute Split: As of 2024, the US accounts for ~70% of global AI compute deployment, while China's share has fallen to under 10% due to export controls and domestic supply constraints.
• Chinese Scaling: China is projected to begin a "hockey stick" expansion in domestic compute production around 2027–2028, potentially adding 30–50 gigawatts by 2029, though these chips will likely offer significantly lower performance per watt than US counterparts.
• Strategic Divergence: While the US faces political and regulatory headwinds slowing deployment, China is expected to aggressively accelerate compute scaling, potentially creating a significant gap in effective AI capabilities by the end of the decade.

Forward-Looking Projections and Risks

• Revenue Per Watt Forecasts: Revenue per megawatt for frontier labs is projected to reach $70–$80 million per megawatt by late 2027, assuming continued model improvements and lack of regulatory suppression.
• Regulatory Impact: Safety regulations (e.g., withholding best models like "Astra" or "Mythos" for external release) may cap revenue growth and slow the deployment of AI, preventing labs from fully monetizing their compute capacity.
• Labor Substitution: The "effective AI labor force" at the frontier is growing at an exponential rate (projected 10x annually), potentially surpassing the total human population count within a single lab by the end of the decade.
• Centralization Trajectory: Fundamental economic forces (economies of scale, RSI potential, and compute scarcity) drive extreme centralization, leaving only a few private entities or the government as the primary allocators of global compute power.
• Economic Reallocation: If the current growth trajectory holds, the AI sector could consume 25–33% of the US economy's output by 2028, forcing a total reallocation of capital away from traditional assets toward AI infrastructure.
• Singularity Risks: There is a risk of a "slow takeoff" scenario where governments delay model releases by months; during Recursive Self-Improvement (RSI), these delays could result in a decade's worth of internal progress being inaccessible to the public.