Gavin Baker: Why AI Demand Is Outrunning Compute Supply

What It Covers

Investor Gavin Baker and a16z's David George analyze AI's supply-demand imbalance, arguing compute capacity is severely constrained through 2028 while actual heavy users number under 10 million against 1.5 billion knowledge workers. They cover NVIDIA's ecosystem dominance, orbital data centers, open source economics, and why the AI buildout resembles no previous technology bubble.

Key Questions Answered

• •Supply-demand asymmetry: Fewer than 10 million heavy AI users currently drive roughly $180 billion in annualized revenue across frontier labs, against a total addressable base of 1.5 billion knowledge workers. Token spend at AI-native companies already reaches 10% of human compensation. When diffusion expands even modestly, existing compute infrastructure faces severe shortages through at least 2028, making undersupply the dominant risk rather than the overbuild most investors fear.
• •Compute payback economics: Neo-cloud operators like Nebius and CoreWeave achieve nine-to-ten month revenue paybacks per gigawatt deployed, with customers prepaying 50-60% upfront. Spot market pricing can compress paybacks further. NVIDIA data centers carry additional financing advantages: roughly $35 billion of a $50 billion deployment can be debt-financed at low rates through Blackstone, KKR, and Apollo, reducing the equity check to approximately $15 billion with residual value guarantees covered by chip gross profit.
• •Positive-sum AI market structure: Rather than zero-sum competition between frontier labs, open source, clouds, and chip companies, Baker argues all layers can expand simultaneously. Frontier models, open-weight models, inference clouds, neo-clouds, application companies, and NVIDIA can each grow because the total addressable market remains vastly underpenetrated. Investors and operators should resist framing AI as a winner-take-all race and instead evaluate exposure across multiple layers of the stack.
• •NVIDIA ecosystem strategy: NVIDIA's defensibility rests on vertical integration combined with horizontal openness. Jensen Huang has locked up TSMC fab allocations, DRAM, NAND, laser, and capacitor capacity years ahead. Critically, NVIDIA data centers are the most financeable in the industry, giving Jensen structural cost-of-capital advantages over competitors. Baker's rule: every 1% accelerator market share is worth roughly $100 billion, so competing chips should seek niche integration rather than direct confrontation.
• •Orbital compute viability: SpaceX's planned orbital data centers are roughly airplane-sized racks using solar panels for power and radiators kept in shadow for cooling. The economic case hinges on Starship reusability driving launch costs below $1 billion per gigawatt, versus roughly $15 billion in terrestrial power, cooling, and labor costs that face structural inflation from electrician shortages and copper constraints. Baker expects an NVIDIA Rubin rack to launch in Q4 2027, with orbital compute becoming swing capacity by 2028.
• •Open source token economics: Open-source tokens require identical compute to frontier tokens for equivalent model sizes — the difference is margin, not physics. The Llama license stipulates a 30% revenue share on commercial deployments, and open-source models are often more token-hungry per task, making them costlier than perceived. For enterprises, the optimal architecture combines a capable open-source base model fine-tuned on proprietary data via reinforcement learning, then routed alongside one or two frontier models — preserving IP while controlling inference costs.