NVIDIA’s entire 2025 Blackwell production was sold out before it shipped. Demand is “extreme.” Customers are ordering in 100,000-GPU quantities – up from tens of thousands for the previous Hopper generation. AWS, Google, Meta, Microsoft, and Oracle are buying every B200 that NVIDIA and TSMC can produce. The GPU chokepoint thesis isn’t a theory – it’s the market reality.
The Data
Morgan Stanley reported that “the entire 2025 production” of Blackwell silicon was “already sold out” by November 2024 – before volume shipments began. Orders are sold out for 12+ months. The scale has shifted: Hopper orders came in tens of thousands; Blackwell orders come in hundreds of thousands simultaneously.
Despite NVIDIA’s efforts to ramp production, supply of PCIe solutions based on the B200 chip remains constrained. The company is prioritizing GB200 series delivery. Cloud GPU supply remains “tight” per NVIDIA’s fiscal Q3 2026 earnings. The shortage persists through December 2025 with no relief in sight.
Framework Analysis
The sellout validates GPU economics analysis. The B200 costs approximately $6,400 to produce and sells for $30,000-$40,000 – implying 75-80% gross margins. But NVIDIA’s pricing power depends on supply chain constraints it doesn’t control: HBM memory (45% of production cost), CoWoS advanced packaging, and TSMC fab capacity.
The 100,000-GPU order quantities reflect hyperscaler infrastructure planning. When Microsoft commits $80 billion and Amazon’s Project Rainier exceeds $100 billion, GPU procurement happens at data-center scale. The AI memory chokepoint compounds the constraint – HBM is sold out through 2026.
Strategic Implications
For AI builders, the 12-month sellout means infrastructure planning extends to 2027. Companies without existing supply agreements face multi-year waits. The hyperscalers who locked in capacity early – through purchase commitments and strategic partnerships – hold structural advantages.
For NVIDIA competitors, the demand validates the market but the supply constraints create opportunity. AMD’s MI300X and custom ASICs from Google, Amazon, and Microsoft gain relevance when NVIDIA can’t ship enough. The CUDA moat matters less when you can’t get GPUs at all.
The Deeper Pattern
Scarcity at chokepoints determines who can scale AI. The companies that secured GPU supply in 2024 can train models in 2025-2026. The companies that didn’t are locked out of the frontier regardless of their AI talent or algorithms.
Key Takeaway
NVIDIA’s 12-month Blackwell sellout with 100,000-GPU orders validates the chokepoint thesis. GPU supply – not algorithms, not data, not talent – is the binding constraint on AI scaling.
Sources:
