GPUs Are Now Geopolitical Assets — Compute Access Is a Structural Constraint

FT reports banned Nvidia AI chips sell for ~2x on China’s black market. A DGX B300 reportedly fetches 8M RMB ($1.1M) versus ~$400K in the US. Over $1B in restricted chips was smuggled in within three months of tighter controls. As compute access becomes a geopolitical risk, founders building on GPUs face shaky cost and supply.

What Happened

The Financial Times reported the prices on China’s gray market. The Nvidia AI chips Washington blocked — Blackwell-based DGX and RTX, plus B200, H100, and H200 — trade at roughly twice their original price there. A single DGX B300 reportedly sells for around 8 million RMB, about $1.1 million. The same system lists for roughly $400,000 in the US. More than double. Tom’s Hardware and others report that over $1 billion in restricted chips was smuggled into China in just three months after controls tightened. When the US Commerce Department issued May 31 guidance closing the loophole that routed high-end chips through foreign subsidiaries, the choked supply drove prices up. It isn’t that controls stopped the chips from arriving — it’s that controls made them arrive more expensively. Washington still treats export controls as working; the market is sending the opposite signal.

What This Means for Founders

This isn’t just a US-China story — it’s a cost-structure problem for everyone building on AI. Compute access has entered a regime where it’s priced by policy and geopolitics, not by the market. Even founders far from China feel it. When controls artificially block one slab of demand, that demand spills into the gray market and into cloud rental, distorting GPU prices and availability worldwide. A startup in Silicon Valley or Bangalore is, in effect, competing for the same finite pool of H100s and B200s as Chinese gray-market buyers and the captive demand of hyperscalers hoarding capacity for their own models. At the same time, the distortion creates openings. First, sovereign compute. Nation- and region-level GPU infrastructure — local data centers, government-backed clouds — is becoming a way to lock in stable supply outside the squeeze, and the founders who secure allocation early step out of the supply risk. Second, chip abstraction. Don’t pin your code to one GPU generation; architect inference so it can move across AMD, custom inference accelerators, and multiple clouds, and a single blocked source won’t halt the business. Third, price pass-through. In an environment where compute cost swings with policy, absorbing pure usage-based pricing evaporates your margin. Design pricing that can absorb cost volatility from day one.

What You Can Do Now

First, audit how tightly your business is bound to one GPU. If all your inference rides on a single cloud and a single chip generation, you have no fallback when prices double. Second, diversify your compute supply lines — domestic sovereign clouds, multiple foreign regions, and inference-specific accelerators — so a blocked source doesn’t take you down. Third, track per-token and per-image cost in real time and build slack into your pricing to absorb cost swings. In an era when compute lurches with geopolitics, supply stability itself is a competitive edge.