~1M unlocked RTX 3090s in China ≈ 140 exaFLOPS BF16, rivaling ~140k H100s for inference
teortaxesTex · x · 2026-10-25
A detailed back-of-envelope estimate circulating around unlocked RTX 3090s:
- Of 2.5–3.5M 3090s worldwide, 30–45% sit in China; after attrition that's roughly 1 million cards (wide error bars).
- An unlocked 3090 delivers 142 TFLOPS dense BF16—roughly matching NVIDIA's China-specific H20 (148 TFLOPS) in matrix compute—at maybe a tenth of the price. Unlocking also shifts its compute-to-bandwidth ratio from 76 to 150 math ops/byte, closer to a datacenter part.
- A million cards total 140 exaFLOPS peak dense BF16 (70 at stock), the peak matrix compute of roughly 140k H100s; aggregate memory bandwidth 0.9 EB/s, equivalent to 280k H100s.
- The big caveat: capacity is scattered across a million machines—24 GB each, no fast interconnect beyond pair NVLink, 350 W each—so it's nearly useless for frontier training, but very useful for many independent inference sessions of small, bandwidth-frugal models.
TeortaxesTex argues that acquiring a substantial share of China's 3090s would be very useful if cornered with cheap electricity; side discussion notes Ada parts can just run fp8, and the rumor concerned the 5090 anyway.
More from Infra
- Nebius cached-token pricing sends its OpenRouter traffic vertical, from under 20B to 67B tokens a day — AccBalanced · 2026-10-25
- Best setup to run Qwen3.8 27B locally on an M4 Pro 48GB Mac? — 9kSs · 2026-10-25
- Security researchers find 32 NVIDIA B300 GPUs on Chinese networks in scan of 12,000+ exposed GPUs — AccBalanced · 2026-10-25
- Xinjiang emerges as an AI compute hub with half-price power and 16.2ms latency to Chongqing — 96Stats · 2026-10-25
- Reddit cost table: cheap-first-then-escalate beats Sonnet-only for agent runs — lucasbennett_1 · 2026-10-25
- Anthropic's $518B compute commitments are only 22% NVIDIA — TPUs take 53% — AccBalanced · 2026-10-25