Claims that major Chinese models train on Blackwell chips draw scrutiny over scale

Around the claim that "all major Chinese models have been trained on Blackwell-class chips," a dense cluster of rumors and back-of-envelope calculations emerged on November 13. tszzl posted that "all major Chinese models are trained on Blackwell-class chips," calling it "pretty bad"; teortaxesTex quoted this and inferred it implies DeepSeek, Kimi and other top labs already have working Blackwell training stacks—unsurprising for DeepSeek, since its open-sourced TileLang/DeepInfer explicitly supports Blackwell and Ascend but not Hopper, showing its software stack is ready for new hardware, though related costs carry a 3-6x premium.

Confirmed

• DeepSeek's open-sourced TileLang/DeepInfer supports Blackwell and Ascend, not Hopper (public code cited by teortaxesTex).
• Inkling 2S completed its first-stage pre-training in just 68 hours (data cited in the post); teortaxesTex extrapolates that on a GB200-class cluster, V4.1F's pre-training should take no more than two weeks—for comparison, V3 took about 2 months on 20,000 GPUs back then (per the original post's comparison).
• SemiAnalysis data: China's total installed compute (including small, low-utilization clusters) is roughly 24 GW; even if 5 labs each used 10 MW clusters for training, that would only account for about 0.2% (cited by teortaxesTex).

Unconfirmed

• The claim that "all major Chinese models are trained on Blackwell-class chips" remains a rumor with no direct evidence.
• Will Depue estimated Kimi K3 and DeepSeek-V4-Pro each used 2,000-3,000 Blackwell GPUs, with a single cluster costing about $150-200 million, inferring Chinese labs still obtained large-scale compute under export controls; teortaxesTex explicitly rebutted these figures as exaggerated.
• Scale of smuggled incremental compute: some claim about 1.5 GW; teortaxesTex estimates frontier labs like DeepSeek and Kimi got only about 3.3% of it, questioning where the rest went (hoarded by ByteDance or Alibaba?).

Why it matters

• If true, it would mean China's top labs built latest-generation Blackwell training stacks despite export controls, but the scale (labs' share of smuggled compute, cluster sizes) is far smaller than some analysts' exaggerated claims; cost premiums and compute allocation are key variables for watching China's AI frontier capability.

2026-11-13 ~ 2026-11-13 · 6 related posts

• [source] Chinese major models reportedly trained on Blackwell-class chips, though scale disappoints — teortaxesTex · 2026-11-13
• [source] Analyst argues labs already run Blackwell stacks, with 3-6x smuggling premium — teortaxesTex · 2026-11-13
• Back-of-envelope: GB200 cluster should pretrain DeepSeek V4.1F in under two weeks — teortaxesTex · 2026-11-13
• Did Kimi K3 and DeepSeek-V4-Pro each train on 2,000-3,000 Blackwells? Claim disputed — teortaxesTex · 2026-11-13
• [source] China's total installed compute is ~24GW; alleged 'data-center' runs are only 0.2% of it — teortaxesTex · 2026-11-13
• Smuggled compute estimated at 1.5 GW, yet Chinese frontier labs may hold only ~3% — teortaxesTex · 2026-11-13