DEV Community

Alex Chen
Alex Chen

Posted on

A Blackwell Fingerprint in Open Weights: What It Proves, and What It Doesn't

A Blackwell Fingerprint in Open Weights: What It Proves, and What It Doesn't

This week the timeline stopped arguing about which model is cheapest and started arguing about evidence again. On November 12, Will DePue at Unemployment Labs published a forensic report claiming that six open-weight frontier models from Chinese labs carry the fingerprint of NVIDIA Blackwell GPUs — chips that need a U.S. export license for China, and for which no licenses have been reported. The report's own words are careful: the fingerprint "supports, without proving" that Blackwell also did the training.

That careful sentence is the whole post. I have spent the last month writing about evidence standards for LLM routing — how many closed findings before you trust a number, what counts as closed, how a headline can move on one observation. This report is the same discipline, applied to the biggest question in AI hardware this year. So let us do what we do with routing claims: separate the measurement from the inference, count what the number is standing on, and say plainly what is still unknown.

What a fingerprint is

Every large model starts training from billions of pseudo-random initial values. In the PyTorch stacks most frontier labs use, those values come from Philox, a deterministic random-number generator, run through NVIDIA's cuRAND library. The work is split across the GPU's streaming multiprocessors (SMs): a sufficiently large initialization uses 2,048 threads per SM. Change the SM count and the same seed lands the same numbers in different positions.

The three counts that matter here:

Hardware class SMs Threads
Hopper H100 SXM / H200 / H800 132 270,336
Blackwell B200 / B300 148 303,104
Blackwell GB200 / GB300 152 311,296

Training changes most weights, but some survive untouched — rows tied to tokens the model rarely saw — and even trained weights keep a faint trace of their starting order. The report recovers the seed and the SM layout that reproduce those surviving values, then tests that combination on weights it held out of the search. A match on held-out weights is the difference between a finding and a coincidence.

Two details make this more than a clever search. First, DeepSeek reused one seed, 953049583, for both V3 and V4-Flash — so the same seed reproduces Hopper layout in the old model and Blackwell layout in the new one. Changing the seed by one reproduces nothing on either machine. Second, the authors verified on physical hardware: unmodified PyTorch on a B300 reproduced 461 of 469 selected rows in V4-Flash, 445 of them bit for bit; the same code on an H100 reproduced all 57,344 preserved values in V3 bit for bit, and neither chip reproduced the other model's rows. A rented GB200 reproduced all 13 preserved rows of Kimi K3. And engineer Peter Liu has since published a CPU-only script that recovers the same SM peaks and seeds for V4-Flash, Kimi K3, and MiniMax-M3 — no Blackwell rental required to check the check.

The table, and what each number is standing on

Lab Model Released Fingerprint Evidence
DeepSeek V4-Flash Apr 2026 148 SM, B200/B300 Exact copy, 1,385σ
DeepSeek V4-Pro Apr 2026 152 SM, GB200/GB300 Partial survival, 49σ
Moonshot Kimi K3 Jul 2026 152 SM, GB200/GB300 Exact copy, 305σ
Moonshot Kimi K3.1-Pro Nov 2026 152 SM, GB200/GB300 Exact copy, 305σ, same seed as K3
MiniMax MiniMax-M3 Jun 2026 148 SM, B200/B300 Exact copy, 323σ
Ant Group Ling-3.1-flash Oct 2026 152 SM, GB200/GB300 Untrained rows, 244σ

Every lab's earlier model still reads as Hopper: DeepSeek-V3 (Dec 2024), Kimi K2/K2.5 (through Jan 2026), MiniMax-M2.7 (Apr 2026), Ling-3.0-flash (Aug 2026). And not every lab flipped — the newest models checked from Alibaba (Qwen), Zhipu (GLM), Tencent, Xiaomi, and StepFun still read as Hopper. The switch is recent, it is visible inside each lab's own lineage, and it is not universal. That pattern is much harder to explain away than a single hit.

The sigma numbers need the same treatment I give success rates: ask what they are standing on. The search covered roughly six trillion seed-and-layout combinations. In a search that large, the strongest accidental match would be expected to reach about 8 sigma. The strongest evidence for each of the six models ranges from 49 to 1,385 sigma, and even the weakest individual test reported reaches 20 sigma. That is the hold-out logic from my last post, applied at industrial scale: the seed and layout were chosen on one set of weights and tested on another that played no part in choosing them.

The method also gets known answers right, which is the calibration check that matters most. In 24 cases where a developer publicly named the training hardware, the method agrees every time: Meta's Llama 3.1 on H100s reads as 132 SMs, EleutherAI's Pythia on A100s reads as 108, NVIDIA's own Nemotron-3 Ultra — whose technical report names the GB200 for one phase — reads as 152, the same as Kimi K3. A method that reproduces 24 known answers earns the right to be heard on the unknown ones.

What it proves

It proves, for DeepSeek-V4-Flash conclusively in the authors' words, that Blackwell-class hardware generated the starting weights. The evidence is strongest exactly where it should be: an exact copy, reproduced on a physical B300, with the seed reused from the previous generation, at 1,385 sigma against an 8-sigma chance ceiling.

It also proves the AMD alternative does not explain V4-Flash. An MI300X can be split into two groups of 152 processing units, or limited to 148, and its random-number software can generate the same underlying sequences — so the layout alone does not settle the question. The authors tested it on a physical MI300X: in the 377 values where AMD's and NVIDIA's arithmetic differ, the published weight matches NVIDIA in 363 and AMD in none. For the other five models the same test could not settle it, and the authors call the AMD explanation "very unlikely" rather than excluded — the distinction is stated, not smoothed over. Note the irony for anyone who runs AMD hardware: the chip that could theoretically fake the fingerprint is the one we measured DeepSeek-V4.1-Flash on ourselves, at $0.016 per million tokens on a real MI300X (SemiAnalysis InferenceX, November 6 run, vLLM, FP4).

What it does not prove

It does not prove Blackwell did the training. The fingerprint is set at step zero. A lab could initialize weights on one system and train on another, or continue training an existing model on new hardware. Kimi K2.7-Code, posted six weeks before K3, is the live example: mostly Hopper, with a small part matching the 148-SM layout. Mixed clusters blur the read the same way. The report's sentence is the honest one: the fingerprint "supports, without proving" that Blackwell also performed the training.

It does not prove where the chips were, who owned them, or that any law was broken. Blackwell systems can be rented from data centers outside China. The weights cannot reveal geography or ownership, and the report states plainly that its findings "do not establish where the chips were located, how access was obtained, or whether export controls were violated."

It does not prove anything about DeepSeek-V4.1-Flash — and that is the most instructive part. The authors searched exhaustively and found no fingerprint in V4.1-Flash. That is a null result, not a clean bill of health. The method needs a recognizable, hardware-dependent pattern in the starting weights. Weights initialized on a CPU leave no GPU trace — several Alibaba Qwen3 models were initialized that way. Baidu's PaddlePaddle framework leaves none either; the authors recovered ERNIE-5.5's seed (2760) but could not identify its hardware. V4.1-Flash's hardware is an open question until DeepSeek documents its seeding stack or the authors publish Ascend generator tests. Any claim that the report "confirmed" Ascend training for DeepSeek is unverified — the report's only Ascend reference is a deleted fine-tuning configuration in Meituan's LongCat-2.0, read as Huawei Ascend 910.

In the language of my last post, V4.1-Flash sits in the third bucket: timed out, still unclosed. An unclosed finding is not a success and not a failure. It is an unknown, and it must be displayed as unknown — never folded into a column to make a story look complete.

The scale math — and the revision that makes it credible

From model sizes and token counts, the report estimates the final training runs at roughly 1,500 Blackwell GPUs for V4-Flash (50 days), 3,000 for V4-Pro (70 days), and 4,500 for Kimi K3 (70 days) — about 12,000 GPUs across the five runs, with the largest cluster costing roughly $200–300 million to buy. The ranges are wide: K3's is 1,000 to 9,500 GPUs. The authors say so themselves, and list the assumptions the weights cannot confirm: utilization, token counts, run lengths, and that training stayed on the initialization hardware.

Here is the detail that actually increased my trust: the authors revised these numbers on the same day. The first version put 2,000–3,000 GPUs on the largest runs and $150–200 million on the largest cluster; the revision raised both after the authors lowered their assumed efficiency, and the revision is dated on the page — "Revised November 12, 2026" — with the old figures still visible. That is the backfill discipline I argued for two weeks ago: backfill corrects the record; it does not rewrite the published numbers. A dated revision with the old numbers preserved is a correction. A silent edit is a rewrite. The report does the former, in public, on day one.

What would renting that run cost at today's prices? V4-Flash's final run — 1,500 GPUs for 50 days — is 1.8 million GPU-hours. At Lambda's H100 SXM rate, which rose to $4.59 per GPU-hour on November 16 (from $3.99), the same hours on B200-class hardware at $7.69 run about $13.8 million; the report's range for the run, 250 to 5,000 GPUs, maps to roughly $2.3–46 million at that rate. Renting the run is an order of magnitude cheaper than owning the cluster the report prices at $200–300 million — which is exactly why "access to hardware" and "ownership of hardware" are different claims, and why the report refuses to collapse them. (For scale: the cheapest H100 hour we track is still RunPod's Community Cloud at $2.69, re-checked November 13. Rental prices are moving weekly; check the day you decide, not the day you read this.)

The same rule, applied to ourselves

This post is written on closed findings, in the sense my series has been building toward. The report is public, reproducible, and dated; the per-model numbers are in its appendix; the limits are in its own limitations section; the revision history is on the page. As of today, no lab, NVIDIA, or the Commerce Department has responded to the report specifically — that silence is a fact about the record, not evidence for either side, and the next forensic check is the next DeepSeek release.

Our own funnel gets the same treatment. Google Search Console, 28 days to November 16: 0 clicks, 47 impressions, and our GPU cost calculator still records zero impressions — it does not appear in the page table at all. That zero is a fact about the window, not a verdict on the page. The discussion that produced this series is the same discipline applied to our own writing: the question that started it arrived when the thread under my November 6 post had 9 comments; by today that article has 66, and across our 10 published articles we have 95 comments and 6 reactions. This morning a reader, Tom Jones, pushed us one step further — every threshold should carry its source, its reading index, and its expiry, so a guess never hardens into a constant — and we answered it the same day. A post written on 9 comments would have been a median of three.

The bottom line

Do not ask "did they train on Blackwell?" Ask three smaller questions, in order: what hardware generated the starting weights, how strong is the evidence for that, and what would change your mind? For V4-Flash the answers are: Blackwell, 1,385 sigma on held-out weights reproduced on a physical B300, and a successful independent reproduction — which already exists, on a laptop, from Peter Liu's CPU-only script. For the training claim itself, the answer is still: strongly suggested, not proven. And for V4.1-Flash, the answer is: unknown, and anyone who tells you otherwise is selling you a conclusion the evidence does not support.

That is the standard I hold our routing numbers to — a 30-closed-finding floor before a routing change, a flip count instead of a percentage, an attribution window that has to close before a number counts. It is the same standard the best forensic work in AI hardware is holding itself to this week. The fingerprint is real. The rest is still evidence, not verdict.

If you want to price the trade-off yourself — what a routing change costs at today's GPU rates before you make it — our free GPU cost calculator runs the same arithmetic: https://qisuanai.com/gpu-cost-calculator/


Sources, all checked November 17, 2026: Will DePue, "Top Chinese Frontier AI Models Carry the Fingerprint of Export-Controlled NVIDIA Blackwell GPUs," Unemployment Labs, November 12, 2026, https://unemploymentlabs.com/checkpoint-forensics (report page, including the November 12 revision note; Table 1, Table 2, Appendix 1, and the limitations section). Explainx.ai write-up with the November 13 updates (TechCrunch recap, Peter Liu's CPU-only reproduction, V4.1-Flash null result), https://explainx.ai/blog/chinese-open-weights-blackwell-gpu-checkpoint-forensics-november-2026. No statement from DeepSeek, Moonshot, MiniMax, Ant Group, NVIDIA, or the Commerce Department in response to the report had been found as of today. Prices: Lambda H100 SXM $4.59 (8x) / $4.89 (1x) and B200 $7.69, effective November 16, 2026, verified from Lambda's pricing page this morning; RunPod Community Cloud H100 $2.69, re-checked November 13. Our own measurement: DeepSeek-V4.1-Flash on MI300X, $0.016 per million tokens (hyperscale own price; $0.022 retail rental), 16,368 tokens/sec per card at 50 tokens/sec per user, SemiAnalysis InferenceX run of November 6, 2026, vLLM FP4 on real MI300X — see our November 10 video. DeepSeek API prices per DeepSeek's Models & Pricing page as reported November 9: deepseek-flash $0.30 input / $1.20 output per million tokens at peak, half off-peak; cache hits $0.006 / $0.003. The rental arithmetic (1,500 GPUs × 50 days × 24 hours × $7.69) is mine, on the report's point estimate — the report's own range is 250–5,000 GPUs, and GB200-class rental is not the B200 rate, so treat it as an order-of-magnitude comparison, not a quote. GSC numbers are our own Search Console pull for sc-domain:qisuanai.com, 28 days October 20–November 16, 2026. The 30-finding floor, 10-finding watchlist threshold, and 7-day attribution window are our operating rules, stated as such; the 8-week replay that would calibrate them has not run yet. For the fuller cost picture, our GPU cost guide: https://qisuanai.com/blog/gpu-cost-guide-2026/

Top comments (0)