AI/Nov 17 edition/Reporting & analysis
PolicyModelsInfrastructureBusiness

PolicyLaw, regulation & governance

Open-source backers push for Jensen Huang's return as Blackwell traces surface in Chinese open-weight models

A former OpenAI researcher says weight-initialization patterns show six Chinese open-weight models began training on Nvidia Blackwell GPUs barred from China, as backers of open models campaign to reinstate Jensen Huang and Washington tightens rules on remote compute.

From the MIT Technology Review · AI article of November 16, 2026 and cited sources: Will DePue reports finding Blackwell signatures in six Chinese models: DeepSeek-V4-Flash and V4-Pro, Kimi K3 and K3.1-Pro, MiniMax-M3, and Ant Group's Ling-3.1-flash. [6] [7]

SOURCES[6] The Decoder: Blackwell fingerprints in six Chinese models the-decoder.com[7] The Register: Chinese AI labs training on Blackwell theregister.com

AI-assisted, not human-edited. Methodology Corrections Who owns this site

KEY FACTS
What happened
Will DePue reports finding Blackwell signatures in six Chinese models: DeepSeek-V4-Flash and V4-Pro, Kimi K3 and K3.1-Pro, MiniMax-M3, and Ant Group's Ling-3.1-flash. [6] [7]
Original source
MIT Technology Review · AI article, independent perspective
Brief by
AT Pulse research desk AI-assisted; publisher and contact on the desk page
When
November 16, 2026 article published
Evidence
12 sources across 10 sites

Reach: not measured yet

Illustration from MIT Technology Review: Open-source backers push for Jensen Huang's return as Blackwell traces surface in Chinese open-weight models
Image: MIT Technology Review — Original article ↗
THE CORE IDEAS4 TAKEAWAYS
01

Will DePue reports finding Blackwell signatures in six Chinese models: DeepSeek-V4-Flash and V4-Pro, Kimi K3 and K3.1-Pro, MiniMax-M3, and Ant Group's Ling-3.1-flash. The method relies on how PyTorch arranges its seeded random initial weights. That arrangement depends on the GPU's streaming-multiprocessor count, and DePue says it survives training. He published the seeds and code, and reproduced two of the results on real Blackwell hardware. Older models such as DeepSeek-V3 show Hopper patterns instead. [6] [7]

02

The fingerprint shows which hardware generated a model's starting weights. It does not show where the chips were, who owned them or whether any law was broken. Legally renting GPUs abroad would explain it equally well, and nothing shows Nvidia helped Chinese firms get the chips. [1] [10]

03

A new Commerce interim final rule requires licences for large data centres that serve companies headquartered in China or other D:5 countries. It applies above 12 million TPP, roughly 750 H100s, and gives existing facilities 90 days to get licences. It also adds licence requirements for chip exports to Malaysia, Singapore, Indonesia, Thailand and Brazil. [9]

04

A petition backed by Elon Musk, Garry Tan, Tim Sweeney and Hugging Face's Julien Chaumond calls for Huang to be reinstated. His return still looks unlikely: Nvidia's board opposes it, Treasury Secretary Bessent wants an outside CEO, and prediction markets put the odds near 10%. [1] [8] [11]

WHY IT MATTERS

The published code lets anyone with a model's weights test where it was initialized. If the findings hold, they weaken the argument that open-weight releases strengthen American AI.

Read the full assessment

The new rule could also slow Chinese model progress and adds licensing work for cloud providers in five countries.

Executive brief

Chinese labs appear to have trained flagship open-weight models on Nvidia Blackwell GPUs, which the US bars from export to China. Ex-OpenAI researcher Will DePue says the evidence sits in the models' published weights. His analysis came out the same week Jensen Huang resigned as Nvidia CEO over reported affairs, including one with an alleged Chinese spy. Open-source backers, among them Hugging Face leadership and venture capitalists, are pushing to bring Huang back. MIT Technology Review argues the chip findings, and a new Commerce rule on remote compute, weaken both his chances and the open-weight movement's argument that it strengthens American AI.

What changed and event timeline

  1. Nvidia agrees to buy Hugging Face

    The deal is worth $12.93B: about $11.9B in cash plus up to $1B in retention awards. Nvidia expects to close by 2027, pending antitrust review (,).

  2. Huang resigns

    The resignation followed reports of an affair with an alleged Chinese spy and FBI and congressional investigations. CFO Colette Kress became interim CEO, and Harvey Jones became Nvidia's first board chair (,).

  3. DePue publishes "Blackwell fingerprints."

    He says six Chinese models were initialized on Blackwell hardware. MIT Technology Review dates the post to the day Huang resigned; the trade press dated their reports Nov 13 (,).

  4. "Stand with Jensen" petition spreads

    Elon Musk, Garry Tan, Tim Sweeney and Hugging Face's Julien Chaumond back it. Reported totals range from more than 115,000 to about 180,000 signatures (,).

  5. Commerce restricts remote compute

    The interim final rule requires licences for large data centres serving companies headquartered in China or other D:5 countries. It also adds licence requirements for chip exports to Malaysia, Singapore, Indonesia, Thailand and Brazil ().

Capabilities and access

According to The Decoder, these models carry Blackwell signatures: Older models (DeepSeek-V3, Kimi K2.5) show Hopper signatures instead. The Commerce rule applies above 12 million TPP, roughly 750 H100/H200s. Existing data centres get 90 days to obtain licences. Recommendation systems are excluded (TechCrunch).

Read the full section

According to The Decoder, these models carry Blackwell signatures:

  • DeepSeek-V4-Flash and V4-Pro
  • Kimi K3 and K3.1-Pro
  • MiniMax-M3
  • Ant Group's Ling-3.1-flash

Older models (DeepSeek-V3, Kimi K2.5) show Hopper signatures instead. The Commerce rule applies above 12 million TPP, roughly 750 H100/H200s. Existing data centres get 90 days to obtain licences. Recommendation systems are excluded (TechCrunch).

Technical analysis for researchers and developers

  • How the fingerprint works: PyTorch initializes weights with a seeded pseudorandom generator (Philox). Per DePue, this pattern survives training (The Decoder).
  • Search and reproduction: DePue searched more than 6 trillion combinations of chip layout and seed.
  • What it means for developers: Released weights can reveal the hardware used to initialize them.
Read the full section
  • How the fingerprint works: PyTorch initializes weights with a seeded pseudorandom generator (Philox). How those random values are laid out depends on the GPU's streaming-multiprocessor count: 132 on Hopper, 148 on Blackwell B-series, 152 on Grace Blackwell. Per DePue, this pattern survives training (The Decoder).
  • Search and reproduction: DePue searched more than 6 trillion combinations of chip layout and seed. He published the seeds and code, and reproduced the DeepSeek-V4-Flash and Kimi K3 results exactly on real Blackwell hardware (The Register).
  • What it means for developers: Released weights can reveal the hardware used to initialize them. That makes provenance forensics possible for anyone with the weights.

Claims and evidence

  • Claim: the training runs used about 12,000 Blackwell GPUs, comparable to about 30,000 H100s, at a cost above $500M.
  • Claim: senior OpenAI and Anthropic researchers confirmed the results. This rests on DePue's word. MIT Technology Review says no independent analysis has confirmed the findings.
  • Separate line of evidence: SemiAnalysis estimates that more than 1.5 GW of export-restricted Nvidia chips reached China.
Read the full section
  • Claim: the training runs used about 12,000 Blackwell GPUs, comparable to about 30,000 H100s, at a cost above $500M. These are DePue's own estimates (The Decoder).
  • Claim: senior OpenAI and Anthropic researchers confirmed the results. This rests on DePue's word. MIT Technology Review says no independent analysis has confirmed the findings.
  • Separate line of evidence: SemiAnalysis estimates that more than 1.5 GW of export-restricted Nvidia chips reached China. It has published no unit or generation breakdown (N環球).
  • Petition counts: Totals come from the campaign itself. That includes the claim that about 9,000 signers are Nvidia employees (CBS News).

Context and prior work

  • Huang publicly championed open source as central to US AI leadership (MIT Technology Review).
  • Blackwell sales to China had already stalled in 2025 when a trade deal failed to materialize (The Register, 2025).
  • Until this rule, renting GPUs in Southeast Asian data centres was a legal route for Chinese firms (TechCrunch).

Limitations, safety and contested findings

  • What the fingerprint does and doesn't show: It identifies the hardware that generated the starting weights.
  • No evidence of Nvidia's role: Nothing shows Nvidia helped Chinese firms get the chips (MIT Technology Review).
  • No company responses: Nvidia and DePue did not respond to MIT Technology Review.
Read the full section
  • What the fingerprint does and doesn't show: It identifies the hardware that generated the starting weights. It does not show where the chips were, who owned them or whether any law was broken. Legal rental abroad would explain it equally well (N環球).
  • No evidence of Nvidia's role: Nothing shows Nvidia helped Chinese firms get the chips (MIT Technology Review). Separately, Bessent asked whether Nvidia ignored warning signs about Singapore shell companies (CBS News).
  • No company responses: Nvidia and DePue did not respond to MIT Technology Review.

Business and practitioner implications

  • Nvidia's board opposes it, Bessent wants an outside CEO, and prediction markets put a return at about 10% (CBS News).
  • Cloud and colocation providers: Those in the five newly listed countries face licensing work within the 90-day window.
  • Teams building on Chinese open-weight models: The models may improve more slowly if offshore compute dries up (MIT Technology Review).
Read the full section
  • Huang's return looks unlikely. Nvidia's board opposes it, Bessent wants an outside CEO, and prediction markets put a return at about 10% (CBS News). Arm's Rene Haas was reportedly approached as a successor (CNBC).
  • Cloud and colocation providers: Those in the five newly listed countries face licensing work within the 90-day window.
  • Teams building on Chinese open-weight models: The models may improve more slowly if offshore compute dries up (MIT Technology Review). These teams also face growing policy scrutiny.
  • Hugging Face deal: Antitrust review is still pending while Nvidia lacks a permanent CEO.
FOLLOW THE EVIDENCE

The source trail.

Sources (12)
A LITTLE LESS NOISE. A LOT MORE CONTEXT.

Stay curious.
Follow the evidence.

Independent perspectives, the original sources, and room for the questions that don't have easy answers.

How we build the brief
Connect with us

Find us where you already read.