What happened
TechCrunch reported on November 13, 2026, that Will DePue, a former OpenAI researcher who leads the nonprofit research group Unemployment Labs, analyzed the weights of open models released by DeepSeek, Moonshot AI, MiniMax and Ant Group. Weights encode a model’s learned structure; DePue’s analysis looked for traces that could reveal the configurations of chips used during training.
The article says DePue found what it describes as compelling evidence that some closely watched models, including Moonshot’s Kimi K3 releases and DeepSeek’s V4 models, bear fingerprints of training on Nvidia Blackwell chips. The article does not describe the analysis as proof that every training run used Blackwell, or that a model was trained entirely on those GPUs. Blackwell chips are restricted from export to China and Chinese-owned firms under US controls.
TechCrunch notes an important alternative: Chinese companies may have accessed chips in offshore data centers. It says this route has become popular for Chinese firms seeking compute and has helped drive data-center activity in Malaysia. The distinction matters because an export restriction based on delivery destination may not settle who can remotely use the installed hardware.
The article cites Epoch AI estimates that about two million advanced chips entered China in 2024 and 2025 through schemes intended to conceal their final destination, and that 275,000 Nvidia AI accelerators were imported into Malaysia over the same period. Those figures concern broader chip flows; they do not identify the hardware used for the specific model runs. TechCrunch also says Chinese firms could have used hardware in an offshore facility without proving that export-control laws were violated.
The article reports that White House technology adviser Michael Kratsios had earlier said Moonshot used export-banned Nvidia chips, citing US intelligence sources. It describes a policy debate over whether controls should track who uses a chip, rather than only its owner or physical location. TechCrunch says US companies currently have limited obligations to track what customers do with controlled goods such as AI chips. It also notes reports of roughly 600,000 H200 chips stockpiled in Hong Kong awaiting shipment into China, though observers cited by TechCrunch disagree on why imports have not been cleared.
What the archive adds
New Runtime's earlier coverage on November 14, 2026, reported details about Nvidia’s objections to proposed H200 export-license conditions and how Jensen Huang’s relationship with Donald Trump figured in policy discussions (Semafor: Nvidia's Huang used his Trump relationship to shape China chip-export rules). This is a new fact on a known story: the earlier note concerned policy and licensing, while this report describes a claim about model-training compute.
Trend signal
The report sharpens the policy question of whether advanced compute controls can work when access may be routed through offshore infrastructure. It does not establish that model providers can freely swap between compute providers or that training provenance is reliably observable. This does not yet fit a pattern New Runtime tracks.
Why it matters
For teams building AI systems, compute provenance can affect supply planning, compliance review and the credibility of claims about how a model was trained. The article’s central distinction is operationally important: a model-weight fingerprint may point to a hardware family, but it does not by itself establish the data-center location, ownership chain, customer, or legal status of a training run.
That distinction should temper both procurement assumptions and policy conclusions. If firms depend on models trained with constrained hardware, they may face uncertainty not just about future chip availability but also about which controls apply to the compute path. Conversely, a finding about hardware traces should not be treated as a verified violation. The report highlights a gap between controlling chip shipments and observing where advanced compute is ultimately used.
For builders evaluating open-weight models, this is not a performance claim and does not change the model-selection process on its own. It is a reminder that training provenance is a separate operational attribute: model cards, vendor statements, hosting geography and contractual representations can all matter when a deployment has regulatory, customer, or supply-chain constraints. Those questions remain relevant even when a model’s weights can be inspected.
Limits and uncertainty
TechCrunch presents the Blackwell finding as an analysis by Unemployment Labs; the article does not provide enough methodological detail to independently reproduce it, nor does it quantify how much of each named model’s training used Blackwell. The article characterizes the evidence as fingerprints in weights, but does not publish a test protocol, confidence thresholds, comparison hardware set, or model-by-model results in the account.
The analysis does not establish that chips were located in China, identify the facilities or operators involved, or prove an export-control violation. It also does not show whether the relevant hardware was purchased, leased, remotely accessed, or used for all versus part of a training run. Epoch AI’s estimates describe wider chip flows rather than the specific training runs. TechCrunch also reports allegations and investigations involving intermediaries, but those claims do not establish responsibility by Nvidia or any named executive.
Reader takeaway
The reported model-weight fingerprints are a signal that export controls may not map cleanly onto where advanced AI compute is used. They are not proof of a specific chip’s route or an unlawful training run. Treat the finding as a reason to ask harder questions about compute provenance, not as a settled account of how these models were trained.