---
title: "Moonshot accessed banned Nvidia chips, White House says"
author: "Asha Iyer"
datePublished: 2026-07-22T18:05:00.000Z
canonical: "https://digitalblog.com.au/post/00tijuo08fylm/moonshot-banned-nvidia-chips-white-house-2026"
---

White House science and technology director Michael Kratsios has accused Moonshot AI of using restricted Nvidia hardware and outside US model output to build Kimi K3, turning a model-launch story into a test of Washington’s AI export controls. The allegation, made in a [White House post against Moonshot AI](https://x.com/mkratsios47/status/2079933645888880708), centres on the Chinese startup’s [Kimi K3](https://apnews.com/article/kimi-k3-china-ai-0d8a5e268deb11a673f4d444fc597cc5), an open model that drew attention last week for its scale and benchmark claims.

In a separate [post on X](https://x.com/mkratsios47/status/2079933645888880708), Kratsios said the administration had information that Moonshot distilled Anthropic’s Fable and had acquired GB300-equipped servers or accessed GB300s in Thailand. Those claims have not been tested in public, and the record available to digitalblog does not show a Moonshot response. The case is being framed as an export-control and terms-of-service matter, not just another fight over model rankings.

> “We have information that Moonshot AI distilled Anthropic’s Fable for the development of its K3 model.”
>
> Michael Kratsios, on X

Last week, [AP reported](https://apnews.com/article/kimi-k3-china-ai-0d8a5e268deb11a673f4d444fc597cc5) Kimi K3 as a 2.8 trillion-parameter open model that had surprised parts of the market. [Axios argued](https://www.axios.com/2026/07/17/china-ai-kimi-k3-open-source-anthropic-opus) the release had narrowed the perceived gap between Chinese and US frontier labs. Kratsios’ claim adds no new benchmark evidence. It asks what inputs sat behind the model, and whether any were supposed to be outside Moonshot’s reach.

The chips named by Kratsios are not a routine supply-chain detail. If a Chinese lab reached restricted accelerators through offshore servers or intermediary access, enforcement becomes the hard part of the policy. Washington can restrict exports on paper, but AI training, model distillation and cloud infrastructure are spread across borders.

Nvidia was not accused of wrongdoing in Kratsios’ statement.

That is the break from last week’s model-race coverage. One question was whether Moonshot had produced a frontier-class open model. The new one is whether US-controlled inputs were used despite rules designed to keep them away from Chinese labs. Only the second points directly to compliance and enforcement.

Moonshot was already trying to turn Kimi K3’s launch into a commercial moment before the White House intervened. [Reuters reported on 20 July](https://d2233.cms.socastsrm.com/2026/07/20/chinas-moonshot-pauses-kimi-subscriptions-amid-hot-demand-ipo-push/) that the company had paused some Kimi subscriptions after heavier-than-expected demand and was pursuing an IPO, with management saying its GPUs were under strain. That made the story look like a product and funding surge. The allegation now makes it a supply-chain story as well.

> “Kimi K3 has received far more love than we expected, and our GPUs are feeling it.”
>
> Moonshot AI, via Reuters

Anthropic’s role matters because Kratsios named Fable rather than making a broad complaint about Chinese model development. That ties the case to model distillation as well as chips. For AI vendors, it raises a difficult enforcement question: how much can output controls, account rules and hardware restrictions be separated when labs train and serve models across several jurisdictions?

Moonshot has said it plans to release Kimi K3’s open weights on 27 July. If that timetable holds, scrutiny is likely to move from benchmark charts to the provenance of the compute and model inputs behind the release. Regulators, cloud providers and enterprise buyers will be watching the same issue from different angles.

For Australian readers, the significance is practical rather than distant. Local businesses buy Nvidia-backed infrastructure, use US model APIs and assess Chinese open-model releases. A dispute over where the chips sat and which model outputs were used is part of the same supply-chain contest shaping the tools Australian firms can buy and deploy.