Moonshot AI's frontier model, the biggest open-weight model ever, was allegedly trained on NVIDIA Blackwell hardware acquired around both US export controls and China's own import restrictions. The compute-smuggling story behind the model.
The story behind Kimi K3 just got more complicated. Moonshot AI's 2.8-trillion-parameter model — the biggest open-weight model ever released — was reportedly trained on NVIDIA Blackwell chips that were acquired by circumventing both US export controls (which bar the latest NVIDIA chips from China) and China's own import restrictions. If accurate, it's the clearest evidence yet that compute controls are leaking, and it reframes the 'China caught up on AI' narrative: the catch-up may owe as much to smuggled hardware as to algorithmic ingenuity. Here's what the reports say and why it's a bigger deal than one model.
Why this matters beyond one model
US export controls on advanced AI chips rest on a simple bet: that restricting access to the best hardware slows rivals' AI progress. If a Chinese lab trained a frontier open model on smuggled Blackwell chips, that bet is leaking — and Kimi K3 isn't a minor model. It topped the Frontier Code Arena leaderboard and is the largest open-weight model anyone has shipped. The reports fit a broader pattern of NVIDIA-hardware smuggling stories, including an ongoing probe reportedly implicating an NVIDIA employee. Whether controls can actually be enforced against determined, well-funded buyers is now an open question — and Kimi K3 is the highest-profile data point that they may not be.
If a frontier open model was trained on smuggled Blackwell chips, the premise of compute export controls is leaking. · Unsplash
The hardware reality hasn't changed
One thing the smuggling story doesn't change: you still can't run Kimi K3 at home. However it was trained, the finished model is 2.8 trillion parameters and needs ~1.4TB of memory to load — a datacenter workload regardless of where the training chips came from. The training-hardware controversy is about geopolitics and enforcement; the inference reality is about physics. If you want to use K3, you'll do it through an API or a host, and if you want frontier-adjacent capability on your own hardware, the runnable open models it inspires are the practical path. The smuggling story is fascinating, but it's a separate question from what you can run.
Quick answers
Was Kimi K3 really trained on smuggled chips?
Reports indicate Moonshot AI trained Kimi K3 on NVIDIA Blackwell chips acquired around both US export controls and China's import restrictions. These are reports rather than confirmed, documented fact, and Moonshot hasn't detailed its training hardware publicly. But the claims fit a broader pattern of NVIDIA-hardware smuggling into China, and Kimi K3's scale (2.8T parameters) implies access to substantial cutting-edge compute. Treat it as credible reporting on a story still developing.
Do export controls on AI chips actually work?
That's exactly the question this story raises. The controls aim to slow rivals' AI progress by restricting access to the best hardware, but if a frontier Chinese model was trained on smuggled Blackwell chips, they're leaking. Enforcement against well-funded buyers using intermediary networks is genuinely hard. The controls likely raise costs and friction rather than fully blocking access — Kimi K3, if the reports hold, suggests determined actors can still get frontier hardware.
Can I run Kimi K3 regardless of how it was trained?
Not on consumer hardware — the training-chip controversy doesn't change the inference reality. Kimi K3 is 2.8 trillion parameters and needs roughly 1.4TB of memory to load, which is a datacenter requirement no matter where the training chips came from. To use it, you'd go through an API or a hosted service. For frontier-adjacent capability you can actually run at home, look to the smaller open models, which run on a single GPU.
If the reports hold, Kimi K3's training on smuggled Blackwell chips is the strongest sign yet that AI compute controls are leaking — a geopolitical story layered on top of an already-remarkable model. Sources: Tom's Hardware and Moonshot. For what the model actually takes to run, see our hardware breakdown.
