Onpode
Cover art for Moonshot acquired 20,000 Nvidia chips through Alibaba to build an AI rival to US labs

Moonshot acquired 20,000 Nvidia chips through Alibaba to build an AI rival to US labs

July 31, 2026 · 9 min

Walt Garner & Nina Park

Chinese AI startup Moonshot AI trained its Kimi K3 model — 2.8 trillion parameters, ranked #1 on Arena.AI above Anthropic — using a cluster of roughly 20,000 Nvidia GPUs accessed through a confidential computing deal with Alibaba, which also holds a 36% stake in Moonshot, raising questions about whether US export controls accelerate Chinese AI consolidation rather than block it.

On July 31, 2026, Bloomberg reported that Chinese AI startup Moonshot AI has a confidential computing agreement with Alibaba Group providing access to a cluster of approximately 20,000 Nvidia GPUs — a substantial portion of the compute underpinning its Kimi large language models.

0:009:05
Get the next episode on Nvidia

Follow it free — new episodes land in your feed.

Or make your own — any topic, in minutes

More Onpode episodes on Nvidia

About this episode

In July 2026, Bloomberg reported that Moonshot AI — the Beijing startup behind the Kimi chatbot — trained its latest model on a cluster of roughly 20,000 Nvidia H200 GPUs, sourced through a confidential computing deal with Alibaba. The result, Kimi K3, is a 2.8-trillion-parameter open-source model that Arena.AI ranked number one globally, above Anthropic and above Alibaba's own Qwen family. Alibaba, which denied the specific H200 designation but did not deny the deal, also owns 36% of Moonshot and is structurally its exclusive compute supplier. This episode works through what that arrangement actually means. The easy headline is that export controls failed. The more unsettling possibility is that they're working exactly as a filter — one that concentrates capability inside the handful of Chinese labs that can negotiate with an Alibaba-scale anchor investor. That's the Cold War COCOM pattern: the workaround doesn't eliminate the barrier, it makes the barrier expensive enough that only incumbents clear it. The episode also holds a careful line between what Bloomberg sourced and what remains White House allegation — the Blackwell GB300 claims are treated separately from the documented H200 arrangement, and that distinction matters. The open-source release of Kimi K3 changes the frame entirely. Once the weights are out, the chip infrastructure that produced them becomes irrelevant to their spread. The episode ends without a tidy answer, which feels right: Alibaba is simultaneously investor, compute supplier, and direct competitor to the model it helped build, and nobody yet knows whether that makes Chinese AI structurally stronger or quietly fragile.

Frequently asked

How did Moonshot AI get Nvidia chips despite US export controls?

Moonshot AI accessed roughly 20,000 Nvidia GPUs through a confidential computing agreement with Alibaba Cloud, according to Bloomberg. Alibaba denied the specific H200 chip designation but did not deny the deal. Alibaba also holds a 36% equity stake in Moonshot, making it simultaneously investor and compute supplier.

What is Kimi K3 and how does it compare to other AI models?

Kimi K3 is a 2.8-trillion-parameter open-source AI model released by Moonshot AI on July 17th. Upon release, Arena.AI ranked it number one globally, placing it above Anthropic's Claude and Alibaba's own Qwen model family — a notable outcome given Alibaba both funds Moonshot and competes directly through Qwen.

Did US export controls on Nvidia chips stop China from building frontier AI?

US export controls did not stop Moonshot AI from training Kimi K3 on roughly 20,000 Nvidia GPUs. Analysts argue the controls transformed — not blocked — Chinese AI development, making access slower and more expensive but ultimately favoring well-capitalized incumbents like Alibaba that can structure compute deals around the restrictions.

Can Kimi K3 be downloaded and used outside China?

Yes. Kimi K3's weights are publicly released and downloadable. Because the model is open-source, the 20,000-GPU cluster used to train it is irrelevant to its spread — quantized versions can run locally, and no export control mechanism can restrict access to already-released model weights on platforms like Hugging Face.

What conflict of interest does Alibaba have with Moonshot AI?

Alibaba holds a 36% stake in Moonshot AI, supplies its compute infrastructure through Alibaba Cloud, and directly competes with Moonshot through its own Qwen model family — with no disclosed firewall between those three roles. Kimi K3's Arena.AI ranking above Qwen makes that conflict structurally visible.

Grounded in 12 sources
Watch Moonshot's Kimi Built With Nvidia Compute · bloomberg.com
The faces behind China's AI rise: Meet the founders of DeepSeek, Z.ai and ... · channelnewsasia.com
Moonshot has Nvidia chip cluster from Alibaba computing deal, Bloomberg News reports - CNA · channelnewsasia.com
Bloomberg: Moonshot used 20K Nvidia chips via Alibaba to rival US · finance.yahoo.com
China’s Moonshot, Z.AI, and DeepSeek are challenging U.S. AI labs—and beating them on cost · finance.yahoo.com
Moonshot has Nvidia chip cluster from Alibaba computing deal, Bloomberg ... · reuters.com
Moonshot AI unveils world's largest open-source AI model ... · scmp.com
What is China’s Moonshot AI and why is it roiling markets? | The Straits Times · straitstimes.com
China's Moonshot, Z.AI, and DeepSeek are challenging ... · fortune.com
Moonshot trained Kimi K3 on 20,000 Nvidia chips from Alibaba’s cloud · thenextweb.com
Who Owns Moonshot AI? Backers, Valuation, Money · capitalandcompute.net
Moonshot AI - Products, Competitors, Financials, Employees, Headquarters Locations · cbinsights.com
Read transcript

Nina Park: Long week — I've been reading about a Carnegie Mellon-trained guy who named his AI startup after a Pink Floyd album, and somehow that's become the most geopolitically loaded story of the summer.

Walt Garner: Yang Zhilin and Moonshot AI — yes, I've been living in this story as well.

Nina Park: So Bloomberg just reported that Moonshot AI got access to roughly twenty thousand Nvidia GPUs through a confidential computing agreement with Alibaba — and multiple sources told Bloomberg those are Hopper-generation H200 processors. While export controls are in effect. That's the headline.

Walt Garner: Hold on — Bloomberg had named sourcing on the H200 designation specifically?

Nina Park: Multiple sources, yeah — Alibaba denied the H200 label but didn't deny the deal itself. And then Kimi K3 drops July 17th, 2.8 trillion parameters, open-source, and Arena.AI immediately ranks it number one globally — above Anthropic, above Alibaba's own Qwen.

Walt Garner: You see, that last detail is the one that genuinely stops me — Alibaba holds thirty-six percent of Moonshot AI and is simultaneously Moonshot's compute supplier, and the product of that arrangement just outranked Alibaba's own model family. If I'm Alibaba's board, I'm asking a very uncomfortable question about what, precisely, we are optimizing for.

Nina Park: Right — but is it actually uncomfortable, or is it the plan? Because that's what I can't figure out.

Walt Garner: That is precisely what we're trying to work out today.

Nina Park: But here's what I keep tripping over — is the 'plan' even one thing? Because there's the documented deal, and then there's the White House saying Moonshot got Blackwell GB300 chips through Thailand, and I don't think those are the same story at all.

Walt Garner: That distinction is precisely where I want to pump the brakes. What Bloomberg sourced — multiple named parties — is the Hopper-generation H200 arrangement. Twenty thousand GPUs, Alibaba Cloud, partly acknowledged. That is documented. The Blackwell GB300 claim comes from White House allegations. Unconfirmed. Moonshot declined to comment. Nvidia declined to comment. Those are two entirely different evidentiary situations.

Nina Park: Hold on — both companies just... said nothing?

Walt Garner: Which is, you see, not the same as a denial. But it is also not confirmation. And when a geopolitically convenient allegation arrives without independent corroboration, the responsible thing is to hold it separately from what we actually know.

Nina Park: Okay, but — actually, I want to test whether the distinction matters. Because if the outcome is identical — Moonshot scales to frontier, Kimi K3 beats Anthropic on Arena.AI — does it change anything whether the chips were Hoppers or Blackwells?

Walt Garner: It matters enormously for what we claim to know, and here is the plain version of why. Think of Alibaba as a landlord who also owns thirty-six percent of your company and — wait, actually the fuller version — who also requires, as a structural feature of the investment, that you rent exclusively from their cloud infrastructure. The captive rental relationship is not in dispute. Whether the apartment is a standard unit or the luxury penthouse — the Blackwell question — that is very much in dispute. You do not want to accuse someone of the penthouse when you can only prove the standard unit.

Nina Park: The landlord thing — yeah, that actually lands. And Alibaba's own denial kind of proves the structure, doesn't it? They denied the H200 label specifically. Not the deal.

Walt Garner: Precisely so. Alibaba disputed the specific chip designation while leaving the compute relationship entirely intact. Which tells you the relationship is not embarrassing to them — the export-control optics of a particular GPU model number is what they're managing.

Nina Park: So the new thing — the actual headline — isn't 'China got the chips.' It's that the investment structure itself is the workaround, and Yang Zhilin building Moonshot in March 2023 already had that baked in from day one.

Walt Garner: And that's where I want to push back on the take that's circulating everywhere right now — 'export controls failed.' That is, I think, precisely the wrong frame.

Nina Park: Wait — you're saying they didn't fail?

Walt Garner: They transformed who can compete. That is a different thing entirely. DeepSeek in early 2025 — matching US frontier models on constrained compute — that was the first signal. Moonshot assembling twenty thousand GPUs to train a 2.8-trillion-parameter model is the confirmation. The controls didn't block the destination; they made the path slower, more expensive, and politically entangled.

Nina Park: Okay but 'slower and more expensive' — isn't that still a meaningful constraint? Like, that's what controls are supposed to do.

Walt Garner: It is — if you don't have Alibaba's balance sheet. That's the mechanism. The expense and the political entanglement filter out every Chinese AI startup that can't get a thirty-six percent anchor investor who also operates the largest cloud infrastructure in the country.

Nina Park: So the controls are — wait, actually this is the part that breaks my brain — the controls are accelerating consolidation. They're making Alibaba more powerful inside Chinese AI, not less.

Walt Garner: Which is, you see, exactly how COCOM worked in the Cold War — it didn't stop Soviet computing, it made them vertically integrated. The workaround requires capital and state relationships that only incumbents possess. The Thailand routing allegation, confirmed or not, illustrates precisely that pattern — third-country intermediaries aren't a scrappy hack, they are an expensive, politically exposed operation.

Nina Park: So smaller independent Chinese labs — no Alibaba, no state relationships — they're actually the ones getting squeezed.

Walt Garner: Precisely. And the part that follows from that — what Kimi K3 being open-source actually does to where this competition is heading — I think that's where this story gets genuinely uncomfortable.

Nina Park: Okay, but 'genuinely uncomfortable' — you mean the open-source part specifically? Because Kimi K3 is downloadable. Right now. Anyone with a laptop.

Walt Garner: 2.8 trillion parameters, released publicly. The concrete scenario I keep returning to: a machine learning engineer in São Paulo, Saturday afternoon, downloads Kimi K3 from Hugging Face, runs it locally, benchmarks it against Claude on the same tasks. Kimi K3 wins on three metrics. She needed a download link. Not a single chip.

Nina Park: Wait — she can actually run a 2.8-trillion-parameter model on a laptop?

Walt Garner: Quantized versions, yes — and the point stands regardless. The trained weights are out. You cannot embargo a GitHub repository. The twenty thousand H200s mattered enormously to produce Kimi K3. They are now irrelevant to its spread.

Nina Park: So the chip controls — I mean, they worked, in a way? They shaped how the model got built. But the moment it ships open-source, that entire apparatus just... the leverage evaporates. And then Moonshot goes from a $4.8 billion valuation to somewhere between $20 and $35 billion, Hong Kong IPO potentially on the table, and Alibaba's compute subsidy is what funded all of it — Alibaba literally enriched the competitor it created.

Walt Garner: Which is precisely why the next policy question cannot be 'how do we tighten chip controls.' That window — actually, no, the window didn't close, it became the wrong window entirely. The real contest has shifted. Compute hosting. Who controls the infrastructure where the next model trains.

Nina Park: Model weight controls, though — is that even enforceable? You can't un-release Kimi K3.

Walt Garner: You cannot. And that is precisely the discomfort. Future policy has to contend with weights, not just hardware — but the Kimi K3 release demonstrates why that policy arrives too late for this generation of models.

Nina Park: So the thing to watch isn't whether Moonshot gets more chips. It's whether the Alibaba template — fund the startup, supply the compute, absorb the capability — whether that spreads to every major Chinese cloud platform. Because if it does, the chip controls just become a forcing function for vertical integration, forever.

Walt Garner: And that's the question I don't have an answer to. Whether it's Alibaba or Tencent or Baidu running the same arrangement — compute as the condition of the investment — you end up with Chinese AI that is, structurally, more durable. Fewer seams. But also far less surprising. The thing that made Kimi K3 possible is the same thing that might make the next Moonshot impossible, because the next Yang Zhilin may not be able to negotiate a deal that Alibaba doesn't control.

Nina Park: Yeah — and Alibaba being simultaneously the investor, the compute supplier, and the direct competitor through Qwen, all at once, with no firewall between any of those three roles... I mean, there's no US analogue to that. That structure doesn't exist here. So I genuinely don't know if that makes Chinese AI stronger or if it's, I don't know, quietly fragile in some way we can't see yet.

Walt Garner: Mm. I don't know either. And I think that is, you see, the honest place to leave it.

Nina Park: The Kimi chatbot's already in millions of hands. The weights are out. And none of that tells us whether Alibaba won or lost.

Walt Garner: Good conversation. Genuinely.