Alibaba's GPU commitment to Moonshot
Alibaba has supplied Moonshot AI with a high-performance computing cluster of approximately 20,000 Nvidia GPUs to support training of the Kimi models. Multiple outlets reporting on 2 August 2026, citing Bloomberg, said the chips are from Nvidia's previous-generation Hopper line, sidestepping the Blackwell export restrictions that bar advanced parts from shipment to China. An Alibaba spokesperson told reporters there is "no basis" for the suggestion that the company supplied H200 chips specifically, while not disputing that it provided resources equivalent to 20,000 Nvidia GPUs to Moonshot. Alibaba is also a major investor in Moonshot and is developing its own competing model called Qwen.
Kimi K3 specifications and benchmark claims
Moonshot released Kimi K3 as what its backers describe as the first openly available model in the 3-trillion-parameter class, with reporting placing the parameter count at 2.8 trillion. The model is designed for coding, long-context reasoning, and multimodal tasks, and carries a context window of roughly one million tokens (1,048,576). Within a day of its launch on 16 July 2026, K3 took first place on Arena's frontend coding leaderboard and placed third on the Artificial Analysis Intelligence Index, where it scored 57, well above the median of 25 among comparable open-weight models. Moonshot says K3 trails Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol overall but beats their second-tier systems on coding and agentic benchmarks; one report cites Fortune as putting the parameter count at 2.7 trillion, a figure that cannot be reconciled with the 2.8 trillion cited elsewhere.
Pricing, open weights, and distribution
The Kimi K3 API is priced at $3.00 per million input tokens and $15.00 per million output tokens, with cache-hit input at $0.30 per million, per Artificial Analysis; Axios separately cites a blended figure of roughly $12 per million tokens, reflecting a different usage tier. The open weights were released on 26 July 2026 under a custom Kimi K3 Licence rather than a permissive open-source licence, with day-zero hosted access via Together AI and Modal, and the download comprising 96 safetensors shards totaling roughly 1.56 TB. Demand was reported as overwhelming capacity, prompting a temporary pause on new API subscriptions.
Policy reaction and the export-controls debate
The release has reopened US debate over the effectiveness of semiconductor restrictions. David Sacks, co-chair of the President's Council of Advisors on Science and Technology after a 130-day stint as President Trump's special adviser on AI and digital assets, called the result "concerning" on X and argued that blocking new data centers and layering state regulation is undermining US competitiveness. Bloomberg, as relayed by regional outlets, noted that despite China's push for semiconductor independence, the episode illustrates continued reliance on advanced US chips for frontier model training. Michael Kratsios, head of the White House Office of Science and Technology Policy, added that Moonshot also accessed Nvidia Blackwell hardware rented through a third party in Thailand for parts of the K3 training; such rentals are generally permitted under US rules, though Bloomberg sources said it was unclear whether the arrangement constitutes a legitimate rental or a direct purchase that violates regulations.
Moonshot's position in China's AI landscape
Moonshot is backed by both Alibaba and Tencent and is valued at roughly $31.5 billion — a fraction of the trillion-dollar-plus valuations attached to Anthropic and OpenAI, but the largest reported figure tied to a Chinese open-weight model developer. The K3 launch was timed just ahead of the 2026 World Artificial Intelligence Conference in Shanghai, where Chinese President Xi Jinping is expected to set out Beijing's AI priorities. Reports indicate some Alibaba insiders have expressed disappointment that Kimi, running on Alibaba's own infrastructure, has outperformed Alibaba's in-house Qwen on several key metrics, highlighting a competitive tension inside the partnership.
Share this article
