A quiet infrastructure announcement may mark a bigger shift than most model releases this year. Baseten, a US AI infrastructure company, says enterprise users can now run Moonshot AI's Kimi K3 inside OpenAI's coding tool Codex — and the charges flow straight into the OpenAI procurement commitments those companies have already signed. No new vendor contract, no fresh procurement cycle, no security review of a foreign supplier. Chinese financial wire Cailianshe reported the news on October 1, calling it the first time a Chinese open-source model has entered OpenAI's mainstream enterprise billing channel.
The mechanics matter as much as the milestone. Enterprise buyers typically lock in annual AI spending commitments with a single vendor to get volume pricing. By routing Kimi K3 through Baseten inside Codex, those same dollars can now pay for a Chinese open-weights model without procurement ever signing a second paper. For open-source labs that give weights away, distribution — not capability — has become the monetization bottleneck, and Baseten just widened a significant pipe.
The Codex integration follows another first. Shortly before, Amazon Bedrock announced support for Kimi K3, which Chinese state-adjacent coverage described as the first revenue-sharing arrangement under which a Chinese large-model company supplies its model through a global top-tier cloud provider. Taken together, Kimi is now reachable through two of the largest enterprise AI channels on the planet — Amazon's and, indirectly, OpenAI's.
The model earned its seat. Moonshot released Kimi K3 in July with 2.8 trillion parameters, the largest open-source model at launch, a 1-million-token context window and native visual understanding. Chinese state broadcaster coverage said it topped the Arena coding leaderboard within hours of release — the first Chinese model to do so — and even Elon Musk left a one-word reply on evaluation coverage: "Impressive."
Kimi K3 has also produced the field's most eye-catching agent stunt: an AI-designed chip. In a 48-hour autonomous agent run, the model used open-source EDA tooling and the Nangate 45nm process library to carry a chip from design through verification — a 4-square-millimeter die with 1.46 million standard cells, 0.277 MB of SRAM and an INT4 MAC array that reached timing closure at 100 MHz, decoding at up to 8,700 tokens per second. Moonshot says it ran the whole flow unattended; there is no third-party silicon to show for it yet.
Commercial plumbing is already public. Baseten's published model catalog lists Kimi K3 at $3 per million input tokens and $15 per million output tokens in fp8 with a 1-million-token context — a premium over DeepSeek's listings on the same platform, but far below frontier closed-model pricing. The same day as the Codex news, IDC released its China AI cloud tracker, sizing the market at RMB 68.4 billion for 2025 with enterprise MaaS token volume up roughly 16-fold year over year.
The strategic read: the open-source frontier is no longer a benchmarking exercise confined to domestic ecosystems. When a Western enterprise can spend its OpenAI budget on a Beijing lab's weights inside OpenAI's own tool, the moat is no longer the model — it is the checkout lane. Expect every major lab to fight harder over which models are allowed inside their procurement gravity wells.
Comments (0)
Log in to join the discussion
Log InNo comments yet