Google CloudはMoonshot AIのKimi K2エンドポイントを10月に廃止します。

Why it matters

US clouds can turn Chinese open-weight models into managed enterprise services. That expands developer choice while raising unresolved questions about commercial ties and model governance.

Yang Zhilin's Beijing-based Moonshot AI still has its Kimi K2 Thinking model available through Google Cloud as a managed API, though Google deprecated the endpoint on 7月21日 and plans to retire it on 10月21日.

The arrangement drew renewed scrutiny Wednesday after Chris McGuire (@ChrisRMcGuire) argued Xへの投稿で that Google was paying Moonshot for the right to host the Chinese developer's model and serve it to US customers using US chips. 「これは明らかに許されるべきではない」とMcGuireは書いた。

Google's public documentation confirms the central technical point: Kimi K2 Thinking has been offered through Google's Model Garden as a serverless, managed API. The model supports pay-as-you-go access, provisioned throughput and processing in Google's US multi-region. Customers call a Google Cloud endpoint without provisioning their own 推論インフラストラクチャ.

The documents do not establish McGuire's claims about payments from Google to Moonshot or the underlying chips. Google's pages describe what customers pay Google and how the service operates, rather than commercial terms between Google and Moonshot. Google prices Kimi K2 Thinking at $0.60 per million input tokens, $2.50 per million output tokens and $0.06 per million cached tokens.

An eight-month-old listing nearing retirement

Google added Kimi K2 Thinking to Model Garden on 2025年11月13日, according to its Vertex AI release notes. The listing gave enterprise developers managed access to an open-weight Chinese model through the same cloud environment used for Google's own models and third-party systems.

Google describes Kimi K2 Thinking as a reasoning and tool-use model capable of making between 200 and 300 sequential tool calls. Its managed endpoint, kimi-k2-thinking-maas, supports function calling, structured output and a context window of 262,144 tokens.

The 7月21日の非推奨化は、Moonshotや中国の開発者に限定された措置ではなく、より広範なエンドポイント廃止の一環だった。GoogleはDeepSeek、Z.ai、MiniMax、AlibabaのQwenに加え、OpenAIのgpt-oss-20bやMetaのLlama 3.3 70Bを同じ10月21日の廃止スケジュールに載せた。

その一括掲載を見ると、Kimiの廃止だけを地政学的な決定として読み取るのは難しい。Googleは縮小期間中もKimi K2 Thinkingとその顧客向け価格を引き続き掲載している。一方でMoonshotは自社製品群を前進させており、同社のAPI platformは現在Kimi K3を旗艦モデルとして宣伝している。

Moonshot's US research roots

Moonshot's founders complicate a simple US-versus-China framing. Yang completed a PhD in language and information technology at Carnegie Mellon University in 2019 and contributed to research including Transformer-XL and XLNet before starting Moonshot in early 2023.

Co-founder Yuxin Wu also studied at Carnegie Mellon after earning his bachelor's degree at Tsinghua University. Wu says he worked on foundation models at Google Brain and computer vision at Facebook AI Research before Moonshot. He created Detectron2 and contributed to research including Group Normalization and Momentum Contrast.

Moonshot has since become one of China's best-capitalized private AI developers. TechCrunch reported in May that Moonshot raised about $2 billion at a valuation of roughly $20 billion, citing a financial adviser involved in the transaction. The report said annual recurring revenue exceeded $200 million in April, driven by subscriptions and API usage.

Google's Kimi listing shows how open-weight models cross national boundaries after release. A US cloud provider can package a model from a Beijing lab as enterprise infrastructure, bill developers by the token and run requests through a familiar API. The public record supports that distribution relationship. It does not support the specific assertion that Google is paying Moonshot, and the endpoint is already scheduled to disappear less than three months from now.