China's Moonshot AI raises $2b at $20b valuation

Moonshot AI, a Beijing-based startup behind the Kimi chatbot, has raised about US$2 billion in a round led by Meituan’s venture arm.

The deal values the company at more than US$20 billion as investors continue backing Chinese AI startups.

HF Capital, which advised some backers, said Moonshot’s annual recurring revenue passed US$200 million in April, while Meituan’s Long-Z Investments confirmed it joined the round.

Moonshot raised US$500 million at a US$4.3 billion valuation late last year, then US$700 million at a US$10 billion valuation earlier this year before seeking another US$1 billion.

Founded by former Tsinghua professor Yang Zhilin, who previously worked at Meta and Google, Moonshot sells paid chatbot subscriptions and AI services to enterprise clients.

🔗 Source: Bloomberg

🧠 Food for thought

Implications, context, and why it matters.

Moonshot’s valuation rests on AI agents built for extended work

Moonshot’s Kimi K2.6 model is built for complex tasks that run over long stretches, well beyond simple chat 1.
In one case, it can update an eight-year-old software project over 13 hours. In another, it can build a full compiler, software that translates code into machine-readable instructions, from scratch in 10 hours without human help 1.
The model also uses “agent swarms,” systems that split one job into 300 smaller AI agents working at once across 4,000 coordinated steps 2.
That approach matches a wider shift in AI. Anthropic and OpenAI are also pushing toward long-horizon agents through multi-session tasks, subagents, and background execution 1.

Long-running AI agents are pushing past current software control systems

Models like Kimi K2.6 can run for extended periods. Moonshot says one internal example lasted five straight days, which strains current software orchestration frameworks, the tools companies use to coordinate tasks and tool access across apps 1.
Those frameworks were built for agents that run for seconds or minutes. They have trouble handling the memory, status, and tool access these longer jobs need 1.
The change is driving talk of new layers such as an “agent runtime,” software that manages how an AI agent operates over time, and an “agent identity provider,” which checks and controls what an AI agent may access 1.
Companies now face pressure to revisit parts of their IT setup and governance. These systems can produce code and system changes faster than many IT and risk teams can review 1.