Moonshot’s $35B Test After Kimi K3

On July 27, 2026, Moonshot AI, the Chinese model lab founded by Yang Zhilin, made Kimi K3’s weights available for public download, modification and hosting. Two days later, Bloomberg reported that Moonshot had closed a $3.5 billion round at a $35 billion valuation. The latest revenue figure supporting that valuation—$300 million in annual recurring revenue in June—predated K3’s open-weight release.

Bloomberg had reported on July 19 that Moonshot’s ARR rose from $200 million in April to $300 million in June. Sources therefore described a 50% increase in two months, evidence that Moonshot’s existing business was growing before developers could host K3 themselves.

Moonshot now has to show that K3’s broad distribution can produce durable revenue after users gain the option to run the model without sending every request through the lab.

Kimi K3 spreads before Moonshot shows who pays

Moonshot’s July 16 announcement described K3 as a 2.8-trillion-parameter model and claimed performance competitive with Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5. The company said it would release the weights by July 27, making distribution part of the launch rather than a later concession.

Bloomberg’s July 27 and 28 accounts establish that users may download, tweak and host K3 under the Kimi K3 License. Those reports do not document separate terms for large hosting companies or display requirements above specified thresholds.

A developer can now evaluate K3, adapt it and operate it on third-party infrastructure without waiting for Moonshot’s product roadmap. Developers can therefore adopt K3 without sending every inference request through Moonshot, so model usage can rise faster than the lab’s own revenue. The reported ARR does not break out K3 revenue or show how self-hosting affects paid usage.

Moonshot still needs a deployment layer

On June 13, VentureBeat reported that Moonshot released Kimi K2.7-Code under a modified MIT license and claimed 30% lower reasoning-token use than K2.6, while noting practitioner disputes over the benchmark evidence. K3 continued that emphasis on model capability and developer access.

DeepSeek moved one layer outward when it released DeepSeek Harness under the MIT license. Its plug-in architecture lets developers swap components, giving the company a role in how builders organize model-driven work rather than only which model they select.

Bloomberg’s July K3 reports identify no named enterprise deployment, managed-agent service or Moonshot-operated workflow platform. An engineering team delegating code changes still needs software to assign permissions, run tools, detect failures and route exceptions. The cited record leaves unanswered who supplies that operational layer and what Moonshot charges for it.

A Hong Kong listing would put revenue in view

Moonshot’s financing expectations rose quickly during the four months before K3’s weight release.

• March 27: The Wall Street Journal reported that Moonshot was considering replacing its Cayman structure with a China or Hong Kong entity, seeking funding at roughly an $18 billion valuation and preparing for a possible Hong Kong IPO.
• May 7: Bloomberg reported a roughly $2 billion round at a valuation above $20 billion and put April ARR at $200 million.
• July 19: Bloomberg reported June ARR of $300 million and said Moonshot had told investors it could list within six months.
• July 29: Bloomberg reported a $3.5 billion round at $35 billion, above an initial target of $1 billion to $2 billion, and an IPO as soon as 2026.

Moonshot has not confirmed the financing or IPO timetable. If it lists, regular disclosures would let investors compare post-K3 revenue with training costs, inference capacity and customer concentration. The June ARR figure establishes momentum entering the release; it cannot establish how the open-weight model performs commercially afterward.

Nvidia access and provenance can interrupt deployment

On July 28, The Information reported that Moonshot trained K3 on Nvidia chips, including Blackwell, and was seeking additional Blackwell capacity for Kimi K4. Moonshot needs that supply to continue training larger models while supporting inference for customers who do not self-host.

White House science and technology director Michael Kratsios has accused Moonshot of distilling an Anthropic model and accessing GB300-equipped servers in Thailand. Neither allegation is an established finding. OpenAI President Greg Brockman called Kimi K3 a “pretty good model” while saying it was too early to determine whether Moonshot had distilled OpenAI models.

A customer building on K3 must assess whether chip restrictions, disputed provenance or geopolitical pressure could interrupt service. Moonshot must address those risks through documented supply arrangements, governance and customer support; benchmark claims cannot resolve them.

The $35 billion figure arrived two days after K3’s weights. It prices Moonshot as though distribution has already become a durable business, while the dated revenue evidence stops in June. Moonshot’s commercial proof begins with customers who can host K3 themselves and still choose to pay the lab.