Moonshot pauses Kimi subscriptions as K3 demand strains GPU capacity

Why it matters

Moonshot's pause shows the constraint facing frontier AI challengers: competitive models can attract users faster than their developers can provision profitable inference capacity.

Moonshot AI paused new subscriptions to its Kimi service on Sunday after demand for Kimi K3 pushed its GPU capacity close to the limit during the previous 48 hours, Kimi.ai said in a July 19th post on X.

Existing subscribers will retain access while Moonshot directs available compute toward them. Moonshot said it is adding capacity and plans to reopen subscriptions in batches, without giving a schedule or disclosing how much additional infrastructure it needs.

The bottleneck arrived three days after Moonshot released K3, the latest model from co-founder and CEO Yang Zhilin. Yang completed his doctorate at Carnegie Mellon University in 2019 and was a co-author of Transformer-XL and XLNet, two projects focused on how language models process context. That research thread runs directly through K3, which Moonshot says can process a 1 million-token context window.

Moonshot described the demand spike with unusual bluntness: "our GPUs are feeling it." The pause turns K3's early reception into an infrastructure problem for Yang. Moonshot can generate attention for a model positioned against the strongest US systems, but serving that model reliably requires enough accelerators, networking capacity and memory bandwidth to absorb sustained consumer and developer use.

A large model meets a finite serving cluster

Moonshot says K3 contains 2.8 trillion total parameters, supports images natively and uses a mixture-of-experts architecture that activates 16 of 896 experts during inference. The developer recommends deployments with at least 64 accelerators, an indication of the hardware footprint required even though only part of the model runs for each token.

K3 is available through Kimi's web and mobile products, the Kimi Work desktop application, Kimi Code and the company's pay-as-you-go API. Moonshot calls K3 an open model, though it has scheduled the release of the full weights for July 27th. Until then, most users remain dependent on infrastructure operated by Moonshot or its inference partners.

That dependence concentrates launch demand on Moonshot's own systems. It also makes capacity management part of the product experience: a model's benchmark position matters little to a paying customer whose requests sit in a queue or fail during peak periods.

Moonshot says K3's API costs $0.30 per million cached input tokens, $3 per million uncached input tokens and $15 per million output tokens. Its membership help center listed monthly paid plans ranging from $19 to $199 on Sunday, with higher tiers receiving additional agent credits, concurrency and Kimi Code usage.

The subscription pause cuts off one source of immediate revenue to preserve service for customers Moonshot has already acquired. It also shows why flat or bundled AI subscriptions become difficult to manage when coding agents, long-context research tasks and ordinary chat draw from the same pool of expensive compute.

Moonshot will separate coding from general Kimi access

Moonshot said it will divide its membership product into two offerings. Kimi Membership will cover the web service, mobile applications and Kimi Work, while Kimi Code Membership will be dedicated to coding workflows.

The split gives Moonshot a cleaner way to price and allocate capacity. Coding agents can run extended sessions, repeatedly ingest repositories and execute multiple tool calls. Those workloads have a different cost profile from short consumer chats or document questions. Separate plans let Moonshot assign quotas and infrastructure without forcing every subscriber into the same usage pool.

The move also changes the pitch behind Kimi's existing bundle. Moonshot's documentation has treated Kimi Code as a benefit inside its broader membership, with access through the terminal, IDE integrations and compatible third-party coding tools. Separating the products gives Moonshot room to charge according to workload as K3 replaces earlier models across its services.

Moonshot has substantial capital behind that buildout. TechCrunch reported in May that Moonshot raised about $2 billion at a $20 billion valuation, citing an investment adviser involved in the financing. The K3 launch makes serving capacity the immediate test of that funding.

Independent interest extends beyond Moonshot's own benchmarks. The Associated Press reported that K3 reached the top of Arena's front-end coding ranking, while Moonshot's technical post acknowledged that the model's overall performance still trails Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol.

For Yang, the next milestone is operational. Moonshot has to add enough capacity to reopen sales, preserve performance for current users and establish separate economics for coding workloads. The pace of that expansion will determine whether K3's launch demand becomes durable, billable usage or a queue Moonshot cannot serve.