Kimi K3 Is So Popular, Moonshot Ran Out of GPUs

Moonshot AI built a model so good, it broke their own infrastructure. Just three days after launching Kimi K3 — the 2.8 trillion-parameter open-weight model that had the AI world buzzing — the Chinese startup hit pause on new subscriptions. The reason? Their GPUs were drowning.

The suspension went into effect July 19, a mere 48 hours after K3's debut. In a statement on X, Moonshot didn't mince words: "Kimi K3 has received far more love than we expected, and our GPUs are feeling it." User demand had pushed the company's compute cluster close to maximum capacity far faster than projections suggested.

The Open-Weight Paradox

Here's the irony that has the AI community talking: Kimi K3 is marketed as the world's largest open-weight model. In theory, open weights should distribute the load. Anyone can download the model and run it themselves, right?

Not yet. Moonshot won't release the K3 weights until July 27 — a week after the subscription freeze began. Until then, the only way to use the model is through Moonshot's own apps and API. All that pent-up demand from developers, researchers, and curious users is landing on one company's servers.

And this isn't software you can run on consumer hardware. Kimi K3 requires a multi-GPU setup — estimated at around eight H100 or H200 chips just to serve inference. Factor in the one-million-token context window and native vision capabilities, and each user session becomes a serious compute commitment.

When Will the Freeze End?

Moonshot hasn't published a specific timeline for reopening subscriptions. What they've said is that new slots will return "in batches" — a controlled rollout rather than a floodgate opening. The company is also reworking its membership structure to manage demand better.

Starting soon, Moonshot will split its plans into two distinct products:

• Kimi Membership for general web, app, and office use
• Kimi Code Membership for programming workflows

The logic is straightforward: coding and agentic tasks consume disproportionately more compute. They involve long reasoning chains and massive context windows. By ring-fencing the heavy users, Moonshot hopes to prevent power users from draining the pool for everyone else.

The Real Pressure Valve: July 27

The more significant date is July 27, when Moonshot plans to release the K3 weights publicly. That's when the open-weight promise actually materializes. Once the files are available, cloud providers and enterprise customers can host K3 themselves, bypassing Moonshot's queue entirely.

Casual users will likely continue relying on Moonshot's hosted service — not everyone has eight H100s sitting around. But the weight release should significantly ease demand on Moonshot's infrastructure by giving serious users an alternative path.

A High-Class Problem

Let's be clear about what this is: a high-class problem. Moonshot built a model so capable that demand overwhelmed supply within days. Kimi K3 ranked second only to Anthropic's Claude Fable 5 on coding leaderboards, beat rivals on front-end development and long-context tasks, and did it at prices that undercut US competitors.

The numbers tell the story. Moonshot's annual recurring revenue hit $300 million in June, up from $200 million just two months earlier. The company is reportedly preparing for a Hong Kong IPO that could value it at over $30 billion.

A model too popular to sell makes for an unusual investor pitch, but it's a persuasive one.

Not Alone in the Squeeze

Moonshot isn't the only lab feeling the heat. In the same week, Anthropic cut usage limits on Claude Fable 5 for subscribers, citing demand that had become "hard to manage." The pattern is becoming clear: the barrier to AI adoption is shifting from model capability to compute availability.

Chinese labs like Moonshot, Zhipu AI, and DeepSeek are delivering frontier-level performance at prices that undercut OpenAI and Anthropic by 80% or more. The models are here. The chips to run them? That's the bottleneck.

What Happens Now

For anyone waiting to try Kimi K3, the options are limited. You can wait for the subscription waitlist to reopen in batches, or you can wait for July 27 and the open-weight release. The latter is likely the faster path for anyone with the infrastructure to self-host.

For Moonshot, the challenge is scaling fast enough to capture the demand they've created. The company says it's racing to add capacity, but GPUs don't materialize overnight — especially not when export controls and supply constraints are affecting the entire industry.

The Kimi K3 capacity crunch is a preview of what's coming. As open-weight models close the gap with frontier closed systems, the limiting factor won't be intelligence — it will be infrastructure. Moonshot just learned that lesson the hard way, three days after launch.