Moonshot Kimi K3 hits capacity limits as U.S. enterprises adopt Chinese open-weight at scale
Moonshot AI's Kimi K3, a 2.8-trillion-parameter open-weight model released July 16, paused new subscriptions within 48 hours due to overwhelming demand. The company wrote on X: "Kimi K3 has received far more love than we expected… demand has pushed close to the limits of our current capacity." K3 is the largest open-weight model ever released and debuted at #3 on Artificial Analysis leaderboard (behind Anthropic Fable 5 and OpenAI Sol), while topping Arena's blind coding evaluation and four of eight task-automation benchmarks. Moonshot is adding capacity in batches but faces intense compute pressure.
U.S. enterprises are adopting Kimi at production scale. DoorDash CTO Andy Fang publicly stated the company delegates "lower-level work" to Kimi K2.6, achieving "better quality at cheaper cost" compared to Anthropic Fable. Cursor (SpaceX's $60B acquisition target) built its Composer 2 agent on Kimi 2.5. Coinbase and AI startup Lindy both confirmed internal use. Per OpenRouter data, Chinese AI models now account for 56–60% of token usage by U.S. companies on the platform—a dramatic shift in just weeks. Kimi K3 pricing: $15 per million output tokens, vs. Fable 5 at $50 and DeepSeek V4 at $0.87.
For production teams, the implication is clear: open-weight models from China have crossed into "good enough and cheaper" territory for high-volume, lower-complexity workloads—and capacity is now the constraint, not capability. U.S. lawmakers (DoorDash subpoenaed July 31) are scrutinizing enterprise adoption, but the economics are undeniable: Chinese models are 3–70x cheaper per token while matching or exceeding prior-gen U.S. closed systems. The inflection point arrived faster than analysts expected (Amodei predicted 6 months; it's happened in weeks). Enterprises face a new decision: build custom tooling to run self-hosted Kimi weights on leased GPU capacity, or pay premium for U.S. API pricing. The capacity crunch at Moonshot suggests demand will outpace supply through 2026.
Sources
- Primary source
- cnbc.com
“DoorDash is delegating lower-level AI work to Chinese AI model Kimi K2.6, leading to better quality at cheaper cost”
- fortune.com
“Chinese AI models' share of tokens used by U.S. firms on OpenRouter in one week in July reached 57%”
- finance.yahoo.com
“Cursor, a coding startup that SpaceX is acquiring for approximately $60 billion, acknowledged it developed its product using a Kimi model”