Kimi Grabs 61% of US Enterprise Open-Source AI Spend
Cointime reports: Chinese open-source models now account for nearly 90% of US enterprise spending on open-weight AI models, with Moonshot AI’s Kimi alone capturing 61% of that market. DeepSeek follows at 15%, GLM at 11%, and Qwen at 3%, while Meta’s Llama — the leading Western open-source model — holds just 7%.
The findings, based on a survey of nearly 400 US enterprise AI decision-makers managing over $4 billion in AI budgets, reveal that total US enterprise AI deployment spending reached $176.8 billion by September 2026, a 5.4x increase from $33 billion a year earlier.
Why Enterprises Are Paying Premium Prices
The shift toward Kimi defies the prevailing narrative that cheaper models would dominate enterprise workloads. Between Q3 2025 and Q3 2026, token share for small closed-source models like Haiku and Luna fell from 29% to 13%, while frontier flagships saw their token share rise from 18% to 42% and spending share jump from 35% to 63%.
The driver is the rise of AI agents. On model routing platform OpenRouter, agent-initiated traffic surpassed human traffic in February 2026, and by September, 80% of platform tokens were generated by agents running autonomously in the background — a 37x increase in just over six months.
Agent workloads compound errors multiplicatively. A model with 99% per-step accuracy running 50 sequential steps completes the task successfully only about 60% of the time. At 95% per-step accuracy, that drops below 8%. Enterprises have therefore shifted from optimizing for sticker price to optimizing for total cost of getting the job done right the first time.
Kimi’s Engineering Moat
Kimi’s 61% share stems from not competing on price. While average open-source model pricing collapsed to $0.13 per million tokens — less than one-fifth of closed-source averages — Kimi priced its K3 model at $3 input/$15 output, matching frontier closed-source flagships.
The justification lies in engineering rigor. Moonshot AI open-sourced a vendor verification tool (KVV) that benchmarks how different inference clouds deploy its models, testing tool-calling regression, JSON schema validation, and 113 coding agent tasks. This addressed a critical pain point: inference providers often quantize models to maximize GPU throughput, degrading performance in multi-step agent loops.
Cursor’s decision to build Composer 2 on Kimi K2.5 underscores the model’s appeal. The coding startup used only about a quarter of final compute on the base model, spending the rest on its own reinforcement learning — a model that scored 61.3 on CursorBench versus Claude Opus 4.6’s 58.2.
The New Open-Closed Equilibrium
The data reveals a maturing ecosystem rather than a winner-take-all battle. Anthropic, OpenAI, and Google still command 69% of enterprise AI spending, and closed-source flagships remain the default for zero-tolerance tasks. But open-weight models like Kimi have carved out a wide middle ground where application companies can build proprietary assets on top of frontier-adjacent foundations.
For crypto markets, the implications are significant: decentralized compute networks and AI agent protocols settling on-chain will increasingly compete — or integrate — with this enterprise inference stack. The model layer’s economics are consolidating around quality and reliability rather than raw cost, a dynamic that will shape how decentralized AI infrastructure positions itself against centralized incumbents.
