fortune.com web signal

Coinbase and DoorDash shift more workloads to Chinese AI models

TL;DR

  • Coinbase CEO Brian Armstrong says the company halved its AI spending by moving employees onto Moonshot's Kimi and Z.ai's GLM models.
  • One million output tokens reportedly cost $50 on Anthropic's Fable, $15 on Kimi K3, $4.40 on Z.ai's GLM-5.2, and $0.87 on DeepSeek-V4-Pro.
  • Chinese models took 57% of US-firm tokens on OpenRouter in one July week, and Nvidia shed nearly $600 billion after Kimi K3's July 16 debut.

The interesting number in Fortune's reporting isn't the leaderboard win, it's the invoice. Coinbase's Brian Armstrong reportedly told the outlet his company halved its AI spending by pushing employees onto Moonshot's Kimi and Z.ai's GLM models. DoorDash's CTO Andy Fang said the company hands what he calls "lower-level work" to Kimi and gets "better quality [at] cheaper cost." Airbnb's Brian Chesky has previously said the same about running customer service on Alibaba's Qwen. When operators at that scale start talking in ratios rather than pilots, the story stops being about a model demo and starts being about pricing power.

The pricing gap is the reason. Fortune reports that one million output tokens from Anthropic's Fable cost about $50; the same million from DeepSeek-V4-Pro cost roughly $0.87, and from Z.ai's GLM-5.2, $4.40. Moonshot's Kimi K3, released July 16 and described in the piece as the largest open-source model ever released, sits at $15. On OpenRouter, Chinese models grabbed 57% of tokens used by US firms in one July week, with six of the top ten and all of the top five coming from Chinese labs including Tencent, Xiaomi, DeepSeek, MiniMax, Moonshot and Z.ai. Nvidia lost almost $600 billion in value after the K3 release and briefly lost its spot as the world's most valuable company to Apple.

Why this matters if you don't buy inference by the billion tokens: the price floor is being set by open weights coming out of Chinese labs, not by the US frontier labs whose economics depend on hosted-API premiums. That has second-order consequences for Anthropic and OpenAI pricing, for Nvidia's revenue mix, and for any application-layer company that priced its product assuming hosted-frontier costs.

The honest caveats are the ones Fortune flags itself. US lawmakers are now investigating American reliance on Chinese AI models on security grounds, so today's cost savings may become tomorrow's compliance headache. What the reporting doesn't give you is what share of Coinbase's actual workload moved, whether K3's benchmark ranking holds in production, or how Anthropic plans to respond on price. The direction is clear enough, though: the cheap tier is where the volume is, and right now the cheap tier is Chinese.