Startups and enterprises are moving a growing share of their AI workloads to lower-cost open models built by Chinese labs, Bloomberg reports, as pricing pressure from US providers squeezes corporate budgets. The AI arms race has opened a new front, and it is the invoice.

Following the money, one token at a time

The shift shows up clearly on OpenRouter, a platform developers use to access a wide range of AI models. Chinese models have captured more than 30% of weekly tokens on the platform every week since February 8, 2026.

Tokens are the small chunks of text that AI models read and write, and providers bill by the token. That makes token share a reasonable proxy for where the actual work—and the actual spending—is going.

The share did not stay at 30%. It peaked at 67% by mid-September 2026, meaning Chinese models briefly handled roughly two of every three tokens moving through the platform.

Price is the most obvious driver. Chinese AI models reportedly cost 10 to 50 times less per token than US offerings, and in some cases prices run 90% lower. Meanwhile, the US side of the ledger has grown heavier: usage-based pricing from OpenAI and Anthropic has driven sharp increases in enterprise AI costs, with reports of some bills climbing as much as 100x. Usage-based pricing ties costs directly to volume: the more tokens a company processes, the larger the bill.

Lindy goes all in, and bigger names hedge

The starkest example is Lindy, an AI startup that moved entirely from Anthropic's Claude models to DeepSeek's V4 model in June 2026. The switch saved the company millions of dollars and cut its inference costs by approximately 90%. Inference—the step in which a trained model generates each response—is a recurring cost that comes with every request a model handles.

Lindy's wholesale migration is the dramatic version. Larger companies are taking a more measured approach, spreading work across several models instead of committing to one. DoorDash is using Moonshot AI's Kimi. Airbnb is working with Alibaba's Qwen. Coinbase, one of the largest US crypto exchanges, is using models including GLM-5.2.

How chip limits shaped a pricing war

The backdrop is US export controls on advanced chips, which restricted Chinese developers' access to top-end hardware. Rather than stalling, Chinese labs responded by building more efficient models and pricing them aggressively.

Labs including DeepSeek, Moonshot AI, and Alibaba have released open models, meaning companies can download and run them directly rather than only renting access through an API.

What it means for US AI providers and their customers

For enterprise customers, the lesson is already clear: avoid lock-in. Companies such as DoorDash, Airbnb, and Coinbase are building systems that can switch between models, giving them leverage the next time a provider changes its pricing.

Risks sit on the other side of the ledger. Relying on Chinese models raises questions for some companies around geopolitics and regulatory exposure, particularly as US-China tech tensions remain a live issue. Firms will have to weigh those concerns against savings that, in Lindy's case, reached approximately 90%.

For now, the figure to watch is OpenRouter's token share. It has stayed above 30% since February and touched 67% in mid-September.

Source: CryptoBriefing