JPMorgan Report Says Chinese AI Models Cost Up to 50 Times Less Per Token Than OpenAI and Anthropic, and Enterprises Are Starting to Switch

Since earlier coverage this week tracked Qualcomm's AI infrastructure ambitions and SpaceX's expanding compute business, a separate front in the AI race has sharpened into focus: the price war between U.S. frontier models and cheaper Chinese alternatives is no longer theoretical. Enterprises are moving money.

What JPMorgan Found

A JPMorgan report released June 25 put hard numbers on the cost gap. Chinese AI models now cost10 to 50 times less per token than comparable U.S. frontier models, according to the analysis, while performing at roughly comparable levels on several standard benchmarks. The report offered a direct head-to-head comparison between an Anthropic Claude model and DeepSeek, citing a nearly 20x cost gap for a performance gap of about 20% on the benchmark tested. (Note: The specific model version numbers cited in earlier drafts of this coverage — "Claude Opus 4.JPMorgan also flagged the specific pricing moves driving enterprise anxiety. OpenAI reportedly doubled token prices between successive GPT-5 model versions. Microsoft raised Copilot pricing and increased charges across multiple AI model families. Some enterprise customers reported AI-related cost increases of up to 100-fold following pricing changes.

Who's Actually Switching The migration is already underway in measurable cases

AI startup Lindy said it moved its services from Anthropic's Claude to China's DeepSeek, claiming millions of dollars in annual savings and improved performance on its specific workloads. Coinbase CEO Brian Armstrong said publicly he expects most AI workloads to shift to significantly cheaper models over the next year. The JPMorgan report named DeepSeek, Alibaba, Xiaomi, MiniMax, and Kimi as the Chinese providers now crowding the efficiency frontier, the zone where performance per dollar is highest.

The SpaceX Data Point

That Fits Here Separately, the Forward Future newsletter reported June 23 that AI startup Reflection AI has signed a deal to access NVIDIA GB300 chips at SpaceX's Colossus 2 data center, paying$150 million per month starting July 1, with the agreement running through 2029. Reflection joins Anthropic, Google, and Cursor as SpaceX compute customers. That contract illuminates the cost math from the other direction. If compute infrastructure costs that much, token prices will rise. The U.S. frontier labs are not raising prices out of greed alone. They are recovering infrastructure costs that their Chinese competitors, operating under a different cost structure and with different government relationships, may not face in the same way.

The Fair Case for OpenAI and Anthropic

The strongest defense of the premium U.S. models is not just about benchmarks. JPMorgan itself said frontier models from OpenAI and Anthropic will remain essential for advanced scientific research, cybersecurity, and high-end AI agents. Benchmark scores measure specific tasks. Real enterprise deployments involve reliability, safety guarantees, legal liability, data privacy, and integration depth with existing U.S. software infrastructure. There is also a national security dimension that the cost-per-token framing ignores entirely. DeepSeek is a Chinese company. Routing sensitive enterprise data through Chinese AI infrastructure raises questions that no benchmark score resolves. Whether U.S. enterprises are weighing that risk seriously or treating AI as a pure cost center is, at this point, an open question.

What's Actually Proven vs

Alleged What is proven: the pricing differentials JPMorgan cited are sourced to specific model comparisons. The Lindy migration is on the record. Armstrong's forecast is an executive's stated expectation, not a verified outcome. What is not yet established: whether the cost savings hold at scale, whether Chinese model performance holds on enterprise-specific tasks rather than standard benchmarks, and whether the data-sovereignty and export-control risks are being priced in by the companies doing the switching. A separate report claiming that Chinese AI models have overtaken U.S. That framing is plausible given the pricing dynamics but should be treated as a directional signal, not a confirmed figure.

The Open Question

JPMorgan's report lands the same week that SpaceX is signing nine-figure monthly compute contracts with U.S. AI companies and Qualcomm is targeting $40 billion in non-handset AI revenue by 2029. The U.S. AI infrastructure build-out is accelerating. But if enterprise workloads are migrating to Chinese models before that infrastructure matures, the revenue base that is supposed to justify those investments gets thinner. The unresolved question: whether federal export controls, data-privacy regulations, or enterprise risk policies will slow that migration, or whether cost pressure will override them. No regulatory action specific to enterprise use of Chinese AI models has been announced as of June 25, 2026.