Chinese AI models now eating 30 to 46 percent of US enterprise tokens

[NorthernKernel]

CNBC dropped an investigation this week showing Chinese built AI models now make up somewhere between 30 and 46 percent of all enterprise API token traffic flowing through US developer platforms like OpenRouter, up from just 4.5 percent in the first half of last year

The number that really got me was Coinbase running 1200 agents on Chinese models and cutting their AI spend nearly in half, and Lindy apparently moved 100 percent of their traffic off Claude and onto DeepSeek

OpenRouter data shows the crossover week was back in February when Chinese models actually out consumed US models for the first time on the platform, and it has not looked back since

The pricing gap is the whole story here, DeepSeek V4 Flash is listed at 14 cents per million input tokens while GPT 5.5 sits at 5 dollars, that is not a rounding error for anyone running high volume workloads

Two US House committees are now jointly investigating whether this counts as a national security issue, so this is not staying a purely technical story for long

Chinese AI models are gaining ground with U.S. companies as OpenAI, Anthropic costs surgewww.cnbc.com

[StringTheory83]

This was inevitable honestly, once the frontier labs started charging enterprise rates for marginal quality gains someone was going to undercut them

The only surprise is how fast it happened

[Odd Arrow]

Cost is cost but running production workloads on models where the origin and training data curation is a black box seems like a much bigger risk than people are pricing in right now

[Margin]

Genuinely asking, has anyone here actually tested GLM 5.2 or DeepSeek V4 side by side against Opus or Sonnet on a real codebase, not a benchmark, an actual messy codebase

[Dave]

I run a small dev shop and switched two of our three internal tools to a Chinese open weight model in June

Honestly the difference on day to day tasks is basically unnoticeable and the bill dropped by 70 percent

[Tel75]

The Booz Allen study mentioned in some of the coverage found Chinese code models add way more vulnerabilities when the prompt implies a US government user, that alone should give enterprises pause before they go all in

[Restless Barrel]

People keep framing this as patriotism versus practicality but I think it is simpler than that, it is just what happens when a market gets commoditized and buyers stop caring about brand

[DQ Eric]

Congress investigating feels like theater at this point, they were not investigating anything when American companies were the only game in town and prices were sky high

[TrainingRun Anvil]

The part nobody is talking about is what happens if Beijing actually restricts overseas access like the article hints at, a lot of these companies built entire workflows around models that could vanish overnight

[ComputeNodeCanopy]

I would rather pay more and know exactly what I am running than save money on something I cannot audit

Maybe that is old fashioned but data residency actually matters to my clients