The short answer
US companies are shifting a growing share of their AI token volume to open-weight Chinese models — DeepSeek, Alibaba’s Qwen and Zhipu’s GLM — because they run roughly 60% to 90% cheaper than the top OpenAI and Anthropic models for comparable work. OpenRouter data reported by CNBC on 7 July 2026 shows Chinese providers have supplied more than 30% of US-routed tokens every week since 8 February 2026, peaking near 46%, versus an average of about 11% over the prior twelve months.
The practical reading for engineering leaders: model cost is now a first-class architectural variable, and the cheapest capable option is increasingly not from a US lab. The catch is that the direct API path for these models often runs through China, which raises data-residency and compliance questions for regulated teams. The winning move is to capture the price drop while keeping data in your own jurisdiction — something the open-weight nature of these models makes possible.
What is actually happening?
For most of the past two years, the default assumption for a serious production workload was a frontier model from a US lab — OpenAI or Anthropic. That assumption is eroding fast. Through the first half of 2026, a wave of capable open-weight models out of China closed enough of the quality gap on coding and general tasks that price became the deciding factor, and price is where they win decisively. The result is a measurable migration of real production traffic, not just hobbyist experiments.
The clearest signal comes from OpenRouter, a routing service that sits between applications and dozens of model providers. As CNBC reported on 7 July 2026, the share of US-routed tokens going to Chinese models has stayed above 30% every single week since 8 February 2026 and has spiked as high as 46%. A year earlier that figure averaged around 11%, and in the first half of 2025 it was as low as 4.5%. For any team running high-volume inference — the kind that dominates cost in AI and data workloads — that is a structural change in where the tokens go.
This is not a single-vendor story either. The gains are spread across DeepSeek, Alibaba’s Qwen, Zhipu (now branded Z.ai), MiniMax, Moonshot’s Kimi and others. By April 2026 the combined weekly share of the major Chinese providers on OpenRouter had crossed 45% of total volume, up from under 2% a year before. The market has moved from “interesting benchmark result” to “line item in the infrastructure bill” in about twelve months.
How big is the shift?
Start with the raw price gap. OpenRouter’s own analysts put open-weight Chinese models at roughly 60% to 90% cheaper than the leading Anthropic and OpenAI models. In one illustrative comparison for a similar body of work, Anthropic’s Claude came out around $4,811, OpenAI around $3,357, and Zhipu’s GLM around $544. Numbers like that do not survive a quarterly budget review unchallenged — once a comparable-quality option is available at a tenth of the cost, finance starts asking why every call routes to the expensive tier.
The adoption curves match the price story. Z.ai’s GLM 5.2, released in June 2026, saw the fastest uptake of any model Vercel tracked this year: daily token volume grew roughly 27x and the number of customers using it grew about 80x in its first full week. DeepSeek’s share of tokens on Vercel jumped from under 1% to about 17% in May alone. And the switch is happening at the level of whole companies: the AI assistant startup Lindy moved 100% of its traffic from Anthropic’s Claude to DeepSeek in June, with its CEO expecting to save millions of dollars as the cost curve fell.
Individual developers tell the same story in miniature. Practitioners quoted in the reporting described dropping from roughly $10 an hour of Claude usage to under 50 cents an hour on DeepSeek, or reserving premium US models for the hardest 10% of tasks while routing the routine 90% to cheaper Chinese models. The pattern is consistent: keep a top-tier model for the work that truly needs it, and send everything else to the cheapest option that clears the bar.
Why are teams switching?
The headline reason is cost, but the deeper reason is that cost stopped being justified by a quality moat. When the best US model was clearly ahead on the tasks you cared about, paying a premium was rational. In 2026 the open-weight Chinese models are close enough on mainstream coding, drafting and tool-use workloads that the premium looks more like a tax than an investment for the bulk of calls. That reframing — from “best model” to “cheapest model that clears my quality bar per task” — is the mental shift driving the migration.
The second reason is optionality. Because these models are open-weight, teams are not locked into a single vendor’s roadmap, pricing or availability. You can run them behind your own interface, fine-tune them on your domain, or swap them out when something better appears — the same portability logic that makes fine-tuning open models attractive for teams that want a durable, owned capability rather than a metered dependency. In a year when frontier access has become staggered and prices have moved in both directions, that optionality is worth real money.
What are the compliance and data risks?
Here is where the enthusiasm needs a governance layer. The cheapest way to try these models — calling a hosted API that runs in China — is also the riskiest for anyone handling regulated or sensitive data. Sending customer records, health data or financial details to an offshore endpoint raises data-residency, security, censorship and geopolitical concerns, and it is not hypothetical: US lawmakers have already opened inquiries into American companies that routed workloads through Chinese models. For a FinTech or healthtech platform, that path can collide directly with GDPR, HIPAA and internal data-handling commitments.
The important nuance — and the reason this is an opportunity rather than a warning — is that the model and the hosting location are separate decisions. DeepSeek, Qwen and GLM publish their weights, so you do not have to use a Chinese-hosted API to use the model. You can pull the open weights and run inference inside your own environment on a US or EU cloud, keeping data in-jurisdiction end to end. That is the same discipline behind sound GDPR compliance work: the question a regulator asks is not only “which model” but “where did the data go.” Answer that with self-hosting and the cost argument survives the compliance review intact.
What it means for US & EU software teams
The first lesson is that model choice is now a cost-engineering discipline, not a one-time vendor pick. With capable models available across a 10x price range, the durable pattern is a provider-agnostic routing layer that sends each workload to the cheapest model that clears its quality bar. Teams that hard-code one model and one price are leaving a large, recurring line item on the table — and are exposed when that vendor changes pricing or availability.
The second lesson is that open-weight changes the deployment calculus. The interesting question for 2026 is not just “US model or Chinese model” but “hosted API or self-hosted weights.” Self-hosting an open model on your own cloud and DevOps stack adds operational work — GPU capacity, serving, monitoring — but it converts a metered, cross-border dependency into an asset you control, with data residency you can prove. For regulated sectors that trade-off increasingly favors self-hosting.
The third lesson is that provenance is becoming an audit artifact. As model sourcing gets more heterogeneous, it matters not just that you used AI but which model, which version and where it ran reached production for each feature. Record that mapping, keep your evaluation evidence, and be ready to explain the hosting path. The teams that will move fastest are the ones whose architecture treats the model as a swappable, well-documented component rather than a permanent, invisible assumption.
What to do this quarter
Here is the shippable version. Treat the Chinese-model cost shift as a prompt to make your AI stack both cheaper and more defensible.
- Add a routing layer. Put a provider-agnostic interface in front of your LLM calls so model, tier and vendor are configuration, not code.
- Benchmark on your own work. Test DeepSeek, Qwen and GLM against your current provider on your real tasks and quality metrics — not on public leaderboards.
- Right-size every call. Reserve premium US models for the hardest workloads; route the high-volume routine tasks to the cheapest model that clears the bar.
- Decide model and hosting separately. For regulated or sensitive data, run open weights inside your own US or EU environment instead of calling an offshore API.
- Map the compliance path. Document data residency for each AI feature and check it against GDPR, HIPAA and your customer commitments before you scale usage.
- Record provenance. Log which model, version and hosting path is in production for each feature, and keep evaluation evidence retrievable for audits.
None of this is legal advice, and your exact obligations depend on your sector and jurisdiction. But the strategic signal is clear: the price of capable AI is falling fast, much of the drop is coming from open-weight models, and the teams that benefit are the ones that can adopt them without importing risk. Treat the model as swappable, keep the data where the rules require, and the savings are yours to keep.
Frequently asked questions
Why are US companies adopting Chinese AI models?
Mainly cost. Open-weight Chinese models such as DeepSeek, Qwen and GLM run roughly 60% to 90% cheaper than the leading OpenAI and Anthropic models for comparable work, while closing much of the quality gap on coding and general tasks. Per OpenRouter data reported by CNBC on 7 July 2026, Chinese providers have supplied more than 30% of US-routed tokens every week since 8 February 2026, up from about 11% a year earlier.
How much cheaper are Chinese AI models than OpenAI and Anthropic?
OpenRouter estimates roughly 60% to 90% cheaper. In one illustrative comparison for a similar body of work, Anthropic’s Claude cost about $4,811, OpenAI about $3,357, and Zhipu’s GLM about $544. Actual savings depend on your workload, caching and quality bar, but the per-token cost drop is large.
Is it safe to use Chinese AI models for regulated data?
It depends on how you deploy them. Sending regulated data to an API hosted in China raises data-residency, security and geopolitical concerns and has already drawn US congressional scrutiny. Because DeepSeek, Qwen and GLM are open-weight, a safer pattern is to run the weights on your own US or EU infrastructure so data never leaves your jurisdiction. Treat the hosting location, not just the model, as the compliance decision.
Are DeepSeek and Qwen open source?
They are open-weight: the trained weights are published under permissive licenses, so you can download and run them on your own servers or a cloud of your choice. That is not the same as fully open training data and code, but it is enough to self-host, fine-tune and keep inference inside your own environment — the property that lets regulated teams capture the savings without sending data to an offshore API.
What should software teams do about the model cost shift?
Put a provider-agnostic routing layer in front of your LLM calls, benchmark open-weight Chinese models on your own tasks, and for regulated data host the open weights inside your own US or EU environment rather than calling an offshore API. Track token cost per feature, keep a capable fallback wired in, and document which model and hosting path each feature uses for audits.
Sources
CNBC — Chinese AI models are gaining ground with U.S. companies as OpenAI, Anthropic costs surge
Rest of World — When Americans choose Chinese AI
Forbes — China’s DeepSeek V4 and Qwen reshape the open-source AI race