Something significant is happening in enterprise AI procurement, and most coverage misses it entirely. US companies are not just complaining about AI costs — they are quietly routing real production workloads to Chinese models.

According to CNBC reporting on July 7, 2026, the share of tokens consumed by US companies on Chinese AI models via OpenRouter has sat above 30% every single week since February 8. At peak, that figure hit 46%. This is not hobbyist experimentation. This is deliberate vendor diversification driven by a simple economic calculation.

The Cost Gap That Drove the Switch

Token prices at the major US AI labs have been climbing. OpenAI announced it would limit the rollout of certain new models at the government’s request late in June. Anthropic’s Fable 5 and Mythos models spent weeks under export controls before returning globally on July 1. The combination of rising prices and availability uncertainty pushed procurement teams to look elsewhere.

What they found surprised many of them.

Chinese open-source models are now 60% to 90% cheaper than leading Anthropic and OpenAI offerings. That gap is not just about cost — it is about performance per dollar. Z.ai’s GLM 5.2, released to significant attention in June, landed within a single percentage point of Anthropic’s Opus 4.8 on one closely watched agentic benchmark. The price difference: roughly one-fifth the cost.

The adoption numbers back this up. Vercel tracked GLM 5.2’s deployment in its first full week. Daily token volume grew approximately 27 times. The number of customers using it grew approximately 80 times. Those are not typical adoption curves.

What “Competitive” Actually Means Now

DeepSeek generated headlines earlier in 2026, but Z.ai’s GLM 5.2 represents a further step. When a model can perform agentic tasks — multi-step reasoning, tool use, autonomous workflows — at near-frontier quality for a fifth of the price, it changes the ROI equation for every business deploying AI at volume.

Enterprise token consumption is not linear. Agentic workflows compound it. A single customer inquiry resolved by an AI agent may involve dozens of model calls: understanding the request, querying a knowledge base, drafting a response, checking compliance, formatting the output. At OpenAI or Anthropic prices, that chain is expensive. At Chinese model prices, it changes the business case entirely.

The Risks Nobody Is Talking About Loudly Enough

Cost wins matter. But this shift introduces genuine complications that business leaders need to weigh.

Data sovereignty. When tokens travel to a Chinese model provider’s infrastructure, so does the content of those tokens. For businesses handling customer data, financial records, or anything sensitive, this deserves a serious legal and compliance review before routing production workloads.

Regulatory exposure. The US government’s relationship with Chinese AI models is unsettled. Export controls on Anthropic’s models and OpenAI’s restricted rollout were both driven by government intervention in the space. The same dynamics could affect access to Chinese models going the other direction.

Support and accountability. When something goes wrong in a production AI workflow — a wrong answer, a compliance failure, a bias issue — you need a clear line of accountability. That line is thinner with offshore model providers.

Benchmarks are not production. GLM 5.2 being within a percentage point of Opus 4.8 on one benchmark is meaningful. It is not the same as being tested on your specific data, your specific workflows, your edge cases.

What This Means for Business

This is not a reason to panic or to immediately switch vendors. It is a reason to run a proper AI vendor evaluation — which surprisingly few businesses do before committing to a model provider.

A few practical questions every business deploying AI should be asking right now:

What are your actual token costs month to month? If you are running agentic workflows at scale, you should know this number to two decimal places. If you do not, you will be shocked when the bill arrives.

Have you benchmarked alternatives on your own data? Generic benchmarks tell you what the industry average looks like. Your use case is not the industry average. Spend a day running your ten most common AI tasks across three or four models before assuming one is definitively best.

What data flows through your AI infrastructure? If the answer includes anything customer-facing, regulated, or proprietary, you need a data classification exercise before you route it anywhere new.

Is vendor concentration a risk for your business? The export control situation with Anthropic in June showed what happens when a single AI vendor becomes unavailable unexpectedly. Enterprises that had already run multi-model evaluations were able to pivot quickly. Those that had not were scrambling.

The market is fragmenting. That is actually a healthy development — competition drives costs down and quality up. But fragmentation also means you need a strategy, not just a default. The businesses that will navigate this well are the ones building genuine AI competency in-house, not just buying subscriptions and hoping for the best.

Enterprise DNA exists for exactly this moment. Whether that means building the data literacy your team needs to evaluate AI tools critically, or working with Omni Advisory to develop an AI procurement strategy that accounts for cost, risk, and capability — the companies that stay passive will keep paying frontier prices for results they could get more efficiently.

The Chinese AI cost shift is a market signal. Treat it as one.