Is China winning the AI race? On the metric that decides which models the world actually builds on — raw usage — the answer as of July 2026 is increasingly yes. Chinese open models now process the majority of traffic on OpenRouter, the largest marketplace developers use to plug into AI systems, and they do it at prices US frontier labs cannot match. But volume is not the same as value, and the gap between those two words is the whole story.

This breakdown draws on a July 2026 analysis by Coin Bureau host Lewis, who traced how a 12-month collapse in American market share happened, who is profiting, and why ordinary investors may be exposed without knowing it.

Key takeaways

  • According to the Coin Bureau analysis, American models ran roughly 72% of OpenRouter traffic 12 months ago; as of mid-2026 that figure is about 33%, with Chinese models north of 50% and climbing.
  • In the week ending June 21, 2026, Chinese models processed 21.37 trillion tokens on OpenRouter versus 5.76 trillion for American models — DeepSeek alone is now bigger than Google and OpenAI individually.
  • US startups including Lindy, Coinbase and Pinterest have shifted large shares of their AI workloads onto Chinese open models, cutting inference costs by roughly 90% in some cases.
  • The likely cause is ironic: US export controls cut China off from top Nvidia chips, forcing labs like DeepSeek to win on efficiency and price instead.
  • China dominates volume, not value — Anthropic still captures ~46% of OpenRouter’s dollar revenue on just 12% of tokens, meaning the premium reasoning market remains American for now.

Is China winning the AI race? What the OpenRouter data shows

The clearest evidence that China is winning the AI race sits in the usage data. According to the Coin Bureau analysis, American models ran about 72% of all traffic on OpenRouter roughly a year ago. Today that share is near 33%, while Chinese models have crossed 50%.

The crossover was sudden. In early 2025, individual Chinese systems like DeepSeek were running under 2% of the platform’s traffic, and the entire Chinese stack sat below 20%. In the week of February 9–15, 2026, Chinese models out-processed American ones for the first time — 4.12 trillion tokens against 2.94 trillion — and the line never crossed back. By the week ending June 21, 2026, the split was 21.37 trillion tokens (Chinese) to 5.76 trillion (American).

This is not a story of a shrinking pie. OpenRouter’s total volume grew roughly 11–12 times year over year. China is capturing the explosion of new, cost-sensitive demand that cheap models created in the first place. DeepSeek by itself now accounts for an estimated 16–18% of all traffic on the platform — more than either Google or OpenAI individually.

The money has noticed. In mid-June 2026, DeepSeek closed a funding round of around $7.4 billion at a valuation north of $50 billion, with Tencent contributing roughly $1.5 billion and battery giant CATL another $735 million. Founder Liang Wenfeng committed nearly $2.75 billion of his own money — and structured the deal so most investors received zero voting rights and a five-year lockup. Investors are paying billions for a seat with no steering wheel.

How US export controls backfired

Here is the irony that ties the story together: China did not choose efficiency because it is elegant. It was forced into it. American export controls cut Chinese labs off from Nvidia’s top chips, pushing DeepSeek and others onto restricted hardware and domestic Huawei Ascend processors. With brute-force scaling off the table, they had to innovate on doing more with less. As Perplexity’s CEO put it, by forcing Chinese firms to engineer around hardware limits, the US may have accidentally turbocharged the very competition it was trying to contain.

Then Washington escalated. In June 2026, an executive order and a Commerce directive extended restrictions from chips to America’s most advanced commercial models — with Anthropic suspending global access to its frontier systems to comply, and OpenAI previewing its newest model to only a small, government-vetted list. It was the first time export-control authority had been pointed at a deployed AI model rather than hardware. Every developer, government and startup outside that approved list was effectively locked out of the best American AI — and reached instead for the free, self-hostable Chinese checkpoint anyone can download.

Why US startups are switching to Chinese AI models

The volume flip is not abstract. Real companies are moving real money onto Chinese engines.

  • Lindy, a US AI-agent startup, found its inference bill had grown larger than the company’s entire payroll. It moved 100% of its traffic off Anthropic’s Claude and onto DeepSeek’s V4 Flash, cutting inference costs by roughly 90%. To manage the security risk, it hosts the Chinese models on US soil through a provider called Atlas Cloud, so data never routes through China.
  • Coinbase CEO Brian Armstrong confirmed the exchange built an internal gateway that defaults engineers to Chinese open-weight models — specifically Zhipu’s GLM 5.2 and Moonshot’s Kimi — for routine work. AI spend fell by about half, the cache hit rate jumped from 5% to 60%, and 91% of engineers stopped hitting their usage caps.
  • Pinterest went further, taking Alibaba’s open Qwen model, gutting its vision layer and fine-tuning on its own visual-preference data. Its CTO reported a 90% cost reduction and a 30% accuracy boost on Pinterest-specific tasks.

The pattern is the point: Qwen and DeepSeek function as foundational platforms developers can build on and own, not just APIs they rent. On Vercel’s developer gateway, DeepSeek’s share of traffic went from under 1% to 17% in a single month. Industry estimates now suggest the majority of new US AI startups are building on Chinese base models underneath. This is the same convergence of cheap open intelligence and application builders driving the AI supercycle in compute and energy.

The mixture-of-experts trick behind DeepSeek’s prices

The price gap is almost absurd. DeepSeek’s V4 Flash costs about $0.14 per million input tokens; the comparable American frontier model, GPT-5.5, runs around $5 — roughly 36 times more expensive for the same unit of work.

The engine behind that number is a design called mixture of experts (MoE), which avoids switching on the whole model for every question. DeepSeek’s architecture, pioneered with V3, uses 671 billion parameters in total but activates only about 37 billion for any given query — an activation rate near 5.5% that slashes compute needs by more than 90%. V4 Flash pushes further, with 284 billion total parameters and just 13 billion activated per query. That V3 foundation was reportedly trained for about $5.6 million in direct compute for a single run, even though the total infrastructure bill runs into the billions.

Inside China, daily token consumption has gone from around 100 billion in early 2024 to roughly 140 trillion by March 2026. That is infrastructure scale — the AI equivalent of becoming the default plumbing everyone else builds on top of.

Volume dominance is not value dominance

If you stop at the usage charts, you walk away with the wrong conclusion. China dominates volume; it does not dominate value.

By platform analysis cited in the video, Anthropic captures around 46% of OpenRouter’s actual dollar revenue while serving just 12% of the tokens. The American frontier still wins the high-stakes, complex reasoning work where cost is the last thing anyone cares about. Analysts reckon the US still leads on deep reasoning and agentic error recovery by anywhere from three to eight months, and Perplexity’s CEO puts the lead closer to a full year. The market has split into a premium American stack that sells reasoning by the carat and a Chinese industrial stack that sells intelligence by the ton, almost for free.

There is also a security catch. When you call a Chinese model’s own API directly, requests route through servers in China — and under China’s 2017 National Intelligence Law, those companies must cooperate with the state on request. That is precisely why Lindy pays to self-host on US soil. The moat fight is turning ugly, too: in June 2026, Anthropic told the US Senate that Alibaba’s Qwen lab had run the largest known distillation attack to date, using roughly 25,000 fake accounts to pull 28.8 million exchanges out of Claude. And a hard ceiling remains — China’s access to advanced AI chips is estimated at about 300,000 units for the year, with Huawei’s best processor running near 60% of an Nvidia H100, while the US still controls roughly 74% of the world’s high-end AI compute.

Why your pension is exposed to the AI bet

This split is a problem for anyone with money in the market, because America has made a colossal wager on the luxury side. The big hyperscalers are set to spend over $725 billion on capital expenditure in 2026 alone — up around 77% year over year — with Goldman Sachs projecting the figure crosses $1 trillion annually by 2027. Sequoia has flagged a roughly $600 billion gap between what is being spent and what AI actually earns, and Oracle recently posted negative free cash flow of nearly $24 billion for the year.

If you hold an S&P 500 index fund, you own a slice of that bet whether you chose it or not. If you have a pension, it is leaning on the assumption that these companies can monetize a frontier lead — a lead that may be only a few months wide and shrinking while the rest of the planet standardizes on free Chinese infrastructure. That is the same fragility we mapped in when the AI bubble might pop, and it is why the plumbing beneath AI — including the crypto rails AI agents increasingly need to transact — matters as much as the models on top. The era of assuming a permanent, unassailable American AI moat is ending; it is starting to look more like a luxury brand about to be commoditized.

Frequently asked questions

Is China winning the AI race in 2026?

On usage, yes — as of mid-2026, Chinese open models process the majority of traffic on developer marketplaces like OpenRouter, and DeepSeek alone handles more than Google or OpenAI individually. But the US still captures the bulk of premium revenue and leads on complex reasoning by an estimated few months to a year, so China leads on volume, not value.

Why are US startups switching to Chinese AI models?

Cost. Chinese open models such as DeepSeek and Alibaba’s Qwen can be roughly 36 times cheaper per token than US frontier models, and because they are open-weight, companies like Lindy, Coinbase and Pinterest can self-host and fine-tune them — cutting inference bills by up to 90% while keeping data on their own infrastructure.

Why is DeepSeek so cheap?

DeepSeek uses a mixture-of-experts architecture that activates only a small fraction of its parameters per query — about 37 billion of 671 billion in V3 — cutting compute needs by more than 90%. US export controls also denied it top Nvidia chips, forcing it to optimize for efficiency rather than brute-force scale.

Is it safe to use Chinese AI models like DeepSeek?

The models themselves are open-weight and downloadable, but calling a Chinese provider’s own API routes data through servers in China, where the 2017 National Intelligence Law compels cooperation with the state. Companies mitigate this by self-hosting the open models on domestic infrastructure so the data never leaves their control.