China's AI Models Won the Token War. The Money Stayed With the GPUs.

On May 25, DeepSeek cut the price of its flagship model by 75% — output tokens, the expensive kind, went from $3.48 per million down to $0.87. That was a model that was a month old. It is the clearest signal yet that the Chinese labs are not trying to win on benchmarks. They are winning on price, and they are winning.

On OpenRouter, the largest neutral router where developers direct their prompts to competing models, Chinese models processed 28.1 trillion tokens in the week of July 28 to August 3. U.S. models processed 4.4 trillion. Nine of the ten most-used models in the world were Chinese. DeepSeek's V4-Flash was the single most-used model on the platform. By one widely cited estimate, Chinese models now handle about 60% of the tokens U.S. companies route through the system — less than 10% a year earlier.

So the answer to "are China's AI models the new global stars" is, by usage: yes. Uncomfortably so.

That's also the wrong question for a U.S. investor to ask. The scoreboard answers a different question: where did the money go? And the answer is visible in a single income statement.

The star is a price story

The adoption story is a pricing story. A token is the unit of text an AI model processes, and everyone bills per million of them. DeepSeek V4-Pro now lists at $0.87 per million output tokens. OpenAI's flagship, GPT-5.6 Sol, lists at $30. Run the same standardized workload — 30 million input tokens and 30 million output tokens per month — and DeepSeek's value model bills about $13. The Pro model bills about $39. OpenAI's flagship bills about $1,050. An AI assistant startup called Lindy said it cut its inference costs by 90% after moving parts of its workload from Anthropic to DeepSeek. Airbnb's CEO has called Alibaba's Qwen "very good" and "fast and cheap."

The wedge is coding. Programming now accounts for more than half of OpenRouter usage, up from about 11% in early 2025, and Chinese models are disproportionately strong — and cheap — at exactly this workload. Research from Andreessen Horowitz puts it at roughly 80% of developers worldwide using open-source tools are building on Chinese models.

The American labs are not standing still. OpenAI cut the price of GPT-5.6 Luna, its cheapest model, by 80%, and Anthropic introduced a lower-priced Claude Opus 5 tier at $5 per million input tokens and $25 per million output tokens. Yet both report record enterprise revenue through the cuts — OpenAI's enterprise business is reportedly running at a roughly $40 billion annualized pace, and Anthropic's preliminary second-quarter revenue topped $11.5 billion. Volume is growing fast enough to absorb lower prices per token.

Two caveats before I draw the conclusion. OpenRouter is one router, and a developer-skewed one — token share is not the same thing as total enterprise AI spending. And usage is not profit: a widely cited MIT study found 95% of organizations saw zero measurable P&L impact from tens of billions in generative AI spending. Treat "China won the AI market" as a usage fact, not a revenue fact. The model layer is where the margin is being competed away. That's the point, not the punchline.

Cheap tokens are a demand multiplier

Here is the mechanism that matters. When the price of a token falls, the number of tokens people run does not fall proportionally — it multiplies. OpenRouter's weekly volume has roughly quadrupled in the past year, from about 5 trillion tokens to more than 20 trillion, and the growth is concentrated in the cheap, open-weight tier that Chinese labs supply. This is the Jevons effect, the old energy-economics rule that efficiency gains increase consumption rather than reduce it: when an input gets cheaper, you run more of it, until use cases that were never economical become economical.

The input that multiplies hardest in the data center is not the model. It's the silicon. When a U.S. company decides Qwen or DeepSeek is "good enough" and hosts the model in its own data center, the weights are free — the inference is not, and that inference still runs on Nvidia-class GPUs in American buildings. The Chinese models are the cheapest thing in the stack. That is precisely what makes the rest of the stack more valuable.

Nvidia's income statement is the proof. The quarter that ended July 26: revenue of $96.2 billion, up 106% from a year earlier. Data center revenue of $89 billion. Gross margin — the share of each sales dollar left after paying for the goods sold — of 75%. And the guide for the next quarter is $108 billion at a 74% margin, built on an assumption of zero data center compute revenue from China. China was already a rounding error: Hopper chip shipments to China were under 1% of data center revenue. Jensen Huang's phrase from the earnings call — "compute is revenue" — describes the whole mechanism. Cheaper inference everywhere is more compute purchased here.

Note what the structural split means. Some Chinese frontier models train on Huawei's Ascend silicon in China, and export controls keep the two markets separate. "China takes the AI market" is not a story that shows up in American income statements. The American buildout continues as if the headline doesn't exist — Nvidia has lined up more than $500 billion of third-party capital for AI infrastructure, and its neocloud partners are exiting 2026 with 8 gigawatts of installed capacity, up from about 3 gigawatts a year ago.

The market is not pricing all of that. Nvidia sits at roughly a $5.3 trillion market cap, about 27 times its trailing-year earnings, and the stock has drifted about 3% over the past 20 days despite another quarter that beat estimates. That is not the market doubting the current quarter. It is the market pricing the next leg — the question is whether the return profile still justifies the allocation, not whether the demand is real.

The generation gap that matters is American vs. American

The contested share of the AI chip market is not being lost to China. It is being contested inside the United States, by the company with the most credible generation answer to the economics of inference: AMD.

Inference rewards memory, cost per token, and time to production — not benchmark prestige. AMD's Helios rack, built around the MI455X GPU, offers 432 gigabytes of high-bandwidth memory versus 288 gigabytes in Nvidia's Vera Rubin NVL72 rack. That is roughly a 50% memory advantage in the exact bottleneck that big-model inference runs into. The commitments are gigawatt-scale: 6 gigawatts with OpenAI and 6 gigawatts with Meta, with deployments beginning in the second half of this year, and Anthropic planning up to 2 gigawatts. Data center was 58% of AMD's revenue last quarter, up from about 42% a year earlier, and it grew 107% year over year. Management expects data center revenue to more than double in 2027.

That is a dual signal, and it cuts both ways. The commitments prove the demand is real — no laggard gets gigawatt orders from the two largest AI buyers on earth unless the product clears qualification. But commitments are leverage too: they assume AMD can manufacture, deploy, and support the Helios ramp on schedule, and the market has already paid for that assumption. At roughly an $840 billion market cap — about 30 times expected 2027 earnings by recent estimates — the 2027 crossover is priced in. A meaningful stumble in Helios shipments pushes the inflection a year out, and a stock priced for perfection has no cushion for that.

So: are China's AI models the new global stars? By usage, they already are. And that star status is precisely what makes the token price war deflationary for the model layer, where the margins are, and inflationary for the compute layer, where the money is. The cheap Chinese token is the demand multiplier for the GPUs underneath it.

For your capital, the decision is not "China versus Nvidia." It is the allocation question the market is currently arguing with its drift: hold the 75%-margin compounder at 27 times trailing earnings and 106% growth, or buy the challenger at 30 times 2027 earnings with the ramp already priced in. I still believe the long-term compute thesis is intact — but much of the remaining return is likely back-half weighted, and the crowded part of the trade is exactly what requires the discipline. The sellers are not seeing Chinese models take the data center. They are deciding whether the next leg is worth the price they are paying for it.