Rhodium: U.S. AI Duo Out-Earns All of China's Models 10-to-1

Research group Rhodium Group said its latest analysis of Chinese and U.S. artificial intelligence companies found that OpenAI and Anthropic together generate 10 times more revenue than all Chinese AI models combined, and that low revenue did not necessarily correspond with lower valuation.

Two American companies are collecting ten dollars of artificial-intelligence revenue for every dollar earned by every Chinese AI model put together. That is the headline finding from Rhodium Group, the research firm whose latest analysis compared Chinese and U.S. AI companies and concluded that OpenAI and Anthropic between them out-earn the entire Chinese model field by a factor of ten.

The second finding is the more awkward one for anyone trying to price this race: according to Rhodium, low revenue did not necessarily correspond with lower valuation. Chinese model developers are being valued on something other than the money currently coming through the door.

A revenue gap, not necessarily a capability gap

It is worth being precise about what Rhodium measured, because the number is about monetization rather than model quality. A ten-to-one revenue gap says that American labs have converted their models into paid products — enterprise subscriptions, developer API calls, seat licenses inside software people already buy — at a scale Chinese developers have not matched. It does not say that Chinese models are ten times worse, and Rhodium's own framing, that revenue and valuation have come apart, implies investors do not read it that way either.

That distinction has been the central confusion of the past two years of AI commentary. Benchmark scores travel fast and cheaply; billing systems do not. The American advantage Rhodium is describing sits in the commercial layer — pricing power, distribution through existing enterprise relationships, and a customer base willing to pay in dollars for tokens.

Why Chinese models monetize thinly

Several structural features of the Chinese market push in the same direction, and none of them require a view on model capability.

Price competition. Chinese developers have competed hard on the per-token price of inference, which compresses revenue even when usage is heavy. A model that is used constantly and charged almost nothing for shows up as a small revenue line.
Open weights. Releasing model weights openly buys developer adoption and geopolitical reach, but it deliberately forgoes the subscription and API revenue that an American closed-model lab books as its top line.
Enterprise software base. The U.S. AI labs sell into a corporate market already accustomed to paying per-seat software fees. That habit is thinner elsewhere, so the same product converts into less cash.
Currency and cross-border reach. Revenue earned mostly at home, in renminbi, against a domestic price floor set by rivals, is a smaller number when stacked against dollar-denominated global enterprise contracts.

Put together, those forces explain how a field can be technically competitive and commercially small at the same time. Rhodium's figure is a measure of the gap between the two.

Valuation stops following the income statement

The part of Rhodium's analysis with the most direct market relevance is the decoupling. If low revenue does not translate into a low valuation, then investors in Chinese AI developers are underwriting something else: expected future monetization, strategic value to a state-backed technology stack, control of distribution inside a domestic market of enormous scale, or simply scarcity of listed exposure to the theme.

None of those are irrational on their own. All of them are harder to check than revenue. A valuation anchored to a shipped product with a price can be tested each quarter; a valuation anchored to an eventual monetization path can only be tested by waiting. That asymmetry is a risk investors on both sides of the Pacific are carrying, and Rhodium's finding is a reminder that the two markets are not being priced on the same evidence. The research was CNBC Finance reported on Thursday.

The tape is still paying up for the American side

Public markets did not wobble on the news — if anything, the AI-heavy end of the market had a good session. The Nasdaq 100 tracker (NASDAQ: QQQ) closed at $716.92, up 1.73% on the day from a prior close of $704.72, with a day range of $713.32 to $718.04, as of the last trade at 20:00 GMT on Sept. 17, 2026. The S&P 500 tracker (NYSEARCA: SPY) finished at $762.60, up 1.13% from $754.05, and the Dow tracker (NYSEARCA: DIA) closed at $518.35, up 0.61%.

Public markets did not wobble on the news — if anything, the AI-heavy end of the market had a good session.

The spread between those three is itself a small illustration of the point. The most technology-weighted of the benchmarks led, the broad index followed, and the industrial-heavy Dow lagged — an ordering that has characterized most of the AI trade's better days. Investors cannot buy OpenAI or Anthropic directly on an exchange, so the enthusiasm Rhodium's revenue figure implies gets expressed through the listed infrastructure and software names that sit inside those indexes.

What would change the picture

A ten-to-one revenue ratio is a snapshot, and the variables that could move it are identifiable.

Pricing discipline in China. If domestic developers stop undercutting each other on inference, revenue could rise sharply without a single new customer.
Export of Chinese models. Open weights are a distribution strategy. If that adoption abroad eventually attracts support contracts, hosting fees or enterprise tiers, the revenue line follows the install base.
Cost of compute. The American revenue lead is expensive to serve. Gross margin, not gross revenue, ultimately decides which side has the more durable business.
Enterprise budget cycles. A large share of U.S. AI revenue sits inside corporate software spending, which is discretionary. A tightening in that budget line would compress the numerator of the ratio rather than lift the denominator.

For now the read-through is narrow but useful. The two best-monetizing AI companies in the world are American and private, and the market's way of owning that fact is indirect. Rhodium has put a figure on how wide the commercial lead is; it has also pointed out that the figure is not what is setting prices on the other side of the comparison.

The measurement problem underneath

Comparing AI revenue across the two countries is genuinely hard, and that is part of why a single headline multiple carries so much weight. Neither OpenAI nor Anthropic is a listed company filing quarterly statements; Chinese model revenue is frequently bundled inside cloud, advertising or handset businesses that report at a higher level. Independent estimates such as Rhodium's are doing the work that audited filings would do in a more mature industry.

That argues for treating the ten-to-one ratio as a direction rather than a decimal. The direction is clear: commercially, the gap is wide. The valuation signal, per Rhodium, is pointing somewhere else entirely — and reconciling the two is the real analytical task for anyone allocating capital to this theme.

Key facts

Rhodium finding: OpenAI and Anthropic make 10x the revenue of all Chinese AI models combined
Valuation note: Low revenue did not necessarily correspond with lower valuation
QQQ (Nasdaq 100): $716.92, +1.73%, last trade 20:00 GMT Sept. 17, 2026
SPY / DIA close: SPY $762.60 (+1.13%); DIA $518.35 (+0.61%)

Frequently asked questions

What exactly did Rhodium Group find?

Rhodium Group's latest analysis of Chinese and U.S. artificial intelligence companies concluded that OpenAI and Anthropic together generate 10 times more revenue than all Chinese AI models combined. The research also found that low revenue did not necessarily correspond with lower valuation, meaning Chinese developers are being priced on factors other than current sales.

Does the revenue gap mean Chinese AI models are worse?

No. Rhodium's measure is about monetization, not capability. A revenue gap reflects pricing, distribution and business model — heavy price competition on inference, open-weight releases that forgo subscription income, and a thinner enterprise software-spending habit — rather than a direct verdict on how well the models perform.

Why would valuations stay high if revenue is low?

Investors may be underwriting expected future monetization, strategic value within a national technology stack, control of distribution in a very large domestic market, or scarcity of listed exposure to the AI theme. Rhodium's point is that those bases are harder to verify quarter by quarter than reported revenue is.

Can investors buy OpenAI or Anthropic shares?

Neither company is publicly listed, so there is no exchange ticker to buy. Exposure to the American AI revenue lead is typically expressed indirectly through listed infrastructure, semiconductor and software companies held inside broad benchmarks such as the Nasdaq 100 and the S&P 500.

How did markets close on the day the research was reported?

As of the last trade at 20:00 GMT on Sept. 17, 2026, the Nasdaq 100 tracker QQQ closed at $716.92, up 1.73% from a prior close of $704.72. The S&P 500 tracker SPY finished at $762.60, up 1.13%, and the Dow tracker DIA closed at $518.35, up 0.61%.

Why is AI revenue hard to compare between the two countries?

OpenAI and Anthropic are private and do not file quarterly public statements, while Chinese model revenue is often bundled inside larger cloud, advertising or device businesses. That is why independent estimates such as Rhodium's carry weight, and why the 10-to-1 figure is best read as a direction rather than a precise decimal.

Sources

OpenAI and Anthropic are making 10 times more revenue than all Chinese AI models combined, research group Rhodium says — CNBC Finance

Photo: Hugovanmeijeren · BY-SA 3.0 — source