OpenAI and Anthropic widen revenue gulf

The new scoreboard is revenue, not raw capability

The latest revenue comparison between American and Chinese AI labs is blunt: according to Rhodium Group estimates reported by CNBC, all Chinese AI models combined generate only about 10% of the revenue reported for OpenAI and Anthropic [1]. The finding lands at an important moment because Chinese models are no longer dismissed as second-tier on capability; several have become competitive enough to force US labs, investors and policymakers to treat them as serious frontier challengers [3]. But the commercial scoreboard tells a different story: capability may win headlines, while distribution, enterprise contracts and access to customers willing to pay Western software prices still decide revenue.

The numbers behind the gap are striking. Rhodium’s estimates put OpenAI at about $40 billion in annual recurring revenue and Anthropic at about $65 billion, while Chinese leaders remain far smaller: DeepSeek at $500 million, MiniMax at $800 million, Moonshot at $1 billion, Z.ai at $1.8 billion, Alibaba at $2.4 billion and ByteDance at $4 billion [1]. Even allowing for different business models, reporting lags and rapid growth in China, the comparison suggests that the two leading US frontier labs have built a monetization machine that China’s model ecosystem has not yet matched.

Annual recurring revenue is an imperfect metric because it annualizes a recent monthly figure and can exaggerate momentum during a hypergrowth phase [1]. Still, it is the metric investors are using to value AI labs, and it is useful precisely because these companies are growing too quickly for traditional annual revenue comparisons to feel current. On that basis, the US advantage is not just bigger; it is qualitatively different. OpenAI and Anthropic have converted frontier models into paid products, enterprise deployments, developer APIs and platform relationships at a scale that Chinese rivals have not yet achieved.

China’s models are catching up, but its business model is not

The most important nuance is that this is not a simple “US models good, Chinese models bad” story. Associated Press reporting this week described China as rapidly closing the AI gap with the United States, noting that Chinese systems from Moonshot, DeepSeek, Z.ai and Alibaba have challenged US leaders in capability, cost and market attention [3]. Moonshot’s Kimi K3, for example, was ranked third globally in model intelligence in July, behind Anthropic and OpenAI systems, before slipping to ninth as newer frontier models appeared [3]. That volatility is the point: benchmark leadership is fluid, and China has shown it can move fast.

DeepSeek helped establish the current pattern by proving that cost-efficient Chinese models could challenge the assumption that only gigantic US spending produces useful frontier performance [3]. Z.ai’s GLM-5.2 and GLM-5.3 releases also drew attention for coding and agentic capabilities, while Alibaba’s Qwen3.8-Max was presented as a direct challenge to Claude and GPT [3]. From a technical perspective, Chinese labs are close enough that every new release can alter developer sentiment. From a commercial perspective, however, the gap remains wide.

That gap exists partly because open-weight and low-cost strategies are double-edged. Chinese labs have often gained adoption by making models cheaper and easier to run outside a closed platform. That helps diffusion, especially among cost-conscious developers and enterprises, but it can limit how much revenue flows back to the model creator. Rhodium noted that Chinese AI labs are exploring ways to capture a larger share of the revenue earned by third parties that provide access to their models [1]. In other words, Chinese AI may be spreading, but the invoice may be captured elsewhere.

The US model is different. OpenAI and Anthropic mostly sell closed or tightly controlled systems through direct subscriptions, APIs, enterprise seats and cloud-platform relationships. That structure can frustrate open-source advocates, but it creates pricing power. It also lets the labs attach revenue to trust, compliance, support, procurement and developer tooling. For large Western companies, the model is not just the intelligence of the chatbot; it is the contract, the uptime promise, the security posture, the integration path and the confidence that the vendor will still be around after the pilot.

Z.ai shows both the opportunity and the constraint

Z.ai is the Chinese company that most clearly illustrates how fast the situation could change. The company raised its year-end annual recurring revenue outlook by 25%, from $2.4 billion to $3 billion, after a $5 billion fundraising round that it said eased near-term computing bottlenecks [2]. Executives said during an investor call that Z.ai’s ARR had reached $1.8 billion so far this year, and that short-term compute supply was no longer the main constraint on rapid revenue growth [2]. That is a meaningful acceleration and one reason the revenue gap should be treated as a moving target, not a permanent law of nature.

Still, the Z.ai update also underlines the central problem for China’s AI labs: monetization depends on capital, compute and paying customers arriving at the same time. A model can become popular before its creator has enough inference capacity, enterprise sales infrastructure or global distribution to harvest the demand. Z.ai’s new funding may reduce one bottleneck, but it does not instantly solve the broader challenge of turning Chinese model adoption into durable, high-margin, recurring revenue.

Valuation makes the issue sharper. Rhodium’s analysis said valuations relative to revenue look especially stretched for Moonshot and DeepSeek, estimating revenue multiples of 50 times and 163 times respectively [1]. By comparison, the same analysis put OpenAI at 34 times and Anthropic at 21 times [1]. Those multiples are still high by ordinary software standards, but the contrast matters: the American leaders appear expensive because investors believe revenue is already material and compounding, while some Chinese rivals appear expensive because investors are pricing in revenue that has not yet arrived.

The AI race is becoming a capital-markets race

The revenue gulf reframes the geopolitical contest. For much of the past two years, the debate focused on chips, benchmarks and whether Chinese labs could match US model quality despite export controls. Those questions still matter. But the new Rhodium comparison suggests that the next phase is a monetization race. The winner will not simply be the country whose models place highest on benchmark tables; it will be the ecosystem that can turn intelligence into contracts, usage, renewals and cash flow.

That is why the US advantage is broader than model weights. OpenAI and Anthropic sit inside a Western software economy with deep enterprise budgets, established procurement channels, venture-backed developer ecosystems, cloud partners and customers accustomed to paying premium prices for productivity tools. Chinese labs have enormous domestic scale and aggressive cost structures, but their ability to monetize globally is limited by geopolitics, trust concerns, procurement barriers and the fact that open-weight adoption does not automatically become vendor revenue.

Rhodium also found that more than 60% of equity investment in Chinese AI chips and servers came from state-affiliated sources [1]. That support can help China build infrastructure, but it is not the same as customer revenue. Government-backed compute spending can keep the race alive; recurring commercial demand is what funds the race sustainably. If frontier models require vast and rising capital expenditure, then the labs with the strongest revenue engines can reinvest faster, subsidize more experimentation and survive more failed training runs.

Safety politics now sits on top of revenue pressure

The revenue story is also unfolding during a new political argument over whether frontier AI development should slow down. This week, AP reported that Anthropic’s Dario Amodei, OpenAI’s Sam Altman and other AI leaders have voiced support for pacing AI development, while competition, profit motives and opposition from President Donald Trump complicate any coordinated slowdown [5]. Le Monde reported that Trump rejected a pause by citing competition with China, saying the US must keep its lead in AI [4].

That debate matters for monetization because safety, regulation and capital markets are becoming inseparable. Frontier labs need investor confidence to fund compute. They also need public confidence to sell into regulated industries. A major safety failure could damage revenue just as quickly as a weak model release. At the same time, a slowdown that only US labs obey could create political fear that Chinese competitors will close the gap faster. The result is a difficult triangle: move fast enough to lead, cautiously enough to be trusted and profitably enough to finance the next generation.

What the gulf really means

The simplest reading is that OpenAI and Anthropic are not merely AI research organizations anymore. They are becoming commercial infrastructure companies. Their models sit inside workflows, coding tools, customer-service systems, office software, analytics stacks and enterprise procurement plans. That is why their revenue lead over China is so large even as China’s technical progress remains real.

For Chinese AI labs, the path forward is not just better benchmarks. They need stronger paid distribution, more enterprise trust, clearer pricing power and mechanisms to capture value when their open models are used by third parties. Z.ai’s new forecast shows this can improve quickly, but Rhodium’s comparison shows how far there is to go [1][2].

The US-China AI contest is therefore not only a race to build the smartest model. It is a race to build the most monetizable model ecosystem. For now, the scoreboard says OpenAI and Anthropic have turned frontier capability into commercial power far more effectively than their Chinese counterparts. Capitalism, as usual, has found a way to make the achievement expensive.

Sources from the last 72 hours

1. [1]OpenAI and Anthropic are making 10 times more revenue than all Chinese AI models combined, research group Rhodium saysSep 17, 2026, 12:00 AM UTC
2. [2]China’s Z.ai raises revenue target 25% after US$5 billion cash injectionSep 16, 2026, 1:31 PM UTC
3. [3]Trump rejects pause in AI race citing competition from ChinaSep 14, 2026, 1:30 PM UTC
4. [4]China is closing the AI gap with the US as concerns rise over safetySep 16, 2026, 12:00 AM UTC
5. [5]Slowing down AI: What would that look like and how possible is it?Sep 16, 2026, 4:14 AM UTC

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.