China's Seven Largest AI Developers Generate a Tenth of OpenAI-Anthropic Revenue

The Revenue Chasm

Seven of mainland China's most prominent foundation-model developers pulled in an estimated $10.7 billion in annualised recurring revenue between March and August this year, according to data from Rhodium Group. That figure represents roughly a tenth of the combined revenue OpenAI and Anthropic are currently generating, a gap that underscores the distance between valuation hype and commercial traction in the world's second-largest AI market.

The gap is particularly striking given the intensity of capital flows into Chinese large-language-model startups over the past eighteen months. At Opentechwire, we've tracked more than a dozen funding rounds north of $200 million across Hangzhou, Beijing and Shenzhen since early 2025, with term sheets often pricing companies at multiples that assume rapid scale. The Rhodium analysis suggests that scale has yet to materialise at the top line.

OpenAI's ARR crossed $100 billion earlier this year when combined with Anthropic's disclosed figures, driven by enterprise API contracts, ChatGPT subscriptions and embedded inference deals with Microsoft, Salesforce and others. The Chinese cohort, by contrast, remains heavily reliant on API calls sold at thin margins and pilot programmes that have not yet converted into multi-year commitments. Several of the seven developers Rhodium examined have yet to disclose pricing for production-grade fine-tuning or batch inference, which in the US market now account for a substantial share of revenue per customer.

Export Controls and the Latency Tax

Part of the monetisation lag can be traced to the semiconductor restrictions Washington imposed in October 2022 and tightened last year. Without access to Nvidia's H100 and H200 clusters, mainland developers have leaned on older A100 inventory, Huawei Ascend 910B accelerators and smaller configurations of domestic chips that deliver lower throughput. The result is higher per-token cost and longer response times, both of which dampen enterprise appetite for production workloads that demand sub-200-millisecond latency.

We spoke to procurement leads at two regional banks in Singapore and one logistics platform in Jakarta, all of which evaluated Chinese models during proof-of-concept phases in the first half of this year. Each cited inference speed and the difficulty of securing dedicated capacity as reasons they ultimately signed with either OpenAI or Google. One procurement director noted that while the Chinese vendor's pricing was 30 per cent lower on paper, the lack of a committed service-level agreement made the discount unattractive for a system handling real-time fraud detection.

Export controls also constrain the ability to train and serve truly large models. Five of the seven developers in the Rhodium cohort run foundation models with parameter counts below 200 billion, compared to the 400-billion-plus configurations OpenAI and Anthropic now deploy in production. Smaller models can be effective for narrow tasks, but they struggle to match the reasoning depth and multi-step planning that enterprise customers increasingly expect, particularly in legal research, code generation and financial analysis.

Domestic Demand and the Price Floor Problem

China's domestic market offers enormous potential, but converting that potential into revenue has proven harder than many investors anticipated. Local enterprises remain cautious about committing large budgets to generative AI, in part because regulatory guidance on data residency and content filtering is still evolving. The Cyberspace Administration of China has issued multiple rounds of draft rules since mid-2025, each requiring adjustments to how models handle sensitive queries and store training data. Compliance overhead eats into margin, and uncertainty delays purchase orders.

Price competition is fierce. At least a dozen smaller model shops, backed by provincial government funds or Tencent-affiliated venture arms, have entered the market over the past year, many offering API access below cost to gain share. That dynamic has made it difficult for the seven leading developers to sustain pricing above $0.002 per thousand tokens for standard inference, roughly a quarter of what OpenAI charges in North America and Europe. Volume can offset low unit prices, but only if volume arrives quickly; so far, query growth has been steady rather than explosive.

Consumer applications, which drove early enthusiasm, have also disappointed. Several chat apps built on domestic models attracted tens of millions of downloads in late 2025, but retention rates dropped sharply once the novelty wore off. Advertising revenue from free-tier users has not compensated for the compute cost of serving them, and converting free users to paid subscribers has proved difficult in a market where Baidu, Alibaba and Tencent offer their own model-backed services at minimal or zero cost as part of broader platform bundles.

The Capital-Efficiency Question

Investors continue to pour capital into Chinese AI, but the Rhodium data raises questions about how efficiently that capital is being deployed. If seven well-funded developers are collectively generating revenue equivalent to a single quarter of OpenAI's trailing ARR, the path to positive unit economics looks long. Training runs for frontier models now cost upwards of $100 million when factoring in cluster time, data acquisition and engineering talent; serving those models at scale adds another layer of expense that only pays off if revenue per customer climbs steadily.

Some of the developers in Rhodium's analysis have begun shifting strategy, de-emphasising consumer chat in favour of vertical solutions for manufacturing, logistics and municipal government. Those segments offer stickier contracts and higher willingness to pay, but they also require deep customisation and longer sales cycles. One Shenzhen-based model shop we follow recently announced partnerships with three provincial transport bureaus to optimise bus-route scheduling, a use case that generates modest but predictable revenue and helps demonstrate return on investment to other government buyers.

Whether that pivot can close the revenue gap with the US leaders remains an open question. OpenAI and Anthropic benefit from mature enterprise go-to-market teams, established relationships with global systems integrators and a regulatory environment that, while tightening, still permits broad experimentation. Chinese developers face a more fragmented domestic market, ongoing export-control pressure and the challenge of building trust with international customers wary of data sovereignty and geopolitical risk.

What the Gap Means for Asia's AI Landscape

The revenue disparity has implications beyond the balance sheets of the seven developers. It signals that access to capital and engineering talent, while necessary, is not sufficient to build a sustainable foundation-model business. Distribution, pricing power and the ability to capture enterprise workloads matter just as much, and in those dimensions the US incumbents retain a significant lead.

For the rest of Asia, the gap creates both risk and opportunity. Southeast Asian enterprises that hoped domestic Chinese models would offer a credible, lower-cost alternative to US platforms may need to reassess their procurement roadmaps. At the same time, the slower-than-expected monetisation in China leaves room for regional players in Seoul, Tokyo and Singapore to carve out niches, particularly in languages and regulatory contexts where neither US nor Chinese models have deep expertise.

Rhodium's analysis covers the period through August, and the Chinese developers have several months left in the year to accelerate revenue. But the ten-to-one ratio is a sobering benchmark, one that suggests the gap between valuation and commercial reality in China's AI sector is wider than many participants have been willing to acknowledge.