Chinese Open AI Models Win Overseas Users, but Cloud Platforms Capture Much of the Revenue

DeepSeek leads usage as the fight shifts to revenue

DeepSeek accounted for 24.5% of text requests on OpenRouter in the week beginning October 5, 2026, ahead of OpenAI at 18.6% and Google at 17.8%. Yet an Epoch AI estimate found that DeepSeek captured only about 8% of customer spending on its models accessed through the platform. Chinese AI developers are winning usage overseas while competing with the companies that host and sell their technology for the money it generates.

OpenRouter's weekly usage rankings, with data through October 13, put DeepSeek V4.1 Flash first at 40.5 trillion tokens processed, followed by Step 5 Preview at 15.1 trillion and GLM 5.3 Flash at 12.9 trillion. Tokens are the small pieces of text or data that an AI model processes when it receives a question and produces an answer.

Chinese models had led US models in weekly token usage on OpenRouter for 24 consecutive weeks through October 11. Their appeal combines lower prices, improving performance and the freedom for businesses to adapt models and run them through a provider of their choice.

Cynthia Zhang, founding partner of FutureX Capital, says Chinese model developers are unlikely to match the revenue of OpenAI and Anthropic soon, although their sales are growing rapidly. Epoch AI estimated that six leading Chinese developers together generated about 10% as much AI revenue as those two US companies combined as of September.

What the usage figures measure

OpenRouter connects applications to many AI models through an application programming interface, or API. Its request chart counts how often models are called. Its token rankings count the amount of material processed, including both questions and answers. The figures describe activity passing through OpenRouter, with private requests excluded.

DeepSeek's request share exceeded OpenAI's by 5.9 percentage points in the October 5 week. Z.ai held another 13.4%, Qwen 5.4%, Xiaomi 2.9% and Tencent 1.7%. Together, those five Chinese developers accounted for 47.9% of text requests. Other Chinese developers may also appear in the remaining categories.

OpenRouter's daily token dataset contains the top 50 public models each day and an aggregate for all others. It supports daily, weekly and monthly periods, with filters for such features as modality and context length. Category and language filters use estimates derived from samples. Providers also divide material into tokens differently, so the totals depend partly on their counting methods. Source: OpenRouter (openrouter.ai/rankings), as of October 13, 2026.

Why businesses choose models they can adapt

Harvey, which builds AI tools for legal work, and Genspark, which develops agents using multiple models, are developing systems based on Chinese open technology to reduce reliance on closed providers. An AI agent is software that uses a model to carry out a sequence of tasks, such as planning work, writing code and checking the result.

Open weights are the learned numerical settings that guide a model's responses. Making them available lets another company download the model, run it on its own servers and adjust it for particular work. The exact rights depend on the licence. Some releases also include source code, while others provide weights with restrictions on commercial use.

Businesses can use proprietary data to adapt these models for their needs. They can also choose where the system runs, rather than sending every request to the original developer. That flexibility helps explain why a Chinese model can become part of an American or European product while its hosting and customer relationships stay outside China.

Intel chief executive Lip-Bu Tan described adoption of Chinese frontier models as "very fast" at a DBS panel in Singapore on October 12. He called the performance of DeepSeek, Kimi and GLM "good enough" and said markets would determine which models best fit different uses. Explaining his preference for open technology, Tan said:

I'm a strong believer of open source, that's the best way to drive efficiency,

Lower prices change everyday decisions

Stu Clott, an operations manager and occasional developer in San Diego, said an hour of coding cost about $10 with Claude, compared with less than 50 cents using DeepSeek. Those figures put his DeepSeek spending at less than one twentieth of the Claude cost for that work.

Ruben Garcia Jr., a Dallas developer building websites and mobile applications, divides work between providers. He spends $500 a month on Claude and ChatGPT for difficult planning and review, and $200 on Chinese models including MiniMax, Kimi and Xiaomi MiMo. He says the Chinese models handle 90% of his tasks, including coding and voice recognition.

Lindy founder Flo Crivello said moving his company's AI work assistants from Anthropic models to DeepSeek saved millions of dollars. Lindy accesses DeepSeek through an American provider. Crivello explained why a less expensive model can suit routine work:

You don't need God to write your email,

Kyle Chan, a fellow at the Brookings Institution, identifies businesses of medium size that are beginning to use AI but worry about costs as a growth market for Chinese developers. Zhang says competition can leave a leading US model with a period of high profitability lasting only about six months before prices start falling.

Performance is close enough for many tasks

The Center for Strategic and International Studies describes Chinese models as competitive in many practical tasks in its Chinese AI assessment. Author Yasir Atalan, deputy director and a data fellow at the CSIS Futures Lab, writes that GLM-5.2 has roughly 750 billion parameters and a context window of one million tokens. Parameters are the model's learned settings; a context window is the amount of material it can consider at once.

The assessment cites the Center for AI Standards and Innovation's finding that DeepSeek V4 Pro was about eight months behind leading US models. It also describes strong coding and agent results from GLM and Moonshot. For a business choosing a system, the relevant question is whether it performs the required task reliably at an acceptable cost.

A US government comparison gives a concrete example. The US-China Economic and Security Review Commission's Two Loops report lists January 2026 prices of $1.20 per million tokens for Kimi K2.5 and $4.81 for GPT-5.2. Both scored 47 on Artificial Analysis's Intelligence Index, putting Kimi's listed price at about one quarter of GPT-5.2's.

Chinese labs have pursued more efficient training and serving methods. DeepSeek said V3's final training run used 2.788 million H800 GPU hours, priced at about $5.6 million. CSIS also describes knowledge distillation, in which answers and judgments from a stronger model become training examples for another. OpenAI and Anthropic have accused Chinese labs of obtaining such material through fraudulent accounts, making the practice a point of commercial and political conflict.

A growing ecosystem reaches global platforms

Hugging Face download data cited by CSIS put Chinese models at 41% of downloads over the previous year. The commission's report says they overtook US models in total downloads in August 2025 and occupied seven of the platform's ten most downloaded positions from November to December that year.

Alibaba's Qwen family illustrates the scale of adaptation. The commission recorded more than 100,000 derivative models at publication. By summer 2026, Qwen had about 2.045 billion downloads and 151,448 derivatives, roughly 2.6 times the number based on Meta models. A derivative is a version built from an existing model, often adjusted for a language, industry or specific job.

Amazon Web Services has added six Chinese open weight models to Amazon Bedrock, including DeepSeek V3.2, MiniMax M2.1 and Qwen3 Coder Next. Kimi K3 joined in September, followed by GLM-5.3 in early October. Bedrock gives businesses a route to use models through Amazon's cloud infrastructure. Baseten has also made Kimi K3 and GLM-5.3 available within OpenAI's Codex coding tool, handling hosting and the computing needed to generate responses.

The expansion has moved through several stages:

• August 2025: Chinese models overtook US models in total Hugging Face downloads.
• November to December 2025: Chinese models held seven of Hugging Face's ten leading download positions.
• July 2026: Moonshot released Kimi K3 with revenue sharing requirements for some large service providers.
• September 2026: Kimi K3 became available on Amazon Bedrock.
• Early October 2026: GLM-5.3 reached Bedrock through a commercial arrangement involving revenue sharing.
• October 12, 2026: Intel's chief executive described rapid adoption at the DBS panel in Singapore.

Who gets paid when an open model spreads?

Z.ai's experience shows how a model can gain popularity while its creator loses hosting business. Epoch AI found that 12 outside providers offered GLM-5.3 Flash through OpenRouter within a day of its weights being released in August. After 20 days, 26 providers offered it, with eight charging less than Z.ai.

Z.ai's share of tokens served for that model fell from 88% to 22%. Its own daily token volume declined 70%, even as total use of the model rose 17%. Customers were increasingly buying access from other hosts.

Vercel recorded a similar gap between activity and income: DeepSeek's token share rose from under 1% to 17% in May, while its revenue share stayed near 1%. Cheap models can account for a large amount of work while producing a much smaller share of spending.

Z.ai reported an API gross margin of 24.6% in the first half of 2026, up from 18.9% in 2025. Gross margin is the share of revenue left after direct service costs, before other expenses. Epoch AI estimated an 85% margin for Anthropic's closed Opus 4.8 offering and 70% to 80% for DeepSeek, whose low serving costs make it an exception among open model developers.

Moonshot is trying to capture more value through its Kimi K3 licence, which requires some large providers selling access to enter revenue sharing agreements. It reportedly sought up to 30% of related service revenue from major cloud platforms. Z.ai's Bedrock arrangement also involves sharing revenue according to use. The precise commercial terms and the revenue these deals will generate have not been disclosed.

Cloud partnerships bring customers and dependence

Wang Peng, a researcher at the Beijing Academy of Social Sciences, describes the emerging arrangement as a division of responsibilities: model developers build the technology, global clouds supply infrastructure and corporate access, and developers receive income when customers use their models.

Fireworks AI illustrates the opportunity for infrastructure providers. Zhang says the California company, whose core management team is largely of Chinese descent, deploys Chinese open models for major overseas enterprises and has surpassed $1 billion in annualized revenue. Annualized revenue expresses the current revenue pace over a full year.

CSIS explains that third party hosting also reduces the burden on Chinese developers to supply computing capacity for a global customer base. However, prompts, feedback and usage records can stay with the host or customer. The original lab then has less access to information about how people use its model and where it needs improvement.

The commission's Two Loops report identifies another route for learning: deployment in factories, logistics networks and robotics. Those uses can generate specialized data that improves later systems. Chinese labs may gain feedback from industrial deployments and community adaptations even when direct access to overseas customer interactions is limited.

Shen Yang, a professor at Tsinghua University, describes different opportunities abroad: stronger commercial certainty in Southeast Asia, greater value per customer in Gulf markets and substantial unmet demand in Latin America. He also notes growing interest in AI sovereignty, meaning the ability to control important AI infrastructure, data and access within a country.

Security and customer trust shape adoption

Ecosia, the Berlin search company, is ending its partnership with France's Mistral AI and working with Melious to use models including Qwen, GLM and Kimi. Founder and chief executive Christian Kroll says the change halved operating costs while improving quality and performance. He cited Mistral performance problems, server outages and concerns about its energy supply and independence. Mistral cofounder Guillaume Lample invited Ecosia to test the forthcoming Large 4 model.

Kroll believes targeted adjustments can address political weaknesses in Chinese models. Rasmus Rothe of the German Federal Association for Artificial Intelligence is more cautious, arguing that retraining can remove direct censorship while subtler influences in training data remain difficult to change. A NewsGuard study also identified failures by Chinese systems to correct misinformation aligned with Beijing's interests.

Atlassian has taken a different approach. Chief executive Mike Cannon-Brookes says it uses Nvidia Nemotron, Mistral and other models, with an AI gateway team assigning tasks to suitable systems. He explained its production policy:

We don't use any Chinese models in production, we're very clear about that,

Cannon-Brookes says some customers resist Chinese models but accept open models. US lawmakers have also investigated Airbnb and Anysphere after disclosures involving Qwen and Kimi. Airbnb chief executive Brian Chesky said the company was not sending data to model developers.

The commission cites NIST findings that DeepSeek open models were more susceptible to cyber risks than comparable US models. Gu Zhaoqin, a partner at Duan & Duan Law Firm's Hongqiao branch, says overseas compliance challenges combine privacy and security concerns with trade protectionism driven by geopolitics. Local hosting gives customers control over processing location; model behaviour, security testing and legal compliance still require separate attention.

Cheaper AI puts infrastructure spending under scrutiny

Jefferies estimates that Microsoft, Alphabet, Amazon and Oracle have about $2.3 trillion in contracted backlog, roughly half associated with OpenAI and Anthropic. Backlog is business committed under contracts for delivery over time. Falling model prices create pressure on the revenue that AI developers expect to earn and use to purchase cloud capacity.

Jefferies cites OpenRouter figures showing Chinese models processed 76.12 trillion tokens in the week ending September 27, compared with 4.37 trillion in late April. The top four US providers processed 14.2 trillion that September week. Those figures describe a different period from OpenRouter's October rankings.

The investment bank questions spending plans of $600 billion to $800 billion on data centers in 2026 and a proposed $993 billion in capital expenditure for 2027. It also says the four major cloud companies' annualized earnings reached $581 billion through the second quarter of 2026, with nonoperating income accounting for all of the $222 billion increase. Investment gains and income from selling computing services are different parts of that earnings picture.

Bain & Company's 2026 Global Technology Report estimates that AI needs $6 trillion in annual revenue by 2031 to fund infrastructure for anticipated demand. It assigns $1.8 trillion to existing applications and $1.5 trillion to emerging ones, leaving a $2.7 trillion gap for new products and uses.

US companies retain major advantages in subscriptions, customer relationships and popular applications. The commission cites Microsoft's adoption estimates of 28.3% in the United States and 16.3% in China in the second half of 2025, with less robust coverage for China. Chinese models' expanding role inside other companies' products creates a different commercial opportunity: becoming widely used foundations while negotiating a larger share of the income earned above them.

Key Points

• DeepSeek led OpenRouter text requests at 24.5% in the week beginning October 5, 2026.
• Chinese models attract overseas customers through lower prices, practical performance and flexible deployment.
• Third party hosts capture substantial spending on models developed by Chinese labs.
• Amazon Bedrock partnerships and new licensing terms provide routes to revenue sharing.
• Security, censorship, privacy and political scrutiny affect enterprise purchasing decisions.
• Falling model prices put pressure on developer revenue and the economics of cloud infrastructure investment.