The distinction matters because the open-model market can grow while its model developers remain poor businesses. Chinese systems accounted for 41 per cent of downloads on major platforms by late 2025, while their estimated global market share rose from about 1 per cent to 15 per cent in a year.[1, 3, 4] Open models also narrowed one benchmark gap with proprietary systems to 0.3 per cent during 2025, although proprietary laboratories have since extended their lead on some measures.[5, 6] These figures establish technical credibility and distribution. They do not identify who earns the margin.
Open access lowers the cost of trying a model and makes substitution easier. The 15-person Korean start-up Univa said using Qwen cut its costs by 30 per cent, while Indian companies have turned to models from DeepSeek, Alibaba and Moonshot AI to contain AI spending.[7, 8] Each new evaluation, download or local deployment can create demand for inference, storage, security and enterprise support. Yet the model may run on infrastructure owned by someone else: Amazon, Microsoft, Google, Oracle or Alibaba can host open systems regardless of which laboratory built them.[9] Inference spending therefore tends to reward the operator that can keep expensive hardware busiest and deliver the workload most efficiently, not automatically the creator of the underlying model.[9, 10]
Alibaba’s integrated cloud model captures the value that open access disperses
Alibaba has built the commercial layer needed to capture that spending. Its cloud external revenue grew 40 per cent year over year; AI-related products produced 30 per cent of that revenue and recorded triple-digit growth for an 11th consecutive quarter.[11] Annualized AI-product revenue exceeded Rmb35.8bn, and the customer base of Model Studio, its managed model platform, increased eightfold.[11] The later segment disclosure was stronger still: Cloud Intelligence revenue reached Rmb41.63bn, while adjusted EBITA rose 57 per cent.[2] Adoption has therefore moved beyond downloads into reported revenue and segment earnings, even though Alibaba has not disclosed how much of either came specifically from Qwen.
That is the economic advantage of being both model supplier and infrastructure owner. Qwen can be free or cheap enough to attract developers, while Alibaba charges for the computing environment in which those developers build and operate applications. Red Hat established the same commercial principle with Linux: freely available software widened adoption, but support, training, consulting and enterprise features formed the paid layer.[12] The analogy supports Alibaba’s structure, not a comparable financial outcome for every Chinese AI company.
Independent laboratories face the opposite arithmetic. Z.ai increased revenue 132 per cent to Rmb724mn last year, but its net loss widened 60 per cent to Rmb4.7bn.[13] Its open-API business has reached Rmb1.7bn in annualized recurring revenue, yet that run rate is not the same as booked full-year sales and the company remains lossmaking.[14] Open-model competition accelerates commercial use while compressing prices across the system, leaving laboratories with rising demand but uncertain long-term profitability.[9] Z.ai may become a large supplier. Its accounts do not yet show that scale producing durable returns.
Overseas distribution widens reach without yet revealing durable returns
The same caution applies to China’s overseas expansion. Huawei has created a regional cloud network and partnerships with STC, e& enterprise, du, Zain, Omantel and Ooredoo; its Riyadh hub says it has added more than 1,000 customers in two years.[15] Alibaba entered Saudi Arabia through a venture with STC, and the local operation says it has adapted more than 100 Alibaba Cloud products and continues to update the service stack.[15, 16] These are concrete distribution channels in a market where government agencies must consider cloud services before buying equipment and providers have used local partnerships to meet data requirements.[15] They show that Chinese companies can place models close to customers. Customer counts and localized catalogs, however, do not reveal regional revenue, retention, workload growth or operating profit.
The capital burden sharpens the divide between integrated platforms and stand-alone labs. China remains years from producing frontier AI chips at scale through a wholly domestic supply chain, constraining the compute available to model developers and cloud operators alike.[17] Alibaba expects to exceed its earlier three-year capital expenditure target of Rmb380bn, after spending Rmb27bn in the March quarter.[18] Its established businesses can fund that build-out, and its cloud segment has already shown rising earnings. Smaller laboratories must finance model development while competing against systems whose open availability keeps prices under pressure.
This does not leave U.S. laboratories untouched. Capable, lower-priced open models can force OpenAI, Anthropic and others to reduce prices and make their training costs harder to recover.[9] A temporary restriction on Anthropic’s Fable and Mythos models gave Z.ai an immediate opening, and Zhipu’s shares rose 33 per cent after the curbs.[13, 19] That was a market reaction, not evidence of lasting customer or revenue losses at U.S. providers. Meta’s former dominance of open-weight AI is more directly challenged: Chinese models had captured 41 per cent of platform downloads by late 2025, while Qwen passed 700mn downloads by January 2026.[1, 3] Even here, ecosystem influence has moved faster than demonstrated financial damage.
The paid layer will determine whether ecosystem influence becomes durable profit
Over the next reporting cycle, the decisive figures will come from the paid layer. Alibaba would strengthen the case if AI-related products rise above half of cloud external revenue while adjusted EBITA continues to grow and AI-product sales remain at triple-digit rates.[2, 11] Named Middle Eastern production contracts, workload data or regional profit disclosures would show whether overseas distribution is economically material rather than strategically useful.[15, 16] A narrower loss or positive operating cash flow at Z.ai would show that an independent laboratory can retain some of the value its models create.[13] Until then, China’s open-model expansion is already a commercial success for Alibaba’s cloud infrastructure; the unresolved pressure sits with the laboratories whose most valuable workloads can still enrich the hosts instead.