On a Monday in late February 2025, Alibaba's Hong Kong-listed shares had already climbed more than 68% since the start of the year, yet the company stepped to the microphone with a figure that reset the entire conversation.
The pledge: at least 380 billion yuan (≈US$52.4B / HK$408B) for cloud computing (云计算) and AI hardware infrastructure across the following three years.
The kicker is that this single commitment is larger than everything Alibaba spent on cloud and AI across the previous decade combined.
Key takeaways
- Alibaba announced a three-year infrastructure budget of at least 380 billion yuan (≈US$52.4B), the largest such pledge ever made by a Chinese private company in this category.
- The spending spans AI and cloud hardware, the foundational model platform, and AI-native applications layered on top of existing businesses.
- The move lands as Alibaba's open-source (开源) Qwen (通义千问) family becomes one of the most widely adopted model series outside the United States.
- Investors treated the shift as a strategic reboot, pushing the stock up sharply before the announcement even landed.
What Alibaba actually announced
The commitment was disclosed by CEO Eddie Wu (吴泳铭) on 24 February 2025. In his framing, the build-out is a "digital power grid" for the AI era: compute supply so abundant that downstream developers and enterprises stop treating tokens as a scarce resource. Wu tied the number directly to demand signals inside the business. Alibaba reported that AI-related product revenue had grown at triple-digit rates for six consecutive quarters heading into the plan, and that cloud revenue had returned to double-digit year-on-year growth.
The budget breaks into three lanes. First, raw infrastructure — data centers, servers, and the networking fabric that stitches them together. Second, the foundational model platform, where Qwen sits. Third, AI-native applications and the retrofitting of existing commerce, logistics, and enterprise services with model capabilities.
Why the number matters
Scale is the headline. Three hundred and eighty billion yuan is not a research grant; it is capital expenditure (capex), the kind of money that buys land, concrete, GPUs, and power contracts. For context, the sum dwarfs the annual R&D budget of most Chinese internet peers and signals that Alibaba intends to own the stack rather than rent it. Reuters, reporting the plan, noted that ByteDance (字节跳动) had separately earmarked over 150 billion yuan in capex for 2025, much of it AI-centered — useful context for just how aggressively the largest Chinese platforms are racing to secure compute.
There is also a competitive logic. Alibaba's cloud is already the largest in Asia by several measures, operating 87 availability zones across 29 regions. The new money is meant to widen that lead precisely as enterprise customers begin moving production workloads onto large models (大模型). Whoever supplies the cheapest, most reliable inference at scale captures the default layer of the next computing platform.
The open-source (开源) engine
The infrastructure bet is underwritten by a model strategy that is unusually generous. Qwen (通义千问) is released under open weights, and third-party trackers repeatedly place Qwen-derived models among the most downloaded on global repositories. This is not charity. Open releases pull developers into Alibaba's toolchain, where they eventually pay for hosting, fine-tuning, and enterprise support on Alibaba Cloud.
The framing matters for readers outside China. Qwen is one of the few non-US model families that a European or Southeast Asian startup can download, audit, and deploy without a US cloud account. That portability is exactly what turns a domestic infrastructure spend into an export channel.
Risks and the spending-discipline question
A three-year, 380-billion-yuan commitment is also a three-year, 380-billion-yuan liability if demand softens. Alibaba has been candid that the original figure looks conservative and could be raised if supply stays tight — a rare case of a company warning that it might spend even more. The risk is not the spending itself but the depreciation curve: GPU fleets lose value fast, and a mis-timed procurement wave can weigh on margins for years.
There is also the demand-side puzzle. Chinese enterprises have been slower than their US counterparts to pay per-seat for AI software. Alibaba is betting that inference volume, not subscription revenue, is the nearer-term monetization path — charge for the compute, not the chatbot.
Honest limitations
This article relies on Alibaba's own disclosure and Reuters reporting of the February 2025 announcement. The 380-billion-yuan figure is a stated intention, not audited spend; actual outlays will surface only in quarterly capex line items over the coming years. We have not independently verified how much of the budget is already committed versus planned. The triple-digit AI revenue growth figure comes from the company's earnings commentary and is not broken out in a separately audited segment. Overseas market-share claims reflect Alibaba's own geographic counts rather than an independent benchmark.
What readers can do now
- Track Alibaba's quarterly capex disclosures — if the run-rate approaches roughly 30 billion yuan per quarter, the 380-billion plan is on track and signals sustained GPU demand across the sector.
- Developers should benchmark Qwen variants against closed US models on their own workloads; open weights mean you can run the comparison locally before committing to a cloud contract.
- Enterprise buyers can use Alibaba's stated scale as a negotiating lever when pricing inference contracts elsewhere — abundant capacity should pressure margins and prices across Chinese cloud providers.