A row of prefabricated data halls sits on the edge of a city that built its fortune on storefronts and logistics. Inside, the racks are being filled not with merchandise but with silicon built to think.
The company behind those halls has made a wager that is quietly reshaping how the country's largest internet groups spend money. What was once a back-office line item has moved to the very center of the strategy.
It is a bet on scarcity. The constraint is no longer ideas or models but the physical capacity to run them — the servers, the power, and the compute (算力) that turns a trained system into a product that customers will pay to use.
In February 2025, Alibaba said it would put at least 380 billion yuan into AI and cloud infrastructure over three years, a commitment the company framed as exceeding everything it had spent on those areas across the previous decade (Reuters, Feb 24 2025, citing the company's statement). At the fixed rate used here of 7.1 yuan to the US dollar, that is roughly US$53.5 billion. It is the largest single infrastructure pledge any Chinese internet company has made in the current wave of the large model (大模型) build-out.
Key takeaways
- Alibaba pledged at least 380 billion yuan (US$53.5 billion) to AI and cloud infrastructure over three years, per its Feb 2025 statement. (Reuters)
- The sum exceeds the company's total AI and cloud spending across the previous ten years. (Reuters)
- Its core e-commerce business generates about US$25 billion in free cash flow each year, funding the build-out. (Alibaba Group)
- Rivals are spending too: ByteDance earmarked over 150 billion yuan for 2025 capex, much of it AI. (Reuters)
- Heavy capital spending is pressuring free cash flow even as AI cloud revenue climbs.
Why a retailer is building power plants
The logic is deceptively simple. A trained system is useless until it runs somewhere, and running it at scale requires data centers, electricity, cooling, and accelerators. Alibaba's argument is that owning that foundation is what lets it sell the layers above it — model hosting, inference, and enterprise cloud services — to other companies.
This is why the company now describes its cloud and chip work as a "full-stack" strategy spanning in-house accelerators, infrastructure, models, and applications. The phrase sounds abstract, but the spending is concrete: building capacity ahead of demand, on the belief that customers will arrive once the racks are live. The wager treats infrastructure not as a cost center but as the on-ramp for every downstream product the firm wants to sell.
There is a second motive that rarely appears in the press releases. Whoever controls enough compute (算力) can set the terms for the developers who build on top of it. In a market where the underlying models are increasingly free or open, the scarce asset may be the metal that runs them, not the software that sits on top.
The math of a three-year sprint
Funding such a plan is not trivial, but Alibaba points to a durable engine. According to the company, its core e-commerce business generates about US$25 billion in free cash flow each year, a cushion that most loss-making AI startups do not have (Alibaba Group, remarks at VivaTech 2026). That cash funds a build-out measured in hundreds of billions of yuan without relying solely on outside financing.
The competitive context matters. Reuters reported, citing people familiar with the matter, that ByteDance earmarked more than 150 billion yuan (about US$21.1 billion) in capital expenditure for 2025, much of it centered on AI. When the largest platforms all race to build the same scarce capacity at once, the bill for equipment and power rises for everyone, and the first mover locks in supply that later entrants must chase at higher cost.
Alibaba is not alone among the giants. Meituan's chief executive said on a post-earnings call in March 2025 that the company was pouring "billions" into chips to train AI models (Reuters, Mar 21 2025). The pattern across Chinese internet leaders is consistent: a shift from light, software-led growth to heavy, asset-led investment in capacity.
Where the money actually goes
The pledge is a commitment, not a contracted schedule, but the direction is clear. The capital is being directed at a few places:
- Data centers and the land, power, and cooling required to keep them running around the clock.
- Server stacks built around both purchased and in-house accelerators, as the firm pushes its own chip designs.
- The software layer that turns raw compute (算力) into sellable cloud services rather than idle hardware.
- Long-term power agreements that secure the electricity a multi-gigawatt build-out will eventually demand.
Each of these carries a different payback clock. Real estate and power are slow and lumpy; the software layer is where margins can improve as products mature and utilization rises.
What the spending buys — and what it risks
By 2026 the company's quarterly capital spending had risen to tens of billions of yuan, weighing on free cash flow even as its AI cloud and compute services revenue climbed sharply year on year. Management has said AI-related capital expenditure could pay back within roughly three years at current margin levels, but that assumes steady enterprise adoption that has not yet been proven at this scale.
The contrarian reading is that this is a heavy-asset business dressed in software language. Compute is physical, depreciates, and must be filled constantly. If demand softens or cheaper capacity floods the market, the same infrastructure that looks strategic today can become a drag on earnings tomorrow. The stock's strong run in early 2025, up more than 68% at one point according to Reuters, shows investors initially welcomed the story — but a higher share price does not itself validate the returns.
The more durable question is unit economics. A data hall only earns when its racks are utilized, and utilization depends on enterprises trusting a local provider with mission-critical workloads. That trust is built over years, not quarters, which is exactly the horizon Alibaba has now committed to.
Honest limitations
The 380 billion yuan figure is a company commitment, not contracted spend, and the pace can slip or shift with demand and supply. Reported quarterly capital spending has run at tens of billions of yuan, but exact figures vary by source and quarter, so they are presented qualitatively here. Returns depend on enterprise uptake of AI cloud services, which remains early and unproven at this scale. This article describes a financing and strategy decision, not an investment recommendation.
What readers can do now
- Investors: track Alibaba's quarterly capex and cloud revenue growth as the cleanest signals of whether the plan is executing.
- Operators: benchmark your own compute (算力) roadmap against the multi-year capacity large platforms are adding.
- Researchers: watch whether self-built infrastructure actually lowers unit costs versus renting from incumbents over time.