Alibaba Reframes AI as Manufacturing: 100% AI-Related Product Revenue Growth, Negative Free Cash Flow
The most instructive number in Alibaba's March-quarter release is the one that contradicts the headline. Cloud Intelligence Group's external revenue reportedly accelerated to 40% growth. Free cash flow, over the same three months, swung by roughly twenty-one billion yuan in the wrong direction—from a RMB3.7 billion inflow a year earlier to a RMB17.3 billion outflow.
Two numbers, same page, pointing opposite ways. The market shrugged: the stock reportedly fell 3.22% on the day and then settled, unable to render a verdict on a company that is simultaneously accelerating and bleeding.
AI is more like a manufacturing industry. To obtain more revenue, two core factories must be established—an AI training factory and an AI inference factory.
Into that gap, CEO Eddie Wu dropped one sentence meant to dissolve the contradiction: "AI is more like a manufacturing industry. To obtain more revenue, two core factories must be established—an AI training factory and an AI inference factory. Now, almost no card in our servers is empty."
Read that as marketing and you miss what it is actually doing. Read it as an accounting argument and it becomes the most important thing said on the call.
Last week we tracked Alibaba's ¥30B ARR target as the test of whether tokens behave like a manufactured good. This week the same thesis prints its cost: ARR has reportedly climbed to a RMB35.8 billion run-rate (per Alibaba, US$5.3 billion) on AI-related product revenue, and free cash flow has gone negative. The capex-intensity thesis is no longer hypothetical. It is on the cash-flow statement.
Here is the regime change hiding inside Wu's metaphor. Alibaba booked "almost all AI R&D and computing power infrastructure investments" into current expenses. Adjusted EBITA reportedly fell 84% to RMB5,102 million. The company reportedly posted a RMB848 million operating loss—its first since 2021.
A pure internet burn-for-scale play would have capitalized more of that spend and asked investors to fund the gap as a growth story. Alibaba is doing the opposite. It is running the buildout through the income statement the way a factory expenses its production line, and calling the result capacity, not loss.
That reclassification is the whole bet. Speculative bets get valued on faith and narrative. Factory capacity gets valued on unit economics—yield per unit of installed base, margin per unit of output. By invoking manufacturing, Wu invites the market to stop asking "will this pay off?" and start asking "what's the yield curve?"
It is a smart trade, because a manufacturing frame comes with a scoring rubric that an internet frame does not. So apply it. The test manufacturing invites is simple: does installed capacity map to bookable output at a price that covers the depreciation?
On the output side, the numbers hold up better than the cash flow suggests. Token consumption on Model Studio reportedly grew more than 10x between November 2025 and May 2026. Model and application services ARR has reportedly crossed RMB8 billion and is on track past RMB10 billion this quarter, against a stated RMB30 billion target by the end of calendar 2026.
That is the yield curve Wu wants you to look at. Cards are full, tokens are flowing, and the run-rate is compounding fast enough to make the RMB30 billion target look like arithmetic rather than aspiration.
But full cards prove demand, not margin. "Almost no card is empty" tells you utilization is high. It tells you nothing about whether the token those cards produce sells above cost. A factory that runs at 100% utilization and prices its output below its all-in cost is not a factory—it is depreciation with extra steps.
This is where the manufacturing frame stops flattering the story and starts stress-testing it. The RMB520.8 billion cash reserve—reportedly about $38 billion net—is real, and CFO Toby Xu reportedly says it funds this intensity for two more years. But a cash cushion is runway, not return. It buys time for the yield curve to price correctly; it does not guarantee that it will.
So the whole thing reduces to a pricing question, and this is where Alibaba's situation genuinely differs from a US hyperscaler's—the one-line contrast, and no more than that: Microsoft or Amazon underwrite their capex against dollar-denominated cloud pricing and third-party silicon. Alibaba can underwrite differently, because it holds two levers a hyperscaler renting Nvidia does not.
The first lever is mix. MaaS—revenue from API calls to models on the Bailian platform—carries structurally higher gross margin than renting raw IaaS compute. The bulk of Alibaba's AI product revenue already comes from Bailian API services. As MaaS rises as a share of cloud revenue, blended margin rises with it, mechanically, without a single new card.
This is the difference between selling factory floor space and selling finished goods. IaaS is floor space. MaaS is the manufactured token, priced at a markup. The 10x token surge matters far more if it is arriving through the high-margin door than the low-margin one—and per the call, most of it is.
The second lever is the one no other Chinese cloud has. Pingtouge's self-designed AI chips are reportedly in scale production, with reportedly more than 60% serving external commercial customers across internet, financial services, and autonomous driving. Alibaba calls itself "China's only cloud provider that can supply self-designed AI chips at scale."
Vertical integration is what turns a utilization story into a returns story. If the silicon underneath the inference factory is your own design rather than a purchased margin stack, the cost side of every token bends in your favor—and the depreciation you expensed this quarter sits on assets you built cheaper than a renter can.
Stack the two levers together and Wu's confidence about "very certain" returns over three to five years becomes legible. Rising MaaS mix lifts the price of output; self-designed silicon lowers the cost of capacity; token volume is already there. That is a credible path from full cards to positive unit economics.
But legible is not proven. The same release shows what happens when a "factory" fills without pricing: quick commerce reportedly grew revenue 57% to RMB20 billion and remains loss-making, with positive unit economics only targeted for end of fiscal 2027. Utilization without margin is exactly the failure mode the manufacturing frame is supposed to guard against—and Alibaba is living it in a different segment right now.
The AI factory is not quick commerce. But the cautionary structure is identical: output can scale beautifully and still lose money on every unit if the price sits below the all-in cost. High utilization is a necessary condition for the manufacturing thesis. It is nowhere near a sufficient one.
Which is why the numbers that matter next quarter are not the capex headline. The capex will be large; Wu has reportedly already signaled it may exceed the announced RMB380 billion. Watching capex rise tells you the factory is being built. It tells you nothing about whether it prints money.
The numbers to watch are two: MaaS as a share of cloud revenue, and whether the ARR line hits the RMB10 billion this-quarter checkpoint on the way to RMB30 billion. The first tells you the margin lever is engaging. The second tells you the yield curve is holding.
That RMB30 billion figure is now a hard, dated commitment—measurable at the very next earnings call. It is the number that closes the token-as-manufactured-good thesis we opened on 07/14. Hit it, and reclassifying compute as capacity was analytically sound. Miss it, and the manufacturing frame was a way of expensing a bet before knowing whether it would pay.
The contradiction on the earnings page, then, is not really a contradiction. It is a wager with a due date. Alibaba has chosen to book its AI buildout as factory cost rather than growth burn, and it has staked the honesty of that choice on a specific run-rate arriving on schedule.
A factory that fills is impressive. A factory that fills and prices its output is a business. The next print tells us which one Alibaba has built.