Alibaba Gave the Model Away. Now It Has to Pay for the War.

Alibaba Gave the Model Away. Now It Has to Pay for the War.

This week Alibaba proposed one of the largest share placements in Hong Kong this year: about $10.2 billion of newly issued stock, with the deal routed through its Hong Kong listing and aimed mostly at non-US investors, and earmarked entirely for AI — the Qwen model family and the infrastructure under it. In the coverage, the verb was "entering" — Alibaba, belatedly, buying its way into the AI race. That framing has a problem. You do not usually sell ten billion dollars of new equity to enter a race you are already, by several honest measures, leading.

Alibaba's models are the most-downloaded open AI family around; its newest flagship, with 2.4 trillion parameters, is billed as competing with Anthropic's best; Apple is putting Qwen inside the devices it sells in China; and the cloud business that runs them grew 45% last quarter and is China's largest AI cloud. The company's own quarterly report claims its AI-product revenue has posted triple-digit growth for twelve consecutive quarters. Now hold that picture next to the same report's bottom line: net income down roughly 75% on about 9% revenue growth. The conventional read is that Alibaba overpaid, or got caught a step late. I think the leading position and the profit collapse are the same fact, not two facts that contradict each other. The $10.2 billion is not entry. It is the price of escalation, and it tells you where the scarce resource in AI now lives.

For the whole history of software, the scarce input was the idea. The code itself scaled almost free — the marginal cost of a copy was zero, and the moat was building something people wanted that was hard to replicate. AI inverts that. The model is the part anyone can copy, and Alibaba proved it by publishing its flagship weights for anyone to download. What cannot be copied cheaply is the machine underneath: the chips, the data-center shell, the power contract. So the scarce resource has shifted from insight to scale, and the accounting arrived on schedule: revenue up 9%, capital spending up 75% in a single quarter to roughly $10 billion, with the difference showing up exactly where you would expect it — as depreciation eating the profit line. That is the whole story of the quarter in three numbers. The company is spending close to $40 billion a year at that rate, and the more it spends on machines it then must depreciate over several years, the more of today's growth gets consumed by yesterday's purchases.

The obvious question is why a company sitting on about $70 billion of cash and liquid investments sells another $10.2 billion of stock. The answer is that it is not the number of dollars that is scarce; it is the kind. Chips, servers, and power contracts are bought in dollars, offshore. A large share of Alibaba's cash sits in onshore currency, which under China's capital controls does not move freely to pay for foreign hardware. Issuing equity through Hong Kong converts domestic value into offshore purchasing power, and aiming the deal at non-US investors keeps it out of a US capital market that has become geopolitically awkward as Washington and Beijing pull apart. Alibaba went to the same well once before, for about the same money, in its 2019 Hong Kong secondary listing. So part of this raise is pure balance-sheet plumbing. But the bet is real. Management lists higher chip component prices among the reasons spending ballooned, which is an honest way of saying the input it needs is getting more expensive while it is running to buy it. When the price of your main input is rising, no amount of cash feels like enough.

Here is the part I cannot stop turning over. Alibaba reached the frontier the cheap way: it gave the frontier away. Open weights meant anyone could take Qwen, download it, build on it — and that generosity is precisely what produced the downloads, the ecosystem, the Apple deal. It was the right strategic move at the category's most crowded moment. But generosity carries a hidden tax. When the model is a commodity everyone can run, the model stops being the moat. The advantage migrates to whoever can run it cheapest at the greatest scale, which is another way of saying whoever can spend the most. So the very strategy that made Qwen the world's most-shared model family is the strategy that made Alibaba's industry a pure capital contest — and then handed Alibaba the bill. You can read the entire $10.2 billion as that bill arriving.

The escalation is visible in the installments that came before it. Alibaba fed this fire for a while in small checks — a record $4.5 billion of convertible bonds in 2024, when a convertible is debt that quietly turns into new shares, and another $3.2 billion of them last September. Now it has stopped pretending a war this expensive can be funded in installments and gone straight to equity investors for the whole $10.2 billion at once. In the same quarter it did that, it bought back $162 million of stock — a rounding error next to the $25 billion buyback plan it ran as recently as 2022. Management reversed its view of its own shares in a single accounting period. The people closest to the business are now telling you, in the only language that costs them, that a GPU is a better investment than a share of Alibaba.

The market's maps have not caught up to what this company has become. The standard industry classification still files Alibaba under consumer retail — Broadline Retail, in the GICS taxonomy — which is how a business spending toward $40 billion a year on compute infrastructure ends up categorized alongside department stores. AInvest's aggregate signal labels the stock a Buy, and on the day of the blowup the shares wobbled through a 7% intraday swing before closing slightly higher: a market treating a 75% collapse in profit as the price of admission to a war it has decided is worth fighting. I do not think the crowd is wrong to believe in the war. I suspect it has no idea what winning will cost, and I suspect Alibaba is mostly guessing too, which is what raises of this size are for — buying the right to keep learning.

The strongest objection to all of this is the "circular AI" worry now circulating through markets. The concern is precise: it describes arrangements where a chip vendor or cloud provider invests in, lends to, or takes equity in a customer, who then spends the proceeds on the vendor's own product — demand that looks diversified and durable but is really a loop, ready to unwind the moment the funding stops. Against Alibaba the charge misses. It is spending its own money on its own cloud, and the demand behind it — external cloud revenue up 45%, AI product revenue compounding — reads as real. But there is a second-order version of the concern that lands. The price of the chips Alibaba is buying is being inflated by exactly the loop-dealing happening elsewhere in the trade. You can run a clean operation and still pay the circularity tax, because the tax is built into the cost of your main input. That is the sense in which Alibaba's spending spree and the bubble talk belong to the same story.

Notice, too, how fast the official account of Chinese AI spending turned over. As recently as December, a UBS analyst argued the Chinese giants were keeping their budgets disciplined enough that there would be no AI bubble and no circular financing in China. Less than nine months later, a single Chinese company is moving nearly $10 billion a quarter, and the world's discussion is about whether that is a floor or a ceiling. The "China is being disciplined" thesis did not survive contact with the actual race, which is worth remembering whenever the next calm-sounding pronouncement about this industry arrives.

So here is the scoreboard I would use, because the ambiguity is the only honest thing in the room. This is not a question of whether Alibaba is in the AI race. The question is whether each marginal billion of capex buys a marginal billion of durable cloud revenue after depreciation — that is, after the machines purchased today stop being free to the income statement. Watch two things. First, whether the near-$10-billion quarter is a one-off step-up or becomes a new run rate; the step-up absorbs badly, the run rate transforms the business. Second, whether cloud's 45% growth accelerates enough to outrun the depreciation now flooding the cost base. If the answers are yes, this raise will turn out to be the cheapest thing Alibaba ever did. If they are not, it is a check written at the top of a spending cycle by a management that decided it could not afford to wait. The deeper lesson applies at every AI company right now: when the scarce resource in a field switches from ideas to capital, the game changes faster than the maps do, and the people who notice the switch while everyone else still believes the moat is the model are the ones standing upwind. Alibaba noticed it the expensive way. Whether it noticed it soon enough is the whole question — and it is the question every shareholder in this trade is now paying to answer.