China's AI Stocks Are No Longer Cheap — the Re-Rating Is Real, but the Revenue Is Rented Compute

The cheap-China-AI trade ended in stages this summer, and the stock charts started moving before the bears could restate the story. Alibaba's Hong Kong shares climbed 36% off their June 26 low by mid-August, and the market's reasoning was stated in plain terms: the valuation anchor moved from e-commerce profits and cash to the growth potential of Alibaba Cloud's AI business. For anyone who had spent 2026 treating Chinese AI names as the last discounts in a frothy global market, that is the assumption that just broke.

The problem is that "cheap" was always a statement about a denominator, and the re-rating changed only one of them. The revenue is real, accelerating, and now big enough to shift a company's value. It is also, to a degree investors can verify from the filings, rented compute. The income statement is paying for it. That gap — between the growth the market now prices and the profit the companies now report — is what decides whether the re-rating holds.

The numbers that ended the cheap era

Start with Alibaba. In the quarter ended March 2026, cloud revenue grew 38% year over year, with AI-related products at 30% of cloud's external revenue — RMB 8.97 billion (about $1.3 billion) in a single quarter, after eleven consecutive quarters of triple-digit growth. Management has guided that AI-related products will cross 50% of cloud external revenue within about a year. In the June quarter the pattern held: total revenue grew 9% — Alibaba's fastest rate in roughly three years — with AI-related services up 45%.

Baidu shows the same shape from a smaller base. Its "AI-powered business" produced RMB 12.5 billion in the June quarter — 50% of Baidu Core revenue, up from 38% a year earlier — and within it, GPU cloud revenue rose 283% year over year.

The market's case for a forever-valuation change is built on exactly these numbers, and the numbers are not cooked. The question is what the revenue is made of.

Decompose the 50%

Baidu's own disclosure answers that. The AI-powered business is three pieces: AI Cloud Infra, AI applications, and AI-native marketing. In the June quarter, AI applications grew 3%. AI-native marketing was flat. AI Cloud Infrastructure grew 50%, and the single explosive line was GPU cloud, up 283% — which Baidu itself previously described as "subscription-based revenue from AI accelerator infrastructure." That is rented chips. The headline growth rate is a compute rental business doing exactly what compute rental does in an AI capex boom.

Alibaba's disclosure is vaguer — the company has not broken out its AI revenue between compute, model platform, and software. But the economics leak through the aggregate: Alibaba Cloud's adjusted EBITA margin was 9.1% in the March quarter, against roughly 38% at AWS and 35.6% at Google Cloud. A revenue line growing at triple digits inside a segment earning a quarter of what the U.S. hyperscalers earn is not high-margin software; it is infrastructure sold in a market that just spent two years in a token price war.

The bill for the revenue

Here is where the re-rating and the income statement collide.

Alibaba's June-quarter net income fell 75% to about RMB 10.4 billion. Its capital expenditure rose 75% year over year to RMB 67.68 billion in a single quarter — the price of buying the GPUs whose rental generates the celebrated revenue line. On trailing-twelve-month numbers, free cash flow has turned negative. Baidu spent about RMB 11.4 billion on capex in the quarter, roughly triple the year-ago level, while total revenue declined 4% — a fifth straight quarterly drop — because its advertising business shrank faster than AI could grow.

Then came the capital structure signal. On August 23, Alibaba priced a placement of 710 million new Hong Kong shares at HK$112.70, raising HK$80 billion (about $10.2 billion) — the largest primary follow-on offering ever by a Hong Kong-listed company, with all net proceeds earmarked for AI infrastructure. For comparison, the main U.S.-traded China internet ETF has seen net redemptions year to date. The re-rating was a repricing, not a wall of new retail ETF money — and the company is now asking the equity market to fund the growth it just got re-rated for.

Cheap on the old denominator, priced on the new one

So who is right — the June 2026 headline that China still offered cheap AI stocks, or the claim that the discount is gone? Both, on different denominators. Alibaba trades around 19x expected forward earnings against about 26x trailing, still a fraction of the multiples the big U.S. cloud names carry. It is cheap on the earnings the old e-commerce business used to produce, and nowhere near cheap on the revenue the new AI business is producing. The 26x trailing number itself is a warning label, not a bargain marker: it rose because GAAP earnings fell off a cliff, not because the stock ran away.

Strip away the cheap-vs-expensive framing and the real question is what a yuan of Chinese AI revenue is worth. The market now prices that revenue as growth. The income statement prices it as capital-intensive and margin-thin. The two readings converge only if unit economics improve — and that is now a testable question with live observables.

The first observable is Alibaba's guidance, taken literally: when AI-related products cross 50% of cloud external revenue, does the cloud segment's margin climb with the mix, or fall as more of the revenue becomes compute? The second is the price of tokens. DeepSeek, the symbol of the assumption that Chinese models would get cheaper forever, is planning significant API price increases. If Chinese inference finally stops getting cheaper, that AI revenue becomes more profitable; if the increases don't stick, the re-rating priced a revenue line that cannot earn its multiple.

The "done being cheap" claim is already true on the price chart, and the revenue numbers justify the market's new attention. The unsupported half is the second step: assuming the revenue is worth what it is now priced at. You do not need a view on whether China wins AI to decide how to treat this re-rating. You need to watch whether the revenue earns more than it costs to produce — and that answer arrives in the quarterly cloud margins, the AI mix crossing 50%, and what DeepSeek charges for a token next quarter.