A chip roadmap used to be a slide shareholders skimmed past. This one came attached to a confession: the model that answers your questions is now learning on silicon the company designed itself. Alibaba's latest earnings call had the shape of an infrastructure story wearing a financial-report cover.
On November 19, Group CEO Eddie Wu told investors that the next generation of Qwen large models (大模型) is being trained on processors from T-Head (平头哥), Alibaba's wholly owned chip-design subsidiary. He also set out a three-generation accelerator plan and said AI-related product revenue had reached RMB 17.92 billion (≈ US$2.67B / HK$20.8B) for the quarter.
Key takeaways
- Alibaba says next-generation Qwen training is running on in-house T-Head chips, and that its Bailian platform already offers inference on Chinese-made hardware.
- The roadmap has three steps: Zhenwu M890 fully launched this year, V900 and J900 moving toward mass production in 2027 and 2028. A new training-and-inference chip, the V900, is slated for volume sales in the first quarter of 2027.
- AI-related product revenue hit RMB 17.92 billion (≈ US$2.67B / HK$20.8B), the 13th consecutive quarter of triple-digit year-on-year growth. AI cloud and compute revenue rose 53 percent year on year.
- Adjusted EBITDA for AI cloud and computing services more than doubled to over RMB 8.8 billion (≈ US$1.3B / HK$10.1B); quarterly capital expenditure reached nearly RMB 50.5 billion (≈ US$7.6B / HK$59B), up 60%.
- The chip unit is also heading for a separate Hong Kong listing, according to reports — a test of whether public markets will price a Chinese AI accelerator on its own.
The roadmap, without the jargon
Three names matter here. The Zhenwu M890 is the chip Alibaba says is fully launched this year. The V900 is the one everyone is watching: a training-and-inference accelerator that T-Head unveiled alongside the earnings call, with mass production and sales targeted for the first quarter of 2027. The J900 is the longer horizon item, with its supply chain advancing toward large-scale production in 2028.
What makes the sequence interesting is not the cadence but the instruction underneath it. Wu's argument is that domestic computing power (国产算力) is becoming the support base for the group's full-stack AI technology — carrying training, inference and agent-style workloads for both Alibaba and its cloud customers. Once a company controls that layer, the pace of model iteration becomes "more controllable," in his phrasing. That is engineer-speak for: fewer surprises when someone else's supply chain wobbles.
The Bailian detail is easy to miss and worth noticing. It means the arrangement is not only internal. Outside developers calling Alibaba's model service may already be running on Chinese-made silicon without having chosen it explicitly. For them, the hardware decision has been moved upstream, into a platform they do not control.
Why training on your own chips is a different claim
Buying chips and using chips are the same story from the outside. Training a frontier model on them is not. Training runs are long, brittle and unforgiving of software gaps; a mismatch between a framework and an accelerator can waste weeks. When a hyperscaler says its flagship model is being trained in-house, it is claiming that the entire stack — compilers, kernels, networking, scheduling — now holds together under real load.
That is a heavier claim than shipping an inference chip, where the workload is shorter and more predictable. It also explains the capex number. Nearly RMB 50.5 billion (≈ US$7.6B / HK$59B) in a single quarter, up 60% year on year, is what capacity expansion and rising component prices look like. Alibaba says the money is going primarily into AI infrastructure, with demand also coming from electric vehicles, embodied intelligence and biotechnology.
There is a second layer to the story. In November, Alibaba filed to spin T-Head off for a separate Hong Kong listing, with reports putting a potential raise in the range of US$2–3 billion (≈ HK$15.6–23.4B). The business had previously been folded into Alibaba Cloud, which made it hard to value on its own. A standalone listing would let public investors price the accelerator bet directly — and would put a disclosed number next to a roadmap that has so far been mostly narrative.
What the numbers do and do not say
The revenue figures are strong but need careful reading. AI-related product revenue has grown triple digits for 13 straight quarters, and AI cloud and compute services rose 53% year on year. The faster jump in adjusted EBITDA — more than doubled to over RMB 8.8 billion (≈ US$1.3B / HK$10.1B) — suggests operating leverage is beginning to appear in that segment.
But the same call reaffirmed a target of US$100 billion in external cloud revenue with a margin above 20% by fiscal 2031. That is a distant goal, not a result. Wu also said capital expenditure would stay high while capacity expands through 2027 and 2028, and CFO Toby Xu flagged a focus on return on capital and positive free cash flow "as early as possible." Read together, those lines describe a company spending heavily now and promising discipline later. Both statements can be true; only one is measurable today.
Honest limitations
This article relies on Alibaba's own fiscal second-quarter earnings call as reported by Global Times and several English-language financial outlets, plus reporting on the T-Head listing filing. All of the strategic claims — that Qwen is being trained on in-house chips, that the Bailian platform serves Chinese-made inference, and the roadmap dates — come from company statements. I have not seen an independent technical verification of the V900's performance or a published training log.
The RMB-to-US and RMB-to-HK conversions are approximate and follow the rates implied in the cited reports; they will drift with markets. The V900 volume date and the 2028 J900 date are plans, not shipped products, and plans in this industry slip. The reported IPO raise is an analyst estimate, not a confirmed deal size. I have also not verified the exact definition of "AI-related product revenue," which groups several business lines and is not the same as cloud AI revenue alone.
What readers can do now
- If you build on Alibaba Cloud, check which hardware backs the model endpoint you use. The answer may affect latency, data residency and price in ways the API docs do not spell out.
- If you follow chip stocks, separate a roadmap from a shipment. A production date announced on an earnings call is a forecast; a customer contract is evidence.
- If you are just watching, watch the listing. A public filing would finally put auditable numbers next to a business that has lived inside a cloud division.
The quiet news in this call was not the revenue line. It was the sentence about where the next model is being trained. Hardware decisions made in a data centre today shape what the rest of us can afford to use tomorrow.
