---
title: "Alibaba's CEO Says Most Model Training Will Move Onto Domestic Chips"
date: 2026-11-21
category: Chips & Compute
site: NeuroAI
canonical: https://neuroai.site/a/na-chips-pingtouge-domestic-training
language: en
---

# Alibaba's CEO Says Most Model Training Will Move Onto Domestic Chips

> Alibaba chief executive Eddie Wu told investors that a new generation of Pingtouge chips will support large-scale model training by early next year, and that most training will eventually run on domestic silicon.

The question on the earnings call was about supply chains and export rules. The answer was about a wager: that within a year, the most important work at Alibaba would no longer depend on chips it could not control.

Eddie Wu (吴泳铭), Alibaba's chief executive, did not frame it as a complaint. He gave it a date. The next generation of Pingtouge (平头哥) domestic chips, he said, will be able to support ultra-large-scale model training in the first quarter of 2027.

## What was actually said

The remarks came at Alibaba's results call on November 19, 2026, and were carried by several Chinese financial news outlets including Cailian Press. Wu said the company had already begun training a next-generation model on the current generation of Pingtouge chips.

He then widened the claim. Looking further out, he said most models — Alibaba's own, those of partners, and those of customers — would come to rely entirely on domestic chips for training, in both the pre-training and post-training phases. In Alibaba's 2030 long-term plan, he said, AI infrastructure is built around a Pingtouge-based domestic supply chain.

That is a stronger statement than "we are experimenting with domestic hardware." It puts the company's own roadmap, and the roadmaps of the customers it serves through Alibaba Cloud, on a single silicon bet.

## Why "pre-training and post-training" matters

It is easy to say a chip can "run a model." Inference is the easier half: a trained model executes a fixed sequence of calculations, and a capable accelerator can usually be adapted to it.

Training is different. Pre-training requires moving enormous quantities of data through a network for weeks or months, keeping thousands of accelerators synchronised, and recovering cleanly when one fails. Post-training — fine-tuning and reinforcement learning — is shorter but far less predictable, with frequent restarts and changing workloads.

Wu named both. If the claim holds, it means the Pingtouge platform is intended to carry the entire training lifecycle, not just the cheap, visible part at the end. That is the difference between a chip that can serve a model and a chip that can create one.

It also explains why Alibaba is pairing the hardware announcement with a customer-facing promise. Cloud customers who train on Alibaba infrastructure inherit the same supply chain. If it works, they get predictable capacity. If it does not, they inherit the problem too.

## The numbers behind the confidence

Alibaba's cloud business gave Wu something to point at. For the quarter, AI cloud and computing power service revenue was 60.973 billion yuan (≈ US$8.5 billion / HK$66.3 billion, approximate), up 53 percent year on year. Revenue from AI-related products reached 17.918 billion yuan (≈ US$2.5 billion / HK$19.5 billion, approximate), which the company described as its thirteenth consecutive quarter of triple-digit growth.

Wu said the push into domestic computing power was progressing faster than originally expected. He guided the next fiscal year's cloud operation target to a growth range of 30 to 45 percent, and said that as domestic computing power scales through 2028, growth could return to above 50 percent, with cash flow turning positive in the second half of that year. The longer-range target of US$100 billion in external cloud revenue by fiscal 2031, with a margin above 20 percent, was left unchanged.

These are management projections, not results. But the growth figures explain why the chip strategy has urgency: a cloud business expanding at this rate cannot afford a supply chain that pauses.

## What could change the plan

Wu was candid about uncertainty. Global rules governing the export of computing hardware are changing, and he said they will affect products next year. He also disclosed that Alibaba's long-term target of building 20 gigawatts of data centre capacity could be revised downward, with overall capital expenditure plans reduced accordingly.

Read together, these statements describe a company trading one kind of certainty for another. Overseas capacity becomes less predictable; domestic capacity becomes the thing it can plan around. That is a rational response, but it concentrates risk. A domestic chip programme at this scale depends on fabrication capacity, packaging, memory supply, software toolchains and developer habits all maturing on roughly the same timetable.

Software is the quiet constraint. A chip can be taped out on schedule and still fail to displace an established ecosystem if compilers, operators and debugging tools lag. Wu's claim that current Pingtouge chips are already training a next-generation model is therefore the more interesting sentence: it implies the toolchain is further along than the headline roadmap suggests.

## Honest limitations

All quotations and figures come from Alibaba's earnings call as reported by Cailian Press, The Paper and other Chinese financial media. I did not find a verbatim transcript on Alibaba's investor relations site, so the wording should be treated as reported rather than as a checked corporate statement. The revenue figures are for the reported quarter and are not independently audited here.

"Fully support ultra-large-scale model training" is not a technical specification. No chip model number, process node, memory configuration, interconnect bandwidth or benchmark was attached to the 2027 claim in the coverage reviewed. Readers should not infer a performance comparison with any foreign accelerator.

The 30–45 percent and 50-plus percent growth outlooks, the 20-gigawatt revision and the fiscal 2031 target are management guidance, not commitments. Export rules and supply conditions can change them. I did not cover Alibaba's model releases or its consumer AI products, which are separate questions from the training-hardware strategy.

## What readers can do now

- **If you buy cloud AI capacity, ask your provider which silicon your training jobs will actually run on.** The answer increasingly affects both price and availability.

- **If you are a developer, watch the software stack, not the die.** Toolchain maturity is what determines whether a domestic accelerator is usable for your workload.

- **If you follow the sector, separate a roadmap date from a shipping product.** First-quarter 2027 is a target; the proof will be a named chip in volume production.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-chips-pingtouge-domestic-training
Free to quote with attribution and a link to the original.
