Huawei’s Answer to Better Chips Is More Chips

Picture a procurement meeting somewhere in Shenzhen in late 2026. Somebody’s laptop is open to a spreadsheet with two columns. One column is what Nvidia can sell into China legally, on what timeline, at what price. The other column is a Huawei roadmap with dates on it. Neither column looks great. But only one of them is guaranteed to still exist next quarter regardless of what happens in Washington.

That meeting is the entire story of Huawei’s Ascend push, and it explains why the company just pulled its Ascend 960DT launch forward to early 2027.

What we actually know

The verified pieces are thin but pointed. Huawei is accelerating the 960DT to early 2027. It plans cluster-based SuperPod deployments reaching up to 100,000 chips. The stated goal is to replace Nvidia inside China and compete globally on raw computing power. Huawei has also laid out a longer roadmap — Ascend 960 in 2027, Ascend 970 in 2028, following the 910C from 2025 — and disclosed work on its own high-bandwidth memory alongside new Atlas systems and a UnifiedBus interconnect it claims can link processors into very large clusters.

That’s it. Everything else floating around this topic is speculation, and I’d rather give you a short list of facts than a long list of guesses.

The strategy is scale, not silicon

As someone who spends most of his time testing tools rather than reading press releases, the thing I find genuinely interesting here isn’t the chip. It’s the shape of the bet.

Huawei is not claiming it will beat Nvidia on a per-die basis. It can’t, at least not while its manufacturing options are constrained. So it’s doing the other thing you can do when your individual units are weaker: build a lot of them and wire them together well. SuperPod at 100,000 chips and UnifiedBus as the connective tissue are both bets that system-level engineering can absorb a per-chip deficit.

This is a legitimate engineering strategy. It is also an expensive, power-hungry, operationally messy one. Anybody who has tried to keep a modest GPU cluster healthy knows the failure modes multiply faster than the node count. Scaling to six figures of accelerators means every percentage point of node failure, every interconnect hiccup, every scheduler quirk becomes a budget line item.

Why the toolkit angle matters more than the benchmark

Here’s where I get skeptical, and it has nothing to do with the hardware.

Nvidia’s actual moat has never been purely the transistors. It’s CUDA, and the fifteen-plus years of libraries, kernels, forum posts, Stack Overflow answers, and framework integrations built on top of it. When you pick up a new model repo, it assumes CUDA. When something breaks at 2am, somebody has already broken it before you and written about it.

Huawei’s CANN stack and the Ascend tooling around it are the part I’d want to test before I formed an opinion on any of this. The questions that decide whether these clusters get used in practice are unglamorous:

• How much of a standard PyTorch training script runs unchanged?
• When a custom kernel doesn’t port, how hard is the rewrite, and who documents it?
• What does the profiling and debugging story look like at scale?
• How fast do new model architectures get support after release?
• Is the documentation in a language and format your team can actually use?

None of that shows up in a roadmap slide. All of it shows up in your engineers’ sprint velocity.

The proprietary memory detail

The HBM disclosure deserves a note. High-bandwidth memory has been one of the tighter chokepoints in AI accelerator supply, and going in-house on it suggests Huawei is planning for a world where it can’t rely on anyone else’s components. That’s a defensive move dressed as a technical one, and it tells you something about the time horizon the company is planning against. You don’t build your own memory for a two-year problem.

My honest read

For buyers inside China, the calculation is simple enough that software friction becomes an acceptable cost. If the alternative is uncertain supply, you take the option that ships.

For everyone else, this needs to clear a much higher bar. Competing globally on computing power means competing against an ecosystem, not a spec sheet, and there is no evidence yet either way on how that’s going.

Early 2027 is far enough out that a lot can change. What I’d watch between now and then isn’t chip announcements — it’s whether independent developers outside Huawei start publishing real Ascend results with real numbers. That’s the signal. Everything before it is positioning.

When there’s hardware I can test, I’ll test it. Until then, treat the 100,000-chip figure as an engineering ambition rather than a delivered capability, because that’s precisely what it is.

🕒 Published: September 20, 2026