NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

A senior Huawei chip scientist breaks silence on the industry's physical wall

In a rare four-hour interview, Huawei's Liao Heng argued that the old semiconductor roadmap is nearing its end and set out a different path.

2026-11-17 · 1078 words · NeuroAI
A senior Huawei chip scientist breaks silence on the industry's physical wall

He does not give interviews. For years, the engineer at the center of Huawei's chip effort has stayed out of view while his company worked to design processors without tools it could no longer buy abroad.

Then, in late July, he sat down for four hours with a Chinese technology interviewer and talked about transistors, memory and the future of computing.

The result was a rare look at Liao Heng (廖恒), a senior Huawei semiconductor scientist — and a blunt warning that the way the world has been building powerful chips is running out of room.

A scientist who rarely speaks

Bloomberg reported on the appearance on August 4, describing it as a rare public moment for the engineer. Liao is not a household name outside the industry. Inside it, he is known as the person who helped shape Huawei's design philosophy after US restrictions cut the company off from key suppliers half a decade ago.

That background matters for reading what he said. Huawei cannot simply order the most advanced lithography machines from ASML and follow the same path as Taiwan Semiconductor Manufacturing or Intel. It has to find performance somewhere else. Liao's interview was less a complaint about that constraint than a defense of a specific answer to it.

The core of his argument is that the industry's habits are approaching a limit. For decades, chipmakers have made transistors smaller and packed more of them onto a die. More recently, they have leaned on ever-larger compute dies and stacks of high-bandwidth memory (HBM) to feed AI accelerators. Liao's message, as Bloomberg related it, is that this cannot continue indefinitely. He warned that the physical limit would have serious consequences; Bloomberg's report framed it as a supply-chain risk.

The case for time over size

Huawei's alternative goes by the name Tau Scaling Law (韬定律). The company detailed the approach in a keynote in Shanghai in May, at an IEEE conference, under the title "New Semiconductor Path in Practice." Rather than treating smaller transistors as the only source of progress, it treats the time a signal takes to travel between parts of a chip and a computer system as the quantity to optimize.

The practical technique behind it is called LogicFolding (逻辑折叠). The idea is to spread critical logic circuits across multiple layers, like folding a sheet, so that wires are shorter and signals arrive sooner. The gain comes from architecture rather than from a finer manufacturing process — from moving data less far, not from making switches smaller.

That distinction is the whole point. Liao's argument is not that Moore's Law is meaningless, but that its economic force has faded: shrinking a node no longer reliably lowers cost per transistor in the way it once did. When brute-force scaling stops paying, the interesting engineering moves to how a system is arranged.

He also tied the chip question to model design. Because Chinese companies work with less compute, he said, they have to spend more effort on design and accept greater complexity in exchange for lower compute consumption. Huawei calls this co-design: shaping the chip and the model together so that clever software offsets a hardware ceiling. He pointed to a Chinese AI lab that made efficiency a virtue out of necessity as evidence the approach can work, without claiming the gap has closed.

An 18-level answer to a layer cake

One of the more striking moments in the interview was a metaphor for the whole AI stack. Where Nvidia's chief executive has popularized a "layer cake" image, Liao described an 18-level pagoda (十八层宝塔). His point was that the chain — from materials and equipment through chips, models and applications — has more layers than the familiar picture suggests, and that Chinese players at each layer need to support one another.

The list he emphasized is less about any single product than about relationships:

  • Chip designers and model developers who tune to each other rather than separately.
  • Domestic suppliers who can fill gaps left by export restrictions.
  • Software toolchains that make a constrained chip usable at scale.
  • A manufacturing base that can keep improving without the most advanced imported equipment.

Huawei has said it intends to release a new generation of its Ascend (昇腾) AI chips each year and double performance with each one. Liao suggested the real test will come when outsiders take apart the company's first smartphone chip built with LogicFolding and measure what the architecture actually delivers. That is a useful admission: the claims so far are the company's own, and independent teardowns are still ahead.

Why a quiet engineer matters

The interview is worth reading not because it settles the technical debate, but because of who delivered it. Liao sits close to the center of China's effort to build a semiconductor path that does not depend on Western tools. When a figure like that speaks at length, the message is as much about confidence as engineering: that the constraint is being treated as a design problem rather than a permanent handicap.

Whether the Tau Scaling Law becomes a durable alternative or a clever set of trade-offs will not be decided in an interview. It will be decided in chips that ship, in benchmarks that outsiders can reproduce and in customers who choose them when they have other options.

Honest limitations

This article rests on reporting of a single interview, principally Bloomberg's account and syndicated versions of it. The technical claims — about signal delay, LogicFolding and the direction of chip economics — are Liao's and Huawei's, not independently verified findings. Specific performance figures circulating in secondary coverage could not be confirmed against two authoritative sources and are therefore left out. Neither the Tau Scaling Law nor the 18-level pagoda is a peer-reviewed framework; they are a company's design philosophy and a metaphor. The interview also touched on US-China technology restrictions, addressed here only as documented business context, not as a political claim. Readers should treat the argument as a position from one engineer with a clear stake in it.

What readers can do now

  • If you buy AI hardware, watch for independent teardowns of Huawei's first LogicFolding smartphone chip rather than relying on announced metrics.
  • If you work in chip design, look up the May IEEE keynote and judge the time-based scaling idea on its technical merits, not on the interview.
  • If you follow the AI supply chain, track whether annual Ascend releases actually arrive on the promised cadence; delivery is the real signal.

Related coverage

Cite this article

NeuroAI (脑机网), "A senior Huawei chip scientist breaks silence on the industry's physical wall," 17 November 2026, https://neuroai.site/a/na-people-liao-heng-huawei

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Builders” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories