From Huawei to DeepSeek: the galapagosization of Chinese tech deepens

and do export controls work?

AI is a parallel computing problem, isn’t it? Why can’t they just put 4x, 10x, as many chips together… If they wanted to, they just gang up more chips, even if they’re 7nm.

…You could gang them together, just like we gang them together with NVL72. They’ve already demonstrated silicon photonics, connecting all of this compute together into one giant supercomputer.

… DeepSeek is not an inconsequential advance. The day that DeepSeek comes out on Huawei first, that is a horrible outcome for our nation.

- Jensen Huang, on Dwarkesh Patel Podcast, April 2026

---

At the May 2026 IEEE International Symposium on Circuits and Systems, Huawei unveiled the τ (tau) Scaling Law, sending shockwaves around the world.

For 50 years, the global semiconductor industry has relied on Moore’s Law: shrinking the physical size of transistors to cram more computing power onto a chip. Because US export controls have cut Huawei off from the advanced ASML EUV lithography machines required to shrink transistors past 5nm, Huawei seems to be shifting the goalposts. Instead of optimizing for size, the Tau Scaling Law optimizes for time: reducing the signal propagation delay (represented in physics by the Greek letter τ ) across devices and systems through LogicFolding technology, a 3D architecture that folds and stacks logic circuits vertically. Huawei states it will debut this architecture in its fall 2026 Kirin mobile processors and claims it will achieve 1.4nm-equivalent performance and density by 2031, all using older, legally available manufacturing equipment.

Let’s make it clear first: a lot of this is clever marketing, and Huawei has always known how to craft a compelling narrative.

What Huawei is proposing is not revolutionary in itself. It’s well known in the semiconductor industry that Moore’s Law has lost power. The world has hit the wall of physics in its quest to create ever-smaller transistors. To scale, the industry is shifting from “how small can we make a transistor?” to “how efficiently can we stack and connect them?” To save time and costs for moving data around, sophisticated stacking, advanced packaging, and system-level communication technologies increasingly matter more than merely carving out the tiniest transistors.

Shortly after the Huawei announcement, Jensen Huang noted that while tau scaling is an impressive breakthrough for Huawei, it does not threaten global foundries. TSMC has been pioneering advanced 3D packaging, chiplet stacking, and hybrid bonding for nearly a decade. Huawei is simply applying these existing concepts hyper-aggressively to compensate for its lack of raw node leadership. Yet their peers, through similar and perhaps currently more advanced packaging on top of their certainly more advanced nodes, still maintain a formidable edge over Huawei.

There is also the problem of thermodynamics. Stacking logic circuits directly on top of each other creates immense thermal management challenges. Dissipating heat from the center of a 3D folded chip is notoriously difficult, which will significantly complicate manufacturing processes, lower yields, and make mass production incredibly expensive. As of now, Huawei hasn’t yet demonstrated it can solve this problem with the right economics.

Furthermore, as Vikram Sekar points out, even if Huawei succeeds here, it’s very likely that Huawei still relies on Western bond manufacturers, such as Besi, a Dutch company, to make it work and is thus not free from potential US sanctions. After all, the Dutch are firmly within the orbit of US influence.

I am not belittling Huawei’s achievements here. I put out all these caveats first to suggest one thing: the actual significance of Huawei’s Tau scaling announcement is not that it has broken free from U.S. sanctions (it clearly hasn’t). The significance is that this announcement provides a roadmap for the future, and it’s a future that’s been played out many times already in history.

Overtaking on a bend - the EV experience

It’s extremely hard for a newcomer to fight against a formidable incumbent. To win, it’s often more desirable for a newcomer to enter a new game - a new paradigm - where everyone has a similar starting line.

The best example is China’s EV sector.

Automobiles are the crown jewel of modern manufacturing that any country with any decent dream of becoming a manufacturing powerhouse would have to win. But it’s also a sector dominated by century-old incumbents with deep, formidable institutional knowledge, complex process know-how, and brand moats that are nearly impossible to overcome.

China tried it before, but failed. Despite commonly held misconceptions about “forced technology transfer”, decades after establishing automobile joint ventures with Western giants like Volkswagen and Ford, China still failed to produce competitive cars powered by internal combustion engines.

A main reason is that although Chinese automakers can develop, their foreign peers are developing as well, and they are developing on top of an already more advanced foundation built through decades of iterations. In the end, for China, it became a desperate race with a target that is constantly moving further ahead. If China kept playing this game, it risked being forever stuck in a catch-up role.

Understandably, Chinese companies found this game to be both tiring and futile. So what did they do? They quit that game. They tried to open a new one.

The often-cited term for this in China is “弯道超车Overtaking on a bend”. In auto racing, overtaking on a “straightaway” requires sheer horsepower to beat an opponent who is already ahead. Overtaking on a curve, however, relies on taking a different line, using agility, and exploiting a change in the track's dynamics to bypass the leader.

The “straightaway” for China is the traditional internal combustion engine (ICE), and the “bend” is the electric vehicle.

China knows that the future of mobility belongs to electrification, for both environmental and energy security reasons. Yet, ICE incumbents indulged too much in their past glories and would be slow to adapt to this future.

And the playing field is level in this new game. Back in 2010, neither Toyota nor Volkswagen had a massive head start in lithium-ion battery chemistry, motor controllers, or smart cabin software. The century of ICE IP became suddenly irrelevant. On the other hand, China already had a massive consumer electronics manufacturing base and control over the refining of critical minerals (lithium, cobalt, rare earths), so it arguably enjoyed even some advantages.

What China eventually did is now well-known. It doubled down on this new game by nurturing champions and culling the laggards, building an entire ecosystem of supply chains around it, and eventually establishing an entirely new set of standards and benchmarks. When, finally, China emerged from Covid, the world suddenly woke to a new reality: Chinese EVs have overtaken legacy gasoline carmakers, on the bend.

What keeps Jensen Huang awake at night

Overtaking on the bend is exactly what Huawei is attempting here.

If the bottleneck of the semiconductor industry has shifted from inside the chip to system-level communication, why bother winning the old game? Why not enter this new game directly? After all, China’s disadvantage in communication and advanced packaging technology is a lot smaller than its gap with state-of-the-art lithography. Making the most advanced chips is the “straightaway”; bundling chips more tightly and making them communicate with each other at the speed of light is the new “bend”.

And perhaps China even has an advantage in this new game. As Paul Triolo points out in his latest article on tau scaling:

There is another layer here that many Western analyses still underestimate: China’s structural advantages in systems engineering and infrastructure integration. Huawei is not merely a semiconductor design company. It is simultaneously a telecom infrastructure company, cloud company, networking company, AI systems integrator, smartphone OEM, operating system provider, and increasingly an AI platform company. Few Western firms possess that degree of vertical integration across infrastructure layers. Under a τ scaling framework, that integration becomes strategically valuable because optimization happens across the entire stack simultaneously.

Beyond tau scaling, one of the hottest trends today is CPO, or co-packaged optics, an industry inevitability that would drastically reduce communication costs within AI systems. But hey, don’t you remember that Huawei started out as a telecommunications equipment company? Also from Paul (emphasis mine):

CPO would allow Huawei to bring optical engines closer to switching and accelerator fabrics, reducing electrical trace length and improving bandwidth density. This plays to Huawei’s strengths: it has deep telecom, optical networking, silicon photonics, switching, and systems-integration capabilities that many pure-play AI chip firms lack. The strategic implication is that Huawei may use CPO not simply as a networking upgrade, but as part of a broader attempt to make Ascend clusters behave more like a unified, latency-managed AI computer, compensating for weaker process-node access through superior data movement.

It’s all but certain that what happened in Chinese EVs would happen again in AI. And just like the EV experience, it will ultimately lead to what I call the “galapagosization” of the Chinese semiconductor industry: because Huawei cannot use global-standard tools or fabrication processes, it is forced to build an entirely separate, vertically integrated ecosystem.

For example, in the tau scaling announcement, they specifically mentioned that they are developing bespoke EDA software and interconnected protocols that are completely isolated from and incompatible with the TSMC-Intel-NVIDIA supply chain. It’s also highly likely that in the quest to compensate for more inferior chip nodes with innovative folding and packaging technologies, Huawei would help usher in an entire ecosystem of material suppliers, equipment makers, and testing partners that cluster around their technology principles and standards. Their clients, Chinese AI companies, will also optimize their products around this new ecosystem, which helps it further diverge from the US tech stack.

Another thing we need to note is that while the Galápagos is a small archipelago, China is a continent. It’s the world’s largest single unified market with 1.4 billion consumers speaking the same language and tens of millions of businesses running on the same operating systems. The scale ensures that evolution will happen fast.

This galapagosization on a continental scale frightens people like Jensen Huang.

In the same episode with Dwarkesh Patel I quoted at the beginning of the article, Jensen said:

With respect to China… what we also want is to make sure that all the AI developers in the world are developing on the American tech stack, and making the contributions, the advancements of AI—especially when it’s open source—available to the American ecosystem. It would be extremely foolish to create two ecosystems: the open source ecosystem, and it only runs on a foreign tech stack, and a closed ecosystem that runs on the American tech stack. I think that would be a horrible outcome for the United States.

… A company developed software, developed an AI model, and it runs best on the American tech stack. I saw that as good news. You set it up as a premise that it was bad news. I’m going to give you the bad news, that AI models around the world are developed and they run best on non-American hardware. That is bad news for us.

Does technology containment work?

Jensen has been at the center of controversy lately. His petulant appearance with Dwarkesh Patel in that podcast episode is only adding to the controversy.

At the center of the controversy is the core US policy dilemma: whether semiconductor containment against China actually works.

Both sides of the debate seem to be able to find supporting evidence.

Take the example of DeepSeek V4. Proponents of export controls point to the rounds of delays in its launch, the fact that V4 hasn’t shocked and awed in the same way that DeepSeek R1 once did, and the fact that its gap with leading US models hasn’t narrowed as major evidence that the US containment strategy is working.

But looking more closely, the real significance of V4 is that the galapagosization is expediting faster than most people expected, and that export control will ultimately be a losing bet.

Here I’d like to quote from this excellent article from FundaAI discussing how DeepSeek V4 represented “the first model custom-built for non-NVIDIA chips”.

A core innovation of DeepSeek V4 is its ability to drastically reduce inference costs by introducing a hybrid attention design to:

compress the KV, reduce the tokens that need to be looked at, retain a local window, then handle long-range information through sparse selection.

This matters for non-NVIDIA hardware. The bottleneck on non-NVIDIA hardware isn’t only compute — it’s also memory bandwidth, data movement, kernel fusion, and software-stack maturity. Reducing KV cache and attention reads/writes directly cuts down on the parts non-NVIDIA hardware finds hardest to handle.

The development paths of DeepSeek and Huawei also start to converge, as both start to place emphasis on system-level innovation rather than on chip-level:

V4 looks more like it was designed for very large optical scale-up systems than for single-chip benchmarks.

Single non-NVIDIA chips still trail flagship NVIDIA GPUs. That gap won’t disappear in the short term. The current trend appears to be using “supernodes” to organize many chips into a larger compute unit. If a single chip is weaker, you compensate with all-optical interconnect, clusters, scheduling, communication, and software stack.

Typical examples: Google TPU’s 9000+-card optical scale-up superpod, and Huawei’s 8912 all-optical-interconnect supernode. NVIDIA, by contrast, currently only has NVL72 — and even with Rubin Ultra, it’ll be NVL576.

But for that route to work, the model has to cooperate. The model can’t generate uncontrolled communication everywhere, can’t read a giant KV cache at every step, can’t lean heavily on kernels that are mature in the NVIDIA ecosystem but that Ascend hasn’t fully caught up on.

V4’s design is a clean fit for that direction. Attention is leaner, KV is smaller, MoE is more GEMM-shaped, inference depends more on compression and sparsity than on full brute-force computation. All of these are better suited to being carried by a very large optical scale-up system.

Thus, DeepSeek and Huawei become the core pillars of this great galapagosization of Chinese AI in real time, with one serving as the software side while the other creates the matching hardware. One can well imagine that together, those two will nurture a whole new ecosystem of species independent of the US tech stack.

So the question of whether the containment strategy works really depends on the timeframe you are referring to. In the short run, it’s hard to deny that it works. But in the long run, it will help nurture a competing standard that would eventually cause trouble for the current leader.

This is all quite common sense. But why do so many Americans still refuse to buy it? Why are there so many debates and negativity against people like Jensen Huang? Is there really no better way for the US to compete with China?

I will try to analyze this in the next post.