The financial press is addicted to a simple, comforting narrative: Nvidia has been locked out of the Chinese market by US export controls, and Huawei is surging to fill the void with its Ascend chips. It is a story of national defiance and technological parity. It is also fundamentally wrong.
If you treat this situation as a supply-and-demand problem, you are looking at the math but missing the physics. I have sat in the rooms where CTOs and systems architects evaluate their infrastructure spend. I have watched them look at the specs for Ascend 910B chips, compare them to the H100, and make their decisions. The "surge" in Huawei sales is not a validation of technological superiority. It is a function of a command economy forcing adoption where the free market would otherwise reject it.
Let us dismantle the delusion.
The Software Moat is a Canyon
The industry suffers from a recurring hallucination that a GPU is just hardware. People look at flops, memory bandwidth, and interconnect speeds. They see that Huawei has narrowed the gap on paper. They assume the rest follows.
They are wrong. Nvidia is not winning because of the H100 chip. Nvidia is winning because of CUDA.
CUDA is not just a software library. It is a fifteen-year head start in developer experience, compiler optimization, and library maturity. When an engineer trains a massive language model, they do not just want raw throughput. They want the stack to work the first time. They want the most esoteric tensor operations to be optimized out of the box. They want PyTorch to run without hitting a segmentation fault in the middle of a three-week training run.
Huawei’s software ecosystem, CANN, is a valiant attempt to build a house on quicksand. It is getting better, sure. But every hour an engineer spends porting code from CUDA to CANN or MindSpore is an hour of productivity lost. In the world of high-performance computing, time is the only currency that matters. If your model takes 30% longer to train because the compiler isn't optimized for your specific architecture, you have lost the race, regardless of the hardware you possess.
Companies in China are buying Huawei because they have no other choice. It is a forced pivot, not an innovation.
The Yield and Lithography Wall
Let’s talk about the physical constraints that the boosters ignore. You cannot "sell" your way out of physics. The US restrictions are not just about banning the sale of the chips; they are about choking the supply of the equipment used to manufacture them.
Huawei’s ability to scale is constrained by access to high-end deep ultraviolet (DUV) lithography machines and the inability to secure extreme ultraviolet (EUV) systems. This is the bottleneck that defines the next decade. Even if Huawei’s designs were identical to Nvidia’s, they are hamstrung by yield rates.
Imagine a scenario where you design a perfect sports car engine, but you can only cast the block with inferior sand molds. Your engine will overheat, crack, or fail to hit its power rating. That is the Huawei reality. Their yield rates—the percentage of usable chips per wafer—are a closely guarded secret for a reason. If they were high, we would see broader deployment across non-state sectors. Instead, we see them concentrated in projects where the government writes the check, regardless of the inefficiency.
When a CEO tells me their firm is "migrating to Huawei," I know exactly what is happening. They are accepting a significant performance tax on their compute clusters to avoid the geopolitical risk of relying on a supply chain that might be cut off tomorrow. This is insurance, not progress.
The Subsidy Trap
There is a concept in economics called the "cost of adoption." Usually, this is about the price of the product. In the Chinese AI chip market, the cost of adoption is hidden. It is subsidized by state capital, state-owned enterprises, and mandatory infrastructure requirements.
When you see a report about "surging sales," ask yourself who is buying. Is it a startup competing in the open market? Rarely. It is a state utility, a telecommunications giant, or a research institution funded by public money. These entities are mandated to prioritize domestic vendors.
This creates a perverse incentive structure. If you are a hardware manufacturer, why would you iterate at breakneck speed to satisfy a picky customer base when you are guaranteed sales by the state? Nvidia had to survive the brutal Darwinism of the open market. They had to innovate, optimize, and serve developers who would abandon them in a heartbeat for a better alternative. Huawei is protected from this selection pressure. They are being kept alive by state oxygen.
This does not build a dominant competitor. It builds a bloated, inefficient national champion that will eventually hit a wall when it has to compete with global standards in markets where state-mandated purchasing does not apply.
Addressing the Counter-Argument
The most common retort I hear is this: "China will just catch up through pure scale and grit."
This misunderstands the mechanics of modern AI infrastructure. This is not the space race. You cannot simply throw more engineers at the problem to overcome the lack of access to high-end lithography and the sheer difficulty of matching a mature software stack.
Software compatibility is a winner-take-all dynamic. Even if Huawei matches the hardware specs, they are still running a fragmented stack. If you are a developer, do you spend your time optimizing for the global standard (Nvidia/CUDA) or for a niche, regional standard? You choose the former. This is the "network effect" in full force. The more developers stick to Nvidia, the better the Nvidia software becomes. The better the software becomes, the more developers stick to it. Huawei is outside this loop, and they are struggling to build a parallel track.
The Nvidia Strategy
Nvidia is not stalling. They are pivoting.
Watch their moves carefully. They are not chasing the low-end, commodity hardware market. They are retreating from the markets they cannot legally serve, yes, but they are also doubling down on the high-margin, high-value ecosystem that China cannot access.
By pushing the limits of the Blackwell architecture and focusing on integrated systems (not just chips, but entire racks, interconnects, and cooling solutions), they are raising the bar. They are making it harder for anyone—Huawei included—to compete, even if they somehow miraculously solved the manufacturing bottleneck.
Nvidia is playing a game of raising the floor. If you are a Chinese company trying to replicate what Nvidia did two years ago, you are already losing because the target has moved. You are chasing a ghost.
What Actually Happens Next
The tech industry loves to predict the "fall of the giants." It makes for a compelling story. But look at history. When has a company ever successfully displaced a software-hardware integrated giant through state-led mandate alone? It hasn't happened.
We will see a bifurcated market. On one side, a state-protected, localized, and inherently inefficient Chinese AI ecosystem that relies on "good enough" hardware and forced software adoption. On the other, a global, hyper-competitive, high-performance ecosystem driven by Nvidia and its partners.
The companies that succeed in China will be those that figure out how to bridge this gap. They will run hybrid environments. They will do their heavy lifting on whatever hardware they can get their hands on, and they will likely spend millions of dollars creating internal abstraction layers to make their code run on both Nvidia and Ascend.
This is the hidden cost. It is a tax on innovation. It is a drain on capital. It is not "self-sufficiency." It is expensive, slow, and structurally inferior.
If you are an investor looking at these headlines, stop reading the sales figures. Look at the software adoption rates. Look at the yield reports of the manufacturers. Look at the talent drain. Huawei is not winning the AI war. They are fighting a rear-guard action to stay relevant in a market that is being boxed out of the global standard.
Do not mistake necessity for victory. The "surge" in Huawei sales is the sound of a market settling for second place, not the roar of a challenger taking the throne. When you strip away the nationalistic pride and the state subsidies, you are left with a simple, brutal reality: you cannot legislate your way into matching a company that has spent two decades building the foundation of modern computing.
Nvidia isn't stalling. It’s just leaving the room, and the door is closing behind it.
