For years, the most obvious explanation for Nvidia’s dominance in artificial intelligence has been its hardware.
The company makes extraordinarily powerful accelerators. But hardware alone does not explain why Nvidia became the default platform for much of the AI industry.
The deeper advantage is CUDA—the software ecosystem that allows developers to build, optimize and run AI workloads on Nvidia hardware.
That is the moat China is now trying to attack.
And DeepSeek and Huawei have just taken a meaningful step in that direction.
The chip is only half the battle
On September 30, Chinese AI company DeepSeek announced a collaboration with Huawei to develop software optimized for Huawei's Ascend AI processors.
The companies released open-source tools covering computation and communication, while DeepSeek also adapted its TileLang programming framework to Huawei's Ascend architecture. Reuters described the move as an effort to reduce reliance on Nvidia's ecosystem.
This matters because replacing Nvidia is not simply a matter of buying another processor.
An AI company that has spent years building models around Nvidia GPUs and CUDA cannot simply unplug Nvidia chips and plug in Huawei accelerators.
The software has to be rewritten.
The models have to be optimized.
The communication libraries have to work.
Developers have to learn the new architecture.
Performance has to be tuned.
And the entire computing stack has to become reliable enough for production.
That switching cost has been one of Nvidia's strongest defenses.
China is now trying to reduce it.
DeepSeek is becoming part of Huawei's software strategy
DeepSeek's importance is not that it has suddenly produced a chip capable of beating Nvidia.
It hasn't.
The significance is that one of China's most important AI laboratories is helping make Huawei's hardware easier for other Chinese developers to use.
That changes the problem.
Instead of Huawei having to independently reproduce Nvidia's enormous software ecosystem, DeepSeek can contribute tools that make the Ascend architecture more accessible to China's AI developers.
And because the tools are open source, other developers can inspect, modify and build upon them.
Reuters reported that the collaboration includes DeepGEMM-Ascend for computation, DeepEP-Ascend for chip-to-chip communication and Ascend support within TileLang.
The objective is not necessarily to beat CUDA tomorrow.
It is to make leaving CUDA less painful.
That may ultimately be the more important battle.
This is what a technological moat looks like
Nvidia's advantage can be thought of as a stack.
At the bottom is the physical chip.
Above that is networking and memory.
Above that are drivers and libraries.
Then comes CUDA.
Then thousands of frameworks, applications, optimization techniques and developers who already know how to use the system.
The higher layers reinforce the lower ones.
That creates a network effect.
More developers use Nvidia because the software is mature.
More software is optimized for Nvidia because so many developers use it.
Cloud companies buy Nvidia because customers want Nvidia.
AI laboratories train on Nvidia because their software already works there.
And developers continue learning CUDA because Nvidia remains the dominant platform.
That is much harder to replicate than a processor.
China has discovered the switching-cost problem
China's AI industry has already demonstrated this.
DeepSeek's earlier models were heavily optimized around Nvidia hardware and software, making them harder to run efficiently on Chinese chips.
The Information reported that DeepSeek has subsequently begun using Huawei chips to train some models, although Nvidia remains important for its largest and most powerful workloads.
That is an important distinction.
China has not already replaced Nvidia.
It is building the conditions under which replacement becomes increasingly practical.
And that process appears to be accelerating.
Huawei needs more than chips
Huawei is China's strongest domestic challenger in AI accelerators, but its challenge has never been simply manufacturing silicon.
It needs an ecosystem.
The company has developed its Ascend processors and its CANN software platform. Chinese companies including Huawei, SMIC and domestic memory manufacturers are also attempting to reconstruct different parts of the AI-computing supply chain. CSIS describes Huawei as being in the strongest position among China's domestic AI-chip designers.
DeepSeek's contribution therefore has strategic value.
It brings an AI laboratory that understands large-model workloads directly into the optimization process.
That creates a feedback loop:
Huawei builds the hardware → DeepSeek optimizes AI workloads → software becomes easier to port → other Chinese AI companies can adopt the platform → more developers contribute software → Huawei's ecosystem becomes more useful.
That is how a rival platform begins to emerge.
And DeepSeek itself is going after silicon
There is another layer to the story.
Reuters reported in July that DeepSeek was developing its own AI inference chip.
The project was reportedly focused on inference—the process of running trained models—rather than training them. The effort was still at an early stage, with DeepSeek hiring chip-design engineers and speaking with potential semiconductor partners.
That tells us something important about China's strategy.
The objective isn't necessarily to find one company that becomes China's Nvidia.
It may be to create an entire domestic ecosystem in which Chinese AI companies have multiple hardware and software options.
Huawei can provide accelerators.
DeepSeek can optimize models and develop software.
Chinese foundries can manufacture chips.
Domestic memory suppliers can support the systems.
And Chinese cloud providers can deploy them.
The ultimate goal is technological substitutability.
Export controls may have accelerated the process
There is an uncomfortable irony for Washington.
U.S. export controls were designed to restrict China's access to advanced AI computing.
They have undoubtedly constrained China's access to the most advanced Nvidia processors.
But restrictions also create a powerful incentive to develop alternatives.
Thorbecke's analysis makes this point: imperfect enforcement and continued access to Nvidia chips through various channels have given Chinese companies time to transition while Huawei improves its own technology.
China does not need to replace Nvidia overnight.
It only needs to make dependence progressively less necessary.
That is a much more achievable objective.
The 160,000-chip question
The scale of China's potential transition is also becoming more interesting.
Bloomberg reported in September that DeepSeek planned to deploy at least 160,000 Huawei Ascend 950DT accelerators at a massive data center in Inner Mongolia.
The reported plan was primarily for inference rather than the most demanding model-training workloads, and the timeline depends on Huawei's ability to produce the processors.
Even if the exact deployment ultimately changes, the strategic direction is significant.
The Chinese industry is moving from:
“Can Huawei make an alternative to Nvidia?”
toward:
“Can we build enough infrastructure around Huawei that Nvidia is no longer indispensable?”
Those are very different questions.
Nvidia still has a formidable lead
None of this means Nvidia's moat has disappeared.
It hasn't.
Nvidia's CUDA ecosystem remains extraordinarily mature, and its hardware, networking, software libraries and developer base reinforce one another.
Chinese companies themselves still use Nvidia hardware for important workloads.
Reuters reported that DeepSeek continues to rely on Nvidia for some of its most powerful models.
And semiconductor manufacturing remains another enormous obstacle.
China must contend with restrictions affecting advanced manufacturing equipment, high-bandwidth memory and other parts of the semiconductor supply chain.
So it would be premature to declare victory for Huawei.
The interesting development is different.
The moat is being attacked from the side.
The real battle is switching costs
This is why the DeepSeek-Huawei partnership deserves attention.
The most dangerous competitor to an incumbent does not always begin by producing a better product.
Sometimes it begins by making the incumbent easier to leave.
That is what open-source software can do.
If Chinese developers can take existing AI workloads and increasingly move them from Nvidia's CUDA environment to Huawei's Ascend ecosystem without rewriting everything from scratch, Nvidia's switching costs decline.
Once switching costs decline, hardware differences become more important.
Price matters more.
Availability matters more.
Government procurement matters more.
Domestic supply matters more.
And geopolitical risk matters more.
That could gradually transform Nvidia's position in China.
China doesn't need to beat Nvidia globally to achieve its objective
This is perhaps the most important point.
Beijing's strategic objective does not necessarily require Huawei to become the world's dominant AI-chip company.
It may simply require China to become self-sufficient enough that American export controls lose their leverage.
That is a much lower threshold.
If Chinese AI companies can train some models on domestic hardware, run inference at scale on domestic accelerators and develop software that works efficiently across those systems, then China's vulnerability to Nvidia—and by extension to U.S. semiconductor policy—declines.
That is the strategic prize.
And it explains why DeepSeek's software work may ultimately matter more than another benchmark showing that one chip is faster than another.
Nvidia's moat is entering a different phase
For years, Nvidia's competitive advantage could be summarized as:
best chips + CUDA + ecosystem + developers.
China is now attempting to attack the middle of that equation.
Huawei is developing alternative hardware.
DeepSeek is helping optimize software for it.
Chinese AI companies are being pushed toward domestic infrastructure.
And developers are being given increasingly usable tools for making the transition.
The goal isn't necessarily to prove that Huawei is already better than Nvidia.
The goal is to ensure that Huawei doesn't have to be better for China to choose it.
That is a very different kind of competition.
And if China succeeds in lowering the software barrier between Nvidia's ecosystem and domestic alternatives, one of Nvidia's most formidable advantages could become substantially less formidable—not because CUDA suddenly stops being good, but because Chinese developers will have another viable place to go.
For Nvidia, that is the part of the China story worth watching.
Not whether Huawei beats Nvidia in a benchmark this year.
Whether China can make Nvidia optional.
Comments
Post a Comment