Is China finally assembling a workable domestic AI stack? DeepSeek + Huawei combo may be it

Most investors are still looking at AI through the usual US versus China lens. But that’s not the question that really matters right now. As China signals long-term intent through policy frameworks like its new five-year plan focused on AI and quantum technologies, the real issue becomes more practical.

What matters is much simpler. Can China actually run AI at scale inside its own market without relying on anyone else?

That’s a higher bar than it sounds. “Self-reliance” gets thrown around a lot, but it’s easy to overstate. A domestic AI stack only counts if all the key pieces work together in the real world.

That means four layers have to hold up. The models need to handle Chinese language and business use cases. The chips need to be available in real volume, not just prototypes. The software has to be usable so developers can build without headaches. And deployment has to show up where it counts, across cloud platforms, enterprise tools, and regulated industries.

By that standard, China is getting closer to something that could be very valuable in its own market. It’s building a sovereign AI stack that actually works for inference and real deployment.

And that last part is what investors shouldn’t miss. The first wave of serious money in AI won’t come from who has the most advanced model. It’ll come from who can deploy it at scale.

Deployment is the threshold that matters

A lot of analysis still asks whether Huawei can match Nvidia’s best chips.

That is the wrong benchmark. China does not need to match Nvidia across the global frontier to change its own market. It needs enough local compute, enough software compatibility, and enough model quality to support useful workloads at acceptable cost inside China. That bar is lower.

MERICS makes the gap plain. China’s self-reliance push now reaches across the full AI stack, while Chinese chips still lag Nvidia in absolute performance. At the same time, China has real advantages higher up the stack: a protected home market, a deep pool of domestic AI talent, strong open-source participation, and political pressure to localize strategic technology. Those strengths do not close the hardware gap. They do change the commercial threshold.

For many enterprise workloads, the real question is simpler. Can the system deliver acceptable latency, secure supply, workable software support, and enough throughput to power customer service systems, internal copilots, document processing, coding assistance, and sector-specific automation?

That is the threshold investors should be watching. Once AI becomes workflow infrastructure, availability and adequacy can outweigh technical leadership.

The important development is a working bridge across the stack

The clearest sign of progress is the emergence of a bridge between domestic model demand and domestic hardware supply.

Recent reporting on DeepSeek V4 launching on Huawei AI chips highlights exactly this shift. DeepSeek has been rewriting parts of its model code with Huawei and Cambricon for compatibility with Chinese hardware. That is significant because software adaptation is often the choke point in hardware substitution.

That is the difference between parallel self-reliance projects and an internal optimization loop.

Huawei’s side of the story matters for the same reason. Reuters reported on March 27 that Huawei’s new 950PR is drawing interest from ByteDance and Alibaba and is becoming more compatible with Nvidia’s software environment than earlier Huawei offerings. Reuters also said the chip is optimized for inference rather than representing a major leap in raw compute.

In this context, that lines up closely with market demand.

Inference is where deployment demand compounds fastest. Once companies decide to embed AI into products and workflows, they need chips that can serve requests reliably and at scale. A product optimized for inference, with easier migration, goes straight at the next commercial bottleneck. Better CUDA-adjacent compatibility lowers switching friction. DeepSeek’s adaptation work lowers model friction. Together, those shifts reduce two of the biggest barriers to local deployment.

That is why this evidence carries more weight than another headline about national ambition. It shows a stack becoming easier to use in practice.

At the same time, it’s important not to overinterpret momentum. Events like DeepSeek’s outage raising reliability questions show that operational stability is still being tested, even if they don’t fundamentally change the direction of the market.

What counts as proof that the stack is “usable enough”?

This is where the debate needs discipline. “Usable enough” should mean domestic buyers can deploy AI at scale without treating Nvidia as the default operating system of the stack.

There are five practical markers of that shift.

First, model portability: leading Chinese models must run on domestic chips with tolerable performance loss.

Second, software usability: developers must be able to migrate workloads without completely rewriting their tooling and habits.

Third, hardware availability: chips must ship in enough volume to support more than symbolic pilots.

Fourth, buyer validation: major cloud providers, internet platforms, or enterprises must be willing to adopt the stack for production workloads.

Fifth, deployment fit: the hardware must be suited to the fastest-growing layer of demand, which is inference.

China does not need to clear all five perfectly to change the commercial picture. It only needs to clear them well enough for domestic deployment to keep compounding.

The recent evidence suggests visible progress on all five.

DeepSeek’s adaptation work speaks to portability. Huawei’s improving software compatibility speaks to usability. Reported interest from ByteDance and Alibaba points to buyer validation. Huawei’s inference focus fits deployment demand. Reuters’ report that Huawei plans to ship roughly 750,000 units of the 950PR in 2026 suggests real supply ambition rather than laboratory-scale nationalism.

None of those points alone proves market victory. Taken together, they suggest the domestic stack is moving from theoretical availability toward operational viability.

The first winners will be inference-heavy Chinese buyers

The first beneficiaries are unlikely to be frontier labs chasing world-leading training runs. They are more likely to be Chinese buyers whose workloads reward deployment more than peak performance.

That includes cloud providers, enterprise software vendors, large internet platforms, and regulated sectors such as finance, telecom, healthcare, and public administration. These users care about uptime, localization, policy alignment, controllable supply chains, and the ability to roll AI through existing workflows. Many of them do not need frontier economics. They need a stack that works consistently enough to trust.

That logic already had an early market signal. Reuters reported in February 2025 that Chinese chipmakers and cloud providers rushed to support DeepSeek models, with analysts describing it as a meaningful step toward reducing reliance on US technology. The current Huawei-DeepSeek alignment looks like a more developed version of that trend. The earlier signal was demand for compatibility. The newer signal is coordinated adaptation between model builders and domestic chip vendors.

That is a more advanced stage of stack formation.

The broader market context points the same way. As AI gets embedded into ordinary workflows, the commercial bottleneck shifts away from headline model quality alone and toward whether tools can be deployed broadly, cheaply, and on approved infrastructure. In that environment, a stack does not need global dominance to create local value. It only needs to become the default practical choice for domestic deployment.

If Huawei-class chips plus DeepSeek-class models become the standard infrastructure for Chinese enterprise AI, that is already a meaningful market outcome, even with Nvidia retaining an advantage at the top end.

What this still does and does not prove

Investors should keep the limits in view, because this is exactly where weak analysis turns a deployment story into a parity fantasy.

Huawei is not Nvidia. China has not solved the high-end semiconductor bottleneck. Chinese chips still trail Nvidia in frontier performance, ecosystem depth, and developer entrenchment. Developments like Nvidia securing China approval for H200 chips while adapting alternatives reinforce that global leaders still hold a strong position at the frontier.

That leaves three realities in place.

Nvidia remains stronger where the hardest training economics matter most.

China’s domestic stack is more credible in inference than in frontier training.

The current evidence establishes growing deployability more clearly than full ecosystem independence.

Those are the boundary lines that keep the thesis honest.

The shift investors should watch

The next phase of the China AI story turns on who becomes the default infrastructure inside China’s own walls.

That is the sharper investor question because it points straight to revenue, procurement, cloud demand, enterprise adoption, and software lock-in. If domestic models are tuned for domestic chips, if domestic chips become easier to program against, and if major Chinese buyers treat that stack as good enough for production, then China can build a large and defensible AI market of its own without catching Nvidia at the frontier.

The stack does not need to be the best in the world. It needs to be usable enough for Chinese deployment to keep moving on domestic infrastructure.

Once that threshold is crossed, even partially, the market consequence is operating leverage. That is where the first serious winners will be made.