A striking claim from a non-public discussion

DeepSeek founder and chief executive Liang Wenfeng has made an unusually blunt assessment of the AI hardware market: export restrictions, he argued, are forcing Chinese developers to build domestic alternatives and are therefore helping Nvidia’s future competitors. In remarks reported from a lengthy investor exchange, Liang said Nvidia was “digging its own grave” by becoming less accessible to Chinese customers.

The most consequential part of his argument concerned Huawei’s Ascend 950-based Atlas 950 SuperPoD. Liang said the system could substitute for Nvidia’s GB200 and GB300 offerings in performance and price, while also acknowledging a substantial caveat: roughly four Huawei accelerators would be needed for the work of one Nvidia unit, with the Chinese hardware around two years behind technologically.

Those remarks should be read as a strategic assessment by a customer and potential ecosystem partner, rather than as an independently verified benchmark result. They nevertheless describe an important change in the competitive logic of AI infrastructure. The question is no longer simply whether a single Chinese accelerator can equal Nvidia’s best silicon. It is whether Chinese suppliers can deliver enough performance, software support and manufacturing capacity for domestic AI companies to operate at scale without relying on Nvidia.

Replacement is not the same as chip-for-chip equality

The language of “complete substitution” can obscure the difference between a component comparison and a system-level calculation. If four accelerators are required to match one competing device, then equality has clearly not been demonstrated at the individual-chip level. The economic result may still be acceptable to a buyer if domestic hardware is available, supported locally and can be deployed in sufficiently large clusters.

That calculation includes far more than headline compute throughput. Training and serving large models depend on memory capacity and bandwidth, interconnect performance, power consumption, cooling, reliability, compilers, libraries and the ability to schedule work over thousands of processors. The cost of porting and maintaining model code is also central. A system that requires more processors may require more racks, networking and electricity, potentially narrowing any apparent price advantage.

Huawei’s recent public presentation of the Atlas 950 SuperPoD underlines the scale of its ambition. The company describes a 1,024-accelerator configuration designed for very large model training and high-concurrency inference, alongside a unified memory addressing architecture and its own software environment. Those specifications show that Huawei is pursuing a full-cluster alternative, not merely a standalone Nvidia-like accelerator.

However, neither Huawei’s product announcement nor the reported DeepSeek discussion provides reproducible, workload-specific comparisons against GB200 or GB300 systems. Until such measurements are published and independently tested, claims of equivalence should be treated as commercial and strategic assertions.

Software is the harder competitive frontier

Liang’s comments place unusual emphasis on software. He said DeepSeek had trained its V3 model on Nvidia hardware while moving away from Nvidia’s software ecosystem through a high-level compiler called TileLang. That distinction matters because Nvidia’s advantage has long rested not only on processors but also on a mature developer environment, tools and libraries.

DeepSeek’s own technical work supports the broader point that hardware-aware software can make a major difference. Its published analysis of the V3 infrastructure describes training on 2,048 Nvidia H800 GPUs and details efforts to reduce communication and memory bottlenecks through model and network design. Efficient architectures can reduce the amount of hardware required for a given task, which makes alternative accelerators more practical.

Still, moving a single laboratory’s workloads away from one software stack is not the same as creating a durable industry-wide platform. A domestic ecosystem has to support researchers, cloud providers, enterprise developers and the many frameworks used across the AI market. It must also keep pace as models, precisions and inference techniques change. This is why Huawei’s investment in software and open-source communities may be as significant as the specifications of the underlying hardware.

Export policy creates both pressure and opportunity

Liang is right that restrictions can give domestic substitution a powerful incentive. Nvidia itself has warned investors that changing export controls can push customers to adopt alternatives and expand competing developer ecosystems. Its filings also show the financial cost of policy uncertainty: new licensing requirements for China-bound H20 products led to a multibillion-dollar inventory and purchase-obligation charge in 2025.

But the policy environment is more complicated than a total ban. In January 2026, the US Commerce Department said that licence applications for Nvidia H200, AMD MI325X and comparable products could be reviewed case by case if specified security conditions are met. Access therefore depends on approvals, end users and compliance requirements, rather than on a simple binary of availability or prohibition.

That uncertainty is itself an advantage for local suppliers. Datacentre operators planning multiyear investments may value predictable procurement, technical support and regulatory alignment as highly as peak performance. Even customers that can obtain US hardware have a reason to test domestic alternatives if future access might change.

Nvidia remains the benchmark, but the market is changing

The available evidence does not show that Huawei has matched Nvidia across AI hardware, systems software and global developer adoption. Nvidia remains the reference point that Chinese competitors are trying to reach, and DeepSeek’s published V3 work was trained on Nvidia H800 hardware.

Yet Liang’s central argument is more credible when framed as an ecosystem forecast rather than a declaration of present-day parity. Restrictions, supply risk and national industrial policy can create a protected proving ground for Chinese hardware. If Huawei and its partners can turn large deployments into dependable tools for training and inference, the resulting software base could become competitive beyond China as well.

For Nvidia, the immediate risk is not that one Huawei product suddenly makes its technology obsolete. It is that limited access to Nvidia products accelerates the very customer migration, developer investment and systems integration that make a rival platform progressively more viable. That is the grave Liang was referring to: not a sudden technical defeat, but the long-term cost of losing an ecosystem before alternatives have fully matured.

Sources