Migrating AI from Nvidia to Huawei: Opportunities and trade-offs

When contemplating the shift, several business advantages may drive a final decision. Relying on one major vendor (namely, Nvidia) can incur risks: pricing leverage, export controls, supply shortages, or a single point of failure in innovation. Adopting or migrating to Huawei has the potential to provide negotiation leverage, avoid vendor lock-in, and offer access to alternate supply chains. That’s especially relevant in areas where Nvidia faces export restrictions.

If an organisation operates in a region where Huawei’s ecosystem is stronger (e.g., China, parts of Asia) or where domestic incentives favour local hardware, shifting to Huawei could align with corporate strategy. For instance, ByteDance has begun training a new model primarily on Huawei’s Ascend 910B chips with notable success.

Huawei’s technology focuses on inference and large-scale deployments, and thus may be better suited to long-term use, rather than occasional use of large infrastructures for training, followed by less intensive inference. If an organisation’s workloads are inference-heavy, a Huawei stack may offer advantages in cost and power. Moving Huawei’s internal clusters (e.g., CloudMatrix) have shown competitive results in select benchmarks.

While migration offers potential gains, several challenges exist. Nvidia’s CUDA ecosystem remains unmatched for tooling and community support, with Nvidia established as the go-to solution for most companies and businesses. Migrating to Huawei’s Ascend chips and CANN software stack may require re-engineering workloads, retraining staff, and adjusting frameworks. Those are not considerations to be taken lightly.

Additionally, Huawei hardware still lags Nvidia in high-end benchmarks. One Chinese firm reportedly needed 200 engineers and six months to port a model from Nvidia to Huawei, yet only achieved about 90% of prior performance. The wholesale rebuilding of development pipelines will incur engineering and operational costs. If significant investment in Nvidia hardware and CUDA-optimised workflows exists, switching will not yield short-term savings.

And while use of Huawei technologies mitigates dependency on Western chips, it may introduce other regulatory risks given the controversy around the company’s hardware in critical national infrastructure. That’s particularly relevant in global markets where Huawei hardware faces restrictions of its own.

There are several case studies showing Huawei technologies effectiveness. ByteDance, the company behind TikTok has trained new large models on Huawei’s Ascend 910B hardware. DeepSeek is currently launching AI models (V3.2-Exp, for example) that are optimised for Huawei’s CANN stack.

Companies may wish to consider dual-stack approaches for flexibility. Regardless, any consideration of migration should include the following:

• Assessment current pipeline and dependencies.
• Defining migration scope (training vs inference).
• Evaluation of Huawei’s ecosystem maturity (Ascend, CANN, MindSpore).
• Running pilot benchmarks on the new tooling.

Ongoing activities will need to include:

• Training teams and retooling workflows.
• Monitoring of supply-chain and changing geopolitical factors.
• Measuring performance and productivity metrics.

Migrating an internal AI model development pipeline from Nvidia to a Huawei-centred stack is a strategic decision with potential business advantages: Vendor diversification, supply-chain resilience, regional alignment, and cost optimisation. However, it carries non-trivial risks. With many industry observers becoming wary of what they see as an AI bubble, an organisation’s strategy has to be fixed firmly on an AI future, despite the potential to be affected by financial market fluctuations and geo-political upheaval.