DeepSeek and Huawei collaborate to close the gap with Nvidia hardware

A strategic hardware partnership

DeepSeek, a startup based in Hangzhou, is working closely with Huawei Technologies to strengthen China's internal AI ecosystem. This collaboration is specifically designed to narrow the performance gap between Chinese technology and American hardware. By combining DeepSeek's software expertise with Huawei's hardware, the two companies aim to create a more independent infrastructure.

The partnership is fueled by both internal and external pressures to build a fully domestic Chinese AI stack. This means developing the necessary tools, chips, and models within the country rather than relying on foreign providers. The goal is to ensure that the domestic tech ecosystem remains controllable and less susceptible to international restrictions.

According to a study by the research firm SemiAnalysis, this cooperation is already showing results in how software interacts with hardware. The move is seen as a significant step toward creating a self-sustaining cycle where Chinese models run efficiently on Chinese chips. This synergy is intended to provide a viable alternative to the current global standards set by companies like Nvidia.

Software tools for domestic chips

A major component of this collaboration involves DeepSeek releasing software tools tailored for Huawei’s Ascend AI hardware. These tools include six software modules that are similar to the ones DeepSeek previously developed for Nvidia chips. By creating these specific modules, DeepSeek is helping to make Huawei's hardware more compatible with high-level AI development.

One of the key releases is an Ascend-compatible version of TileLang. This is a custom programming language that DeepSeek designed to simplify the creation of high-performance software kernels. Kernels are the core parts of software that handle the most intensive computations, and making them easier to develop is essential for high-speed AI processing.

DeepSeek described the Ascend version of TileLang as a way to promote a more independent and controllable domestic software ecosystem for GPUs. By providing the software framework, they allow developers to write code that extracts the maximum power from Huawei's hardware. This mirrors the way Nvidia provides software layers to ensure its own chips perform optimally for users.

Bridging the performance gap

The technical progress made by Huawei's engineering teams is also playing a critical role in this shift. Recent testing on the Ascend 950PR chip shows that new Huawei chip kernels are achieving efficiency at 99.7 percent of their theoretical ceiling. This high level of efficiency suggests that the hardware is capable of performing very close to its maximum possible speed.

While the Ascend 950PR was released in March, it is still considered to be behind Nvidia’s H100 chip, which came out in 2022. However, the collaboration with DeepSeek is helping to minimize the performance difference between these two systems. By optimizing software to work perfectly with the available hardware, the gap in practical use becomes much smaller.

SemiAnalysis noted that the ecosystem gap is definitely closer because of DeepSeek's help on top of Huawei's engineering work. This suggests that software optimization is a primary lever for overcoming hardware limitations. If the software can squeeze every bit of power out of the chip, the underlying hardware differences become less of a hurdle for developers.

Model access and inference

Beyond software modules, DeepSeek is providing Huawei with early access to its latest AI models. This includes models such as V4 and V4.1, which allows Huawei to offer immediate inference support upon their release. Inference is the process where an AI model takes in new data and produces an output, such as generating text or an image.

This early access ensures that Huawei's hardware can be tested and deployed with the latest models as soon as they become available. It creates a streamlined pipeline where new models can be integrated into Huawei's hardware ecosystem quickly. This speed is vital for staying competitive in the fast-moving field of artificial intelligence.

The long-term goal of this arrangement is for DeepSeek to rely primarily on Huawei’s technology stack by next year or 2028. This would represent a near-total transition to a domestic hardware and software foundation. Such a shift would mean that the entire lifecycle of an AI model, from training to deployment, would occur within a Chinese-made ecosystem.

The impact on the global market

The move toward a domestic ecosystem has significant implications for the global AI market. Nvidia's leadership has expressed concern about this trend, with CEO Jensen Huang previously calling it a "horrible outcome" if DeepSeek optimized its technology for Huawei architecture. His comments highlight how significant the shift toward non-Nvidia hardware could be for the industry.

As DeepSeek and Huawei continue to work together, the dominance of US-made chips may face increased competition from high-performing domestic alternatives. The success of this partnership depends on whether the combined software and hardware can match the reliability and scale of current global leaders. If they succeed, it could lead to a more fragmented global AI landscape with multiple competing standards.

For researchers and developers, this means more options for running large-scale models on different types of hardware. While Nvidia currently holds a leading position, the progress of the Huawei and DeepSeek partnership shows that the gap is closing. The focus is moving toward how well software can be tuned to maximize the capabilities of any available chip.