Alibaba and Baidu turn to their own AI chips as US export controls tighten

Chinese cloud companies open a second route to AI training

Alibaba is using its own AI processors to train smaller models, while Baidu is testing its Kunlun P800 chip on new versions of Ernie. As of September 13, 2025, the moves mark a practical shift for two of China's largest consumer internet and cloud companies: processors they designed are beginning to take on work long dominated by Nvidia hardware.

Alibaba's newest training accelerator, known internally to employees as Zhenwu, is already handling smaller training jobs inside Alibaba Cloud. People who have worked with it describe performance as competitive with Nvidia's H20 on some tasks. Baidu's experiments focus on whether its latest Kunlun processor can deliver the stability, processing speed and software support needed for new Ernie models, which underpin its generative AI services.

Both companies continue to depend on Nvidia for their largest and most complex training runs. Those jobs need powerful processors, large amounts of fast memory, reliable software and exceptionally quick connections between chips. The immediate change is a broader choice of hardware, with domestic processors taking work suited to their strengths.

Two pressures are driving that choice. United States export controls have restricted China's access to advanced AI semiconductors, while Chinese government and critical infrastructure customers increasingly prefer technology designed and supported at home. For cloud companies, developing their own chips connects the need for dependable supplies with the demands of customers buying computing services.

For China's data centers, the potential benefit is concrete. A company waiting for imported graphics processing units, or GPUs, could have another way to obtain computing capacity. If domestic accelerators can take a growing share of everyday work, they could help keep AI projects on schedule and give cloud operators more control over spending and expansion.

Alibaba moves beyond its earlier chip designs

Alibaba's chip development business, T-Head, already has experience with processors designed for different computing tasks. It previously unveiled the Hanguang 800 accelerator for AI inference and a server central processing unit, or CPU, based on Arm instruction sets. Instruction sets are the basic commands a processor understands.

Hanguang 800 focused on running AI models after they had been trained. Alibaba's latest accelerator extends its efforts into training, the demanding process of teaching a model to recognize patterns and produce useful responses. Employees refer to that processor internally as Zhenwu.

Alibaba Cloud is using the new chip for smaller models while retaining Nvidia hardware for its hardest training work. People with experience of the accelerator say it can compete with Nvidia's H20 on tasks that do not need the most advanced connections between processors or the largest amounts of memory.

The size and structure of a job therefore matter as much as the chip's headline performance. A processor that handles a smaller model efficiently may be useful across many cloud workloads, even while the largest models remain on a different platform. Alibaba is building experience with its silicon before relying on it more widely.

Baidu tests Kunlun P800 with Ernie

Baidu's Kunlun project is one of China's longest running domestic AI accelerator lines. Earlier Kunlun chips concentrated on inference. With the P800, Baidu is testing training workloads for new versions of Ernie and checking whether the processor and its supporting software can operate reliably together.

The tests cover stability, throughput and software readiness. Stability means keeping a job running correctly; throughput measures how much work the system completes in a given period. Both matter when training consumes substantial computing time and interruptions can delay a model's development.

Baidu controls several parts of this process. It develops Ernie, operates large cloud computing clusters and controls PaddlePaddle, the software framework used to build and run AI models. That gives it room to adjust the model, software and hardware together, seeking better efficiency across the whole system.

Baidu's experiments also reflect the careful pace of the transition. Kunlun P800 is being evaluated for new Ernie training work, while Nvidia remains important for the largest and most demanding runs. The practical goal is to establish which jobs the domestic platform can handle consistently.

Export controls narrow access to imported chips

Washington began tightening access to advanced AI accelerators in 2022 through performance thresholds that limited what US suppliers could ship to China. Suppliers responded with processors tailored to the Chinese market, including GPUs with reduced bandwidth in the connections used to exchange data between chips.

Later measures narrowed those routes. Nvidia's H100 and newer Blackwell processors remain out of reach in China under the restrictions described here. Nvidia developed the less capable H20 to comply with US rules, giving Chinese customers another option for AI computing.

Industry research indicates that shipments of H20 chips also require export licenses starting in April 2025. The same research identifies AMD's MI308 as covered by licensing requirements from that month. Licensing adds uncertainty for buyers trying to plan when computing equipment will arrive and how much capacity they can build.

The sequence helps explain why Alibaba and Baidu are pursuing imported and domestic hardware at the same time:

• In 2022, US performance thresholds began limiting shipments of advanced AI accelerators to China.
• Over the following years, further measures narrowed the scope for processors tailored to Chinese buyers.
• From April 2025, industry research indicates that Nvidia H20 and AMD MI308 shipments require export licenses.
• By September 13, 2025, Alibaba is using its own accelerator for smaller training jobs, and Baidu is testing Kunlun P800 on new Ernie versions.

Chinese cloud and internet companies have responded by obtaining overseas chips where possible while investing heavily in domestic accelerators. Government and critical infrastructure customers also increasingly favor local hardware, giving providers a commercial reason to demonstrate acceptable costs and predictable performance on Chinese processors.

Why training is harder than running a finished model

AI training teaches a model by repeatedly adjusting billions of internal values, known as parameters, as it processes data. The work puts heavy demands on computing power, memory capacity and the speed at which information moves through the system.

Large training runs can spread across thousands of accelerators. Those processors must exchange information and stay coordinated. Fast connections between them are essential because a chip waiting for data from another chip cannot use all its computing capacity productively.

Inference happens after training, when the model generates an answer or another output for a user. Here, the priorities include responding quickly, keeping costs low and limiting power consumption. The model has already learned its patterns, so the hardware is carrying out a different job.

A chip can therefore perform well at inference or at training smaller models while still struggling with training very large models over several months. At that scale, the connections between processors and the software coordinating them can become bottlenecks, limiting how much useful work the entire cluster completes.

Alibaba's reported comparison with the H20 concerns particular tasks and smaller models. For the largest models and the fastest training schedules, Nvidia retains advantages in peak computing power, connections between GPUs, mature compiler software and a broad collection of optimized libraries. H100 and the latest Blackwell designs remain the standard for the most advanced model training outside China.

Software determines how much a chip can deliver

Nvidia's position rests on more than processor design. Its CUDA platform, libraries such as cuDNN and TensorRT, and a large range of tools from partners help developers write software and obtain predictable performance. Libraries provide reusable software for common tasks, reducing the amount developers must build themselves.

Compilers are another part of that system. They translate developers' instructions into operations the hardware can execute efficiently. A powerful chip needs software that can keep it busy, move data effectively and coordinate its work with other processors.

Baidu's PaddlePaddle framework and Alibaba's cloud toolchains already support their own chips. Both companies can tune kernels, the small programs that perform specific calculations, and graph compilers, which organize a model's operations for execution. Control of these tools gives them opportunities to improve efficiency for their own models.

Developers outside the two companies may still need to learn different tools or adapt existing code when moving to domestic hardware. Some may have to retune kernels or use tools specific to the vendor to obtain the best throughput. Such adjustments are common in high performance computing, but they take engineering work.

Alibaba and Baidu can reduce that burden by managing the transition within their cloud services. Customers seeking computing capacity may place more weight on price, availability and reliable service than on the manufacturer of the underlying chip, provided their familiar frameworks continue to work well.

Investors see value in a more dependable supply

Shares of Alibaba and Baidu rose after reports of the deployments. Investors welcomed the prospect that the companies could secure more AI computing capacity without depending entirely on imported processors, supporting product development and reducing the risk of delays.

A domestic accelerator that handles a substantial share of training work could also lower operating costs over time if it is easier to obtain and deploy across Chinese data centers. The business case depends on the cost of completing useful work, alongside reliable access to equipment.

Alibaba Cloud and Baidu Cloud sell computing services to businesses and public sector customers. Hardware designed and supported in China can help them compete for sensitive workloads and meet customer preferences concerning security and where data is stored and processed.

Nvidia has acknowledged the stronger competition while stating its intention to keep serving developers around the world. A company spokesperson described its position:

The competition has undeniably arrived. We will continue to work to earn the trust and support of mainstream developers everywhere.

Alibaba and Baidu are pursuing a practical mix of equipment. Nvidia systems remain central to the hardest problems, while domestic processors can take jobs where performance is comparable and the economics are attractive. The companies can expand their alternatives while continuing to use the imported hardware they already rely on.

Manufacturing and memory remain constraints

Alibaba and Baidu must do more than design capable processors. Manufacturing, packaging and memory supply all affect how many chips they can deploy and how efficiently those chips operate. The most powerful AI processors use advanced manufacturing processes and high bandwidth memory placed close to the processor through specialized packaging.

High bandwidth memory supplies large volumes of data quickly. Training large models needs both substantial memory capacity and enough bandwidth to keep processors supplied with information. A fast accelerator can lose useful performance when data cannot reach it quickly enough.

China's foundries have made progress, but they do not yet match the most advanced manufacturing processes overseas. That gap affects the highest performance chips can reach and how much electricity they need. Design choices, effective packaging and efficient software can partly compensate.

Cloud providers also adapt the way they run jobs. They can distribute work across more processors, adjust numerical precision formats and assign tasks to hardware that suits them. Changing precision means choosing how much numerical detail calculations retain, which can affect computing and memory demands.

A domestic processor integrated well with cloud software can consequently carry meaningful workloads even when its theoretical maximum performance is below that of the best foreign chips. Availability, software efficiency and the suitability of the job all contribute to what a data center can deliver.

What customers need and what comes next

Customers of Alibaba and Baidu will judge the transition through three practical measures: how quickly they can obtain capacity, whether jobs run reliably and the total cost of completing training or inference work. Domestic accelerators could shorten waiting lists if they become readily available and work consistently with familiar software frameworks.

Cloud operators are likely to use a mixture of imported and domestic equipment, with scheduling software directing jobs to the hardware best suited to them. The largest frontier models would continue to depend on Nvidia systems, while a growing range of mainstream workloads could run on Alibaba and Baidu chips.

Two developments are central over the coming year: the operation of US export licensing for accelerators such as H20 and MI308, and evidence that the domestic chips can sustain production training at scale. No dates are given for broader deployment milestones or the publication of performance results.

Benchmarks using familiar models, accounts of customer deployments and the speed at which cloud capacity using domestic accelerators sells out would provide practical measures of progress. Support in popular frameworks and tools from other developers will also affect how easily customers can move their work.

Other Chinese platforms face similar spending decisions, whether they design accelerators themselves or buy from domestic suppliers. Their allocation of capital between imported GPUs and local alternatives will help determine how quickly the wider market changes. For Alibaba and Baidu, successful deployments would validate years of chip investment and make their AI services less exposed to interruptions in overseas supply.

Key Points

• Alibaba is using its own accelerator, known internally as Zhenwu, for smaller AI training jobs.
• Baidu is testing Kunlun P800 on new versions of Ernie, supported by its control of PaddlePaddle and cloud clusters.
• Both companies retain Nvidia hardware for their largest and most complex training runs.
• US controls have tightened since 2022. Industry research indicates export licenses are required for H20 and MI308 shipments from April 2025.
• Alibaba's chip is described as competitive with H20 on some tasks, while Nvidia retains advantages for the largest training workloads.
• Software, manufacturing, packaging and high bandwidth memory remain central to domestic chips' performance and availability.
• Shares of Alibaba and Baidu rose after reports of the deployments, reflecting hopes for more dependable AI computing supplies.
• Customers will assess the shift through access to capacity, reliability and the cost of completing their work.