Let's be real. If you're searching for which Chinese companies are using Nvidia chips, you're not just looking for a simple list. You're trying to understand the real landscape—who's genuinely dependent, who's just name-dropping for investors, and how the whole scene is shifting under the pressure of US export controls. Having spent years tracking the semiconductor and AI hardware supply chain in Asia, I can tell you the picture is far messier and more interesting than most headlines suggest. It's a mix of public partnerships, deep-seated dependencies in R&D labs, and a growing gray market for last-generation chips.
What You'll Find Inside
The Landscape: Who's Publicly Using Nvidia?
We can break this down by sector. The most visible users are the giants with the budgets and the public AI ambitions. But even here, the "usage" varies wildly.
Internet & Cloud Titans
These companies built their AI infrastructure in the pre-ban era, and Nvidia GPUs are deeply embedded in their data centers.
Baidu is a classic example. Their PaddlePaddle AI framework was optimized for Nvidia hardware for years. While they promote their own Kunlun AI chips now, go talk to any engineer there—they'll tell you a significant portion of their most demanding model training still happens on A100 and H100 clusters procured before the restrictions hit. It's the workhorse they know best.
Alibaba Cloud and Tecent Cloud offered GPU-accelerated computing instances powered by Nvidia V100 and A10 chips as a core service. Post-ban, they've had to pivot hard to promoting alternative hardware, but existing customer workloads on those older Nvidia instances are still running. The dependency is in the installed base.
Here's a subtle point most miss: for these cloud providers, the real pain isn't just about training their own models. It's about losing a key demand driver from their customers. Startups and enterprises would flock to their cloud services specifically for access to Nvidia GPUs. That attraction is now dimmed.
Autonomous Driving & EV Frontrunners
This is where the dependency feels almost absolute. Nvidia's DRIVE platform, especially the Orin system-on-a-chip (SoC), became the de facto brain for advanced driver-assistance systems (ADAS).
NIO, XPeng, and Li Auto all have vehicles on the road or in development using Nvidia Orin. NIO's Adam supercomputer, which powers its autonomous driving, is based on Nvidia hardware. The design cycle for a car is 3-5 years. These companies committed to Orin years ago, and untangling from that now is a nightmare of re-engineering and re-validation. I've spoken to sourcing managers in this space who describe a frantic search for every last Orin chip they can legally acquire, often through complex multi-tier distributors.
They're publicly exploring Chinese alternatives like Horizon Robotics for future models, but the current generation is locked in.
AI & Computing Startups
Companies like SenseTime, Megvii, and CloudWalk grew up on Nvidia. Their entire algorithm research and development pipelines were built on CUDA, Nvidia's proprietary programming model. Shifting away from that isn't like changing a lightbulb; it's like asking a writer to switch from their native language.
Their public statements now emphasize "full-stack" capabilities and domestic hardware partnerships. But dig into their research papers or job postings for AI researchers, and CUDA experience is still frequently listed as a preferred qualification. The institutional knowledge is a form of dependency that's harder to quantify but very real.
How are Chinese companies using Nvidia chips?
It's not one thing. The application dictates the chip and the level of urgency to find a replacement.
AI Model Training (The Biggest Hunger): This is where the banned A100 and H100 chips were kings. Training massive foundation models requires thousands of these high-end GPUs working in parallel. Companies that secured clusters before late 2022 have a significant, but dwindling, competitive advantage. They're using them to train their next-generation models while trying to parallelize workloads on alternative chips, which is an immense software challenge.
AI Inference & Cloud Services: This involves running already-trained models. Chips like the A10, A30, or even older T4 GPUs are (or were) widely used here. The performance bar is different, and domestic alternatives from Huawei or others are making faster inroads in this segment because the software porting can be somewhat easier.
Edge & Embedded Devices: The Nvidia Orin in cars or Jetson modules in robots are perfect examples. The chip is part of a finalized product design. Replacing it means redoing hardware design, software, safety certification—a multi-year, multi-million dollar project.
Imagine a robotics startup that designed its flagship product around the Jetson AGX Orin. Their entire codebase, power management, and thermal design are centered on it. The US ban doesn't just block a component; it potentially blocks the entire product line unless they can find an identical drop-in replacement, which doesn't exist.
The Gray Areas and Unseen Users
This is the part you won't find in press releases.
First, there's the third-party pathway. Large Chinese companies with international subsidiaries or joint ventures sometimes route purchases through those entities. It's a legal gray zone that's tightening by the month. I know of at least one major AI lab that has a team in Singapore specifically to manage and operate a cluster of H100s, with researchers accessing it remotely.
Second, the data center loophole. US restrictions target chips above a certain compute threshold. There's a market for slightly downgraded versions (like the A800 and H800, which were later also banned) and for amassing huge numbers of older, unrestricted chips like the V100. Some companies are building "GPU farms" out of thousands of these older cards. The performance per watt is terrible, and the electricity bills are astronomical, but it keeps research moving.
Finally, the secondary and reclaimed market. Used Nvidia GPUs from decommissioned mining rigs or older enterprise servers are finding their way into smaller Chinese AI labs and universities. The reliability is a gamble, but the price is right, and they still run CUDA.
How can you spot these less visible users? Look for job ads seeking engineers with specific Nvidia tech stack experience (like CUDA, TensorRT), or scan research paper acknowledgements that thank a company for "computational resources"—sometimes the hardware details slip through.
What are the alternatives to Nvidia in China?
The push for domestic silicon is real, but it's a marathon, not a sprint. Here’s the lay of the land from my perspective.
Huawei Ascend is the 800-pound gorilla. Their Ascend 910B chip is the most credible alternative for AI training in China right now. It's being pushed heavily into cloud services (like Huawei Cloud) and is getting design wins in some government and state-owned enterprise projects. The catch? Its software ecosystem, CANN, is still playing catch-up to CUDA's maturity and breadth. Porting complex models takes time and expertise. It's good, but it's not a seamless swap.
Cambricon and Iluvatar are other players. They have capable inference chips and are making progress on training. Their challenge is scaling and building a robust software community. Developers won't flock to a platform unless the tools are great and the hardware is widely available.
Startups like Biren Technology and Moore Threads showed promise but have been hammered by US sanctions themselves, limiting their access to advanced chip manufacturing. Their long-term viability is under a cloud.
Let's be blunt: for the cutting-edge work on 100-billion-parameter models, there is no true Chinese equivalent to the H100 yet. The gap is in raw compute, memory bandwidth, and—critically—the interconnects that link thousands of chips together. Chinese companies are forced to innovate on system architecture and software to compensate for the hardware gap.
Navigating the New Reality: A Practical Perspective
So, what does this mean if you're a procurement manager for a Chinese tech firm, or an investor trying to gauge risk?
For existing projects on Nvidia hardware, the strategy is conservation and stretching. Companies are optimizing algorithms to be more compute-efficient, using more model compression techniques, and prioritizing which research tracks get the precious remaining GPU cycles.
For new projects, the default is no longer Nvidia. The design starts are happening on Huawei Ascend or other domestic platforms. This is accelerating the ecosystem development through sheer necessity. The quality of the software tools and developer support is now a key competitive battleground for Chinese chipmakers.
The smart companies are building hardware-agnostic software layers. Instead of coding directly in CUDA, they're using higher-level frameworks and investing in compiler technology that can, in theory, target multiple backends (Nvidia, Huawei, Cambricon). This is painful upfront work but provides crucial long-term flexibility.
My advice? Watch the talent flow. Where are the engineers from Nvidia's former China business going? Which domestic chipmaker is attracting the best systems software architects? That will tell you more about the future landscape than any official roadmap.
