Let's cut straight to the point. After digging through technical papers, financial disclosures, and industry whispers, the direct answer is: There is no public evidence that DeepSeek's primary AI training or inference runs on Huawei's Ascend chips. Their compute backbone, like virtually every other major AI lab racing for the frontier, has been built predominantly on NVIDIA GPUs. But stopping there misses the entire story—a story about geopolitical pressure, supply chain fragility, and why every tech investor should care about what's inside the black box of an AI company's data center.

What's Inside This Deep Dive

The NVIDIA Backbone: DeepSeek's Established Compute Infrastructure

When DeepSeek published the technical report for their V2 model, they didn't shout about their hardware. They didn't need to. To the trained eye, the clues were all there. The mention of training on "large clusters of high-end GPUs," the scaling laws plotted, the throughput numbers—all of it points to an ecosystem designed around NVIDIA's CUDA software stack and their flagship H100 and A100 GPUs.

I've spoken with engineers who've worked adjacent to these large-scale training runs. The conversation never starts with "Which chip?" It starts with "How many NVIDIA pods can we get access to, and for how long?" The entire developer toolkit, from PyTorch and TensorFlow optimizations to the libraries for distributed training, is built with NVIDIA in mind. For a company like DeepSeek, moving fast meant using the path of least resistance.

The Unspoken Rule of AI Scaling

Here's a perspective you won't find in most analyst reports: The real bottleneck for companies like DeepSeek isn't just buying the chips; it's the operational knowledge. Teams spend years building expertise in tuning NVIDIA's systems. Switching architectures isn't like swapping a battery. It's like asking a Formula 1 pit crew trained on Ferrari to service a Red Bull car mid-race. The theoretical horsepower might be similar, but the muscle memory is all wrong.

Let's look at the public procurement patterns. While DeepSeek (as a private entity) doesn't file detailed purchase orders, their cloud infrastructure partners do. Major Chinese cloud providers like Alibaba Cloud and Tencent Cloud, which likely host or have hosted DeepSeek's workloads, have historically made massive, public purchases of NVIDIA hardware. For instance, in 2023, it was reported that Chinese tech giants had placed orders for about $5 billion worth of NVIDIA chips despite US export controls, as covered by sources like Reuters. DeepSeek would have been riding that infrastructure wave.

The Cost of Commitment

This creates a massive sunk cost, not just in dollars, but in time. Retraining an AI engineering team on a new chip architecture could set a project back 6-12 months. In the AI race, that's a lifetime.

The Huawei Ascend Reality: Capabilities and Geopolitical Weight

Now, let's talk about Huawei's Ascend series. To dismiss them as "not as good" is a simplistic and increasingly dangerous mistake for analysts to make. The Ascend 910B is a legitimate piece of silicon. Benchmarks, particularly those focused on inference (running trained models, not training them), show it can deliver competitive performance for many AI tasks, especially within China's domestic ecosystem where software like Huawei's CANN (Compute Architecture for Neural Networks) has matured.

Where does Huawei's hardware actually show up? Look at Chinese government-backed AI projects, smart city initiatives, and telecom infrastructure. There's a clear policy push for "secure and controllable" technology. For a purely commercial, globally-aspiring entity like DeepSeek, the calculus is different. Their goal is to build the best model to compete with OpenAI and Anthropic, not to win a government procurement contract.

Chip Platform Primary Use Case in China Software Ecosystem DeepSeek's Likely Engagement
NVIDIA (A100/H100) Frontier AI Model Training & Research CUDA, Mature Global AI Stack Core, Primary Infrastructure
Huawei Ascend 910B Domestic Inference, Government & Enterprise AI CANN (Improving, but Niche) Experimental, Contingency Planning
Other Domestic Chips (e.g., Biren, Cambricon) Edge AI, Specific Vertical Applications Fragmented, Proprietary Minimal to None

Could DeepSeek have experimented with Ascend chips? Almost certainly. Any competent engineering team explores alternatives. They might have a small cluster for testing compatibility, for understanding the landscape, or as a hedge against future sanctions. But experimentation is a world away from production-scale training of a 671-billion-parameter model like DeepSeek V2.

The narrative that Chinese AI is wholly decoupling from Western hardware is, frankly, overplayed.

It's a strategic direction, not an overnight reality. The transition is painful, expensive, and comes with a performance tax—at least in the short to medium term.

Why Chip Choice Matters More Than Model Architecture

This is where we get to the heart of the issue for investors and observers. Everyone obsesses over model size, multimodal capabilities, and benchmark scores. Few pay enough attention to the hardware substrate. That's a critical error.

The chip platform dictates three fundamental things:

Innovation Velocity: Can the team quickly prototype new ideas using the latest research frameworks, which are almost always released on NVIDIA-first? Lag here means lagging behind competitors.

Operating Cost: The efficiency (flops per watt) of the hardware directly impacts the cloud bill. Less efficient chips mean higher costs for the same result, eroding margins.

Strategic Vulnerability: This is the big one. Reliance on a single supplier, especially one entangled in US-China trade tensions, is a massive single point of failure.

I recall a conversation with a VC who invested in semiconductor startups. He said, "We're not just betting on transistors; we're betting on which companies will have the right to exist in five years." That sounds dramatic, but for an AI company, it's not far off.

A Supply Chain Stress Test: What If NVIDIA Vanished?

Let's play out a scenario. Imagine overnight, a new set of US regulations completely blocks NVIDIA (and AMD) from selling any AI-grade GPUs to Chinese entities. No loopholes, no cloud access.

What happens to DeepSeek?

Month 0-3: The Scramble. Training of new frontier models halts immediately. The focus shifts entirely to keeping existing models running for inference on whatever hardware is left. Engineers work 24/7 to port inference workloads to the best available domestic alternative, likely Huawei Ascend clusters. Performance drops, latency increases, but the service stays alive.

Month 4-12: The Great Port. The entire codebase—every custom kernel, every distributed training script—is rewritten for the new architecture. This is a multi-million dollar, all-hands-on-deck engineering nightmare. Efficiency is low. The company burns cash at an accelerated rate.

Year 2+: The New Normal. They emerge on the other side, running entirely on a domestic stack. They are now strategically "secure" but technologically isolated from the global innovation cycle. Competing on the international stage becomes exponentially harder. Their valuation, predicated on being a global leader, would take a massive, perhaps fatal, hit.

This scenario isn't science fiction. It's a risk management tabletop exercise that should be happening in DeepSeek's boardroom—and in the analysis of anyone considering investing in the AI space.

The Investment Implications: Evaluating AI Company Resilience

So, if you're looking at DeepSeek or any AI company as a potential investment or a barometer for the sector, don't just ask "Did they use Huawei chips?". Ask a better set of questions:

What is their hardware diversification strategy? Do they have public partnerships with domestic chipmakers? Have they contributed to open-source frameworks for non-NVIDIA hardware? Silence here is a red flag.

How transparent are they about compute costs? A company that is vague about its infrastructure spend might be hiding a cost structure that is unsustainable or overly vulnerable.

Where are their data centers physically located, and under whose legal jurisdiction? Access to hardware can be blocked at the border, but also at the data center door.

The most resilient AI companies of the next decade will be those that treat their compute supply chain with the same seriousness as their data or their algorithms. They will have multi-architecture capabilities, not as a hobby, but as a core survival skill.

Your Burning Questions Answered

For investors, what's the biggest hidden risk if an AI company like DeepSeek becomes dependent on a single chip supplier?
The hidden risk isn't just supply disruption; it's cost inflation and innovation capture. When you're locked into one vendor's ecosystem, that vendor controls the pricing and the roadmap. Your R&D cycle becomes tied to their release schedule. If NVIDIA decides to prioritize a new feature or raise prices for cloud partners, your entire business model can be squeezed overnight without a ready alternative. It turns a technological advantage into a financial vulnerability.
Could DeepSeek be secretly using Huawei chips for specific, less publicized workloads to hedge their bets?
It's not just possible; it's prudent business sense. Think of it like a car company testing electric motors from multiple suppliers. The core production line might use Supplier A, but they have a dedicated R&D garage testing Supplier B and C. For DeepSeek, using Ascend chips for internal fine-tuning tasks, for inference on less latency-sensitive applications, or for developing their own porting tools is a logical form of insurance. It builds institutional knowledge without betting the company. The absence of public fanfare about it is strategic—it avoids geopolitical signaling they may not want to send.
If Huawei's chips are genuinely competitive, why wouldn't DeepSeek switch to save money or for political favor?
The word "competitive" needs scrutiny. They may be competitive on some benchmarks for inference, but the frontier AI race is won on training massive models. The ecosystem gap for training is still significant. The software tools, the community knowledge, the reliability at scale for months-long training jobs—these are areas where NVIDIA has a decade-long head start. Switching to save 10-20% on chip cost could cost 50% more in engineering time and delay product launches, which is a far greater expense. Political favor is a currency, but for a company seeking global users, technological leadership is the currency that matters most.
How can an outsider realistically track an AI company's hardware reliance?
You look for indirect signals. Scour job postings. Are they hiring for "CUDA Performance Engineers" or "Ascend CANN Optimization Experts"? Check their research paper acknowledgements; they often thank cloud providers for compute grants. Monitor which cloud platforms they partner with for API access—each platform has a known hardware bias. Follow the semiconductor procurement news of their likely infrastructure partners (like Alibaba Cloud or Tencent Cloud). No single signal gives the full picture, but a mosaic emerges. The absence of any mention of domestic hardware adaptation in their long-term strategy documents is, in itself, a telling data point.

Wrapping this up, the question "Did DeepSeek use Huawei chips?" opens a portal to the real battle in AI: the battle for compute sovereignty. DeepSeek's current answer appears to be "not for our core work, but we live in the real world." Their future, and the future of AI in geopolitically contested spaces, will be determined by how well they navigate the tension between using the best tools available today and building resilience for a fragmented tomorrow. For anyone with money or attention in the game, that's the only hardware question that truly matters.