Let's cut to the chase. If you're searching for what type of GPUs DeepSeek uses, you've probably hit a wall of vague statements and marketing fluff. You want specifics. You want to know the actual chips, the reasoning behind the choice, and what it means for the model's capabilities and limitations. After piecing together information from technical disclosures, industry whispers, and the stark realities of large language model training in 2024, the core answer is this: DeepSeek primarily relies on clusters of NVIDIA H800 GPUs for its most intensive training workloads.

But that's just the headline. The real story is in the why and the how. Why the H800 over other options? How many are we talking about? What does this tell us about DeepSeek's strategy and the brutal economics of modern AI? This isn't just a spec sheet review. It's a look under the hood of one of the world's most advanced AI systems.

What's Inside This Deep Dive?

DeepSeek's Primary GPU: The NVIDIA H800 Workhorse

The NVIDIA H800 isn't just a piece of silicon for DeepSeek; it's the fundamental building block of its computational brain. This isn't a guess. It's the logical conclusion drawn from a combination of factors: China's semiconductor import regulations, public procurement data, and the performance requirements for training a model of DeepSeek's scale.

The H800 is essentially the export-compliant version of NVIDIA's flagship H100 GPU, designed specifically for the Chinese market to adhere to U.S. government export controls on compute performance. It features slightly reduced interconnect bandwidth (NVLink) compared to the global H100. Think of it as a supercar with a governed top speed—still incredibly powerful, but with one specific limiter in place for geopolitical reasons.

The Key Detail Most Miss: Many assume DeepSeek would use whatever is "best." But in AI hardware, "best" is a function of availability, cost, and software stability. The H800 represented the optimal point where raw compute power met the reality of what could be legally and reliably acquired in volume within China during DeepSeek's critical scaling phase.

Here’s a breakdown of what the H800 brings to the table for DeepSeek:

Feature Specification (H800) Why It Matters for DeepSeek
FP8 Tensor Core Performance ~1,979 TFLOPS This is the engine for mixed-precision training. It allows the model to train faster while managing memory, crucial for iterating on massive datasets.
High-Bandwidth Memory (HBM) 80GB HBM3 Fits larger model parameters and training batches into each GPU. More memory means less time spent shuffling data, leading to higher efficiency.
Interconnect (NVLink) ~400 GB/s per link (reduced from H100) Even reduced, this is still blisteringly fast. It's the glue that allows thousands of GPUs to act as one giant computer. Communication speed between chips is often the real bottleneck, not the chips themselves.
Form Factor Typically SXM (Server Form Factor) Designed for dense data center deployment, not consumer cards. This means they're built for 24/7 operation and efficient cooling in massive clusters.

I've seen discussions where people conflate inference hardware with training hardware. It's a critical distinction. While the H800 clusters do the heavy lifting of training the models, the inference side (serving answers to users) likely uses a more varied and cost-optimized mix. This could include older A100s, or even custom inference chips for specific tasks. The training cluster is the R&D lab; the inference infrastructure is the factory floor.

Why H800 and Not H100? A Strategic Deep Dive

This is where it gets interesting. On paper, the global H100 is more powerful. So why commit to the H800? The answer isn't about preference; it's about necessity and strategy.

1. Geopolitical and Supply Chain Reality: During the period when DeepSeek-V2 and its predecessors were being scaled, acquiring a large, stable supply of H100 GPUs in China was not just difficult—it was practically impossible due to export restrictions. The H800 was the highest-performance tool legally available for large-scale deployment. Choosing it wasn't an alternative; for a Chinese AI lab aiming for the top tier, it was the only viable main option.

2. Software Ecosystem Lock-in: NVIDIA's CUDA platform is the de facto operating system for AI. Every line of DeepSeek's training code, its custom kernels, and its optimization libraries are built on CUDA. Switching to a different architecture (like AMD or a domestic alternative) would require a monumental, years-long software rewrite. The cost of that transition in lost development time would be catastrophic. The H800, being CUDA-based, provided continuity.

3. The Interconnect Bottleneck Myth: Yes, the H800's NVLink is slower than the H100's. But here's a nuanced point: for many workloads in a distributed training setup, the network between servers (InfiniBand or Ethernet) becomes the limiting factor long before the NVLink speed within a server does. DeepSeek's engineers would have optimized their model parallelism strategy around the H800's specific profile, making the raw NVLink difference less of a decisive factor than outsiders assume.

One perspective I don't see enough: this choice forced a unique optimization path. Facing a known hardware constraint (the interconnect), DeepSeek's team had to become exceptionally clever with their model architecture and training parallelism. This constraint might have directly influenced decisions that led to innovations like the MoE (Mixture of Experts) design in DeepSeek-V2, which is inherently more efficient in certain distributed scenarios.

How Many GPUs Does DeepSeek Actually Use?

Nobody outside DeepSeek's operations team knows the exact number, and it fluctuates daily. But we can make an educated estimate that reveals the staggering scale.

Training a state-of-the-art LLM like DeepSeek-V2 is not a one-and-done job. It involves continuous pre-training, fine-tuning, reinforcement learning from human feedback (RLHF), and ablation studies (testing variations). Each phase requires a massive cluster.

  • Industry Benchmark: We know that training models like GPT-4 or Google's Gemini Ultra required tens of thousands of H100-class GPUs running for months.
  • Cluster Sizing Logic: A single H800 server node typically holds 8 GPUs. A meaningful training cluster for a model with hundreds of billions of parameters would start in the hundreds of nodes. A conservative, low-end estimate for a major training run would be 2,000 to 5,000 H800 GPUs.
  • The Total Fleet: This is the key. DeepSeek isn't running one cluster. They likely have multiple clusters for different purposes: a massive one for flagship model training, smaller ones for research experiments, fine-tuning, and inference staging. The total inventory could easily be 2-3 times the size of a single training cluster.

Let's talk money for a second. An H800 GPU costs significantly more than a high-end car. A cluster of a few thousand represents a capital expenditure in the hundreds of millions of dollars. This isn't just a technical detail; it's the single largest barrier to entry in the modern AI race. When you ask "What GPUs does DeepSeek use?" you're really asking, "What is the multi-hundred-million-dollar foundation of their company?" The answer is rows upon rows of H800 servers.

Beyond the Chip: The Full Infrastructure Picture

Focusing solely on the GPU model is like describing a Formula 1 car only by its engine displacement. The supporting infrastructure is what allows the GPUs to perform.

The Unsung Heroes:

Networking: You could have 10,000 H800s, but if they can't talk to each other fast enough, they're useless. DeepSeek almost certainly uses NVIDIA's InfiniBand networking (like the Quantum-2 platform) to create a low-latency, high-bandwidth fabric. This network is the nervous system of the cluster.

Storage: Training data is measured in petabytes. The storage system must feed data to the GPUs at an unimaginable rate to keep them busy. This means all-flash storage arrays or custom solutions built for massive sequential reads.

Cooling: A dense H800 cluster puts out heat like a small furnace. Advanced liquid cooling (direct-to-chip or immersion cooling) is likely in use. This isn't about comfort; it's about preventing thermal throttling that would slash performance and hardware lifespan.

Power: A single server rack with 8 H800s can draw over 10 kilowatts. A full-scale data center for DeepSeek requires a dedicated power substation. The electricity bill alone is a major operational cost.

How Does This Stack Up Against Other AI Giants?

Putting DeepSeek's choice in context helps us understand the global AI hardware landscape.

Company / Lab Primary Training Hardware Key Differentiator / Constraint
OpenAI NVIDIA H100 (and likely H200/B100 soon) Early and massive investment, direct partnership with NVIDIA and Microsoft Azure. Faces fewer procurement restrictions.
Google DeepMind Custom TPU v5e/v5p Vertical integration. They design the chip and the software stack together for optimal efficiency, avoiding NVIDIA's ecosystem costs.
Meta (FAIR) NVIDIA H100, moving to custom MTIA chips Massive scale for both research and production. Investing heavily to reduce long-term dependence on NVIDIA.
DeepSeek NVIDIA H800 Operates under export restrictions. The H800 is the highest-performance attainable tool, leading to unique optimization challenges and strategies.
Anthropic / xAI NVIDIA H100 (via Cloud Providers) Reliant on cloud hyperscalers (AWS, Google Cloud, Oracle). More flexible but at a higher operational cost per FLOP.

DeepSeek's position is unique. Unlike Google, it doesn't control its silicon fate. Unlike OpenAI and Meta, it couldn't access the unrestricted H100. This placed it in a box. Yet, its performance shows that within that box, through exceptional software and model architecture work, it achieved world-class results. It's a testament to the fact that while hardware is the canvas, the algorithm is the painting.

The Future: What's Next for DeepSeek's Hardware?

The H800 is today's answer, but the landscape is shifting rapidly.

1. The Next-Generation Export Chip: NVIDIA has already announced the H20, the successor to the H800 for the China market. It's a given that DeepSeek will evaluate and likely adopt these as they become available to stay competitive.

2. Domestic Chinese GPUs: Companies like Biren Technology and Moore Threads are developing alternatives. The performance gap is still significant, but the pressure for technological self-sufficiency is immense. I expect DeepSeek to run pilot projects and research collaborations with domestic chipmakers, but a full-scale switch for core training remains a distant, high-risk prospect due to the software ecosystem hurdle.

3. Specialized Inference Hardware: This is a more likely near-term diversification. Designing or using custom ASICs for serving inference (answering user queries) can drastically reduce cost and energy use. DeepSeek may already be experimenting with this for its public API.

The biggest trend isn't a new chip, but a new approach: algorithmic efficiency. The real breakthrough for DeepSeek will be making a model smarter with fewer computations. The focus will increasingly shift from "How many more H800s can we buy?" to "How can we make our training 10x more efficient?" This is where the real competitive battle is moving.

Your Burning Questions Answered

Does DeepSeek use AMD or other non-NVIDIA GPUs for any part of its work?

It's highly unlikely for their core model training. The software stack—CUDA, cuDNN, NCCL—is so deeply entrenched that switching even a portion of the workflow would create a compatibility nightmare. The engineering overhead would outweigh any potential cost benefit. They might use CPUs or other accelerators for peripheral tasks like data preprocessing, but the critical path of neural network training runs on NVIDIA silicon.

Is DeepSeek locked into NVIDIA's ecosystem, and is that a risk?

It's a valid concern. They are profoundly locked in, as is almost every major AI lab outside of Google. The risk is twofold: cost (NVIDIA commands high margins) and supply chain fragility (geopolitics can disrupt access). The mitigation is long-term investment in software abstraction layers (like OpenAI's Triton) that could, in theory, allow code to run on different hardware. But decoupling from CUDA is a decade-long project, not a quick fix.

How does the GPU choice affect the quality or "intelligence" of the final DeepSeek model?

It doesn't directly affect the potential intelligence ceiling. A brilliant algorithm trained on sufficient data will produce a brilliant model, whether it runs on H800s, H100s, or TPUs. Where the hardware matters is in the pace of innovation. Faster hardware (or more of it) allows for more experiments, larger-scale training runs, and quicker iteration. The H800 constraint may have slowed DeepSeek's experimental velocity compared to a lab with unrestricted H100s, forcing them to be more selective and clever with their research direction—a constraint that may have inadvertently led to more elegant solutions.

Could DeepSeek have built its model using cloud GPUs instead of buying its own?

For the scale they operate at, owning (or long-term leasing) the infrastructure is almost certainly more economical. Cloud GPU costs are astronomical for continuous, large-scale training. The break-even point for building your own data center is reached very quickly when you're consuming compute at the petascale. Cloud is great for flexibility and startups, but for a sustained, top-tier research effort, capex beats opex.

What's the single biggest challenge with using H800 clusters at this scale?

Reliability. When you have thousands of GPUs running flat-out for months, hardware failures are a constant, not an exception. A single GPU failing can crash a training job that has already run for weeks. The engineering challenge isn't just writing the training code; it's building the orchestration, checkpointing, and fault-tolerance systems to keep this fragile, massively parallel computation alive long enough to complete. The operational complexity is mind-boggling.

So, what type of GPUs does DeepSeek use? They use the GPUs they could get, the GPUs they could afford, and the GPUs around which they built a world-class software fortress. The NVIDIA H800 is more than a component; it's a symbol of the current era of AI—an era defined by astronomical compute costs, geopolitical tensions over technology, and the relentless pursuit of efficiency within constraints. Understanding this hardware foundation is the first step to understanding DeepSeek's position, its strategy, and the immense challenges it has overcome to stand among the leaders in artificial intelligence.