Demystifying DeepSeek - Export Controls Killed NVIDIA (Part 3/3)

In our previous posts, we debunked concerns about DeepSeek’s data privacy and the misleading claim that it was trained for just $5 million. Now, we turn to one of the most persistent myths: "Export Controls Killed NVIDIA - DeepSeek Proves China Doesn't Need U.S. Chips!“

This claim suggests that DeepSeek’s success under U.S. export restrictions proves China’s self-sufficiency in AI chips and renders NVIDIA irrelevant. However, this oversimplifies the situation. While China has adapted, DeepSeek’s achievements still rely heavily on NVIDIA hardware and software.

The Role of High-Flyer and DeepSeek's Hardware Arsenal

DeepSeek is not an isolated startup—it is backed by High-Flyer, a hedge fund that recognized AI’s potential early on. High-Flyer invested heavily in computational infrastructure, acquiring 10,000 NVIDIA A100 GPUs in 2021, forming what was then China's largest known AI computing cluster. In May 2023, DeepSeek was spun off as a dedicated AI company but continues to share High-Flyer’s resources.

According to SemiAnalysis, DeepSeek still operates with a formidable fleet of NVIDIA GPUs:

• Estimated 50,000 GPUs, including Hopper-based architectures.

• ~10,000 H800s and ~10,000 H100s, many likely acquired before stricter regulations.

• Significant orders for NVIDIA H20 GPUs, which comply with the latest U.S. export restrictions.

These figures illustrate that DeepSeek’s success is not a result of bypassing NVIDIA, but rather of working within constraints while still utilizing its technology.

Why NVIDIA Still Holds the Advantage

DeepSeek has undertaken impressive low-level engineering optimizations to make the most of its available hardware. Their ability to develop innovations like Multi-Head Latent Attention (MLA) and Mixture of Experts (MoE) shows exceptional engineering talent. DeepSeek has even optimized CUDA-level communication scheduling to compensate for the reduced interconnect bandwidth of H800 chips. These achievements demonstrate deep technical expertise.

However, this also illustrates just how much engineering effort typically comes pre-packaged within NVIDIA’s software ecosystem. Much of the work that DeepSeek has done manually—such as optimizing GPU communication and maximizing hardware efficiency—is typically handled by NVIDIA’s advanced CUDA software stack. NVIDIA’s CUDA, cuDNN, and TensorRT provide deeply integrated AI acceleration, while Nickel (NCCL) ensures seamless multi-GPU training—capabilities that have no direct Chinese equivalent. DeepSeek’s ability to push its hardware to the limit is impressive, but it also highlights how dependent AI research still is on NVIDIA's foundational technologies.

Conclusion: The Reality of DeepSeek's Success

DeepSeek’s engineering feats are real, and its efficiency gains are significant. However, the idea that DeepSeek has eliminated NVIDIA’s relevance is a myth. The company:

• Has a GPU cluster worth well over $500 million USD, demonstrating its reliance on NVIDIA hardware.

• Has adapted to export controls but still lacks access to the most cutting-edge AI chips, which limits scaling potential.

• Is deeply embedded in NVIDIA's software ecosystem, benefiting from CUDA, cuDNN, and other proprietary optimizations crucial for large-scale AI training.

Rather than proving NVIDIA’s irrelevance, DeepSeek’s innovations highlight how companies optimize within limitations—while still depending on NVIDIA's hardware and software ecosystem.

Ultimately, NVIDIA is still the clear king on the hill, years ahead of competitors. No other company currently offers the same combination of cutting-edge hardware, optimized software, and deep industry integration.