Broker Check
Is DeepSeek Already a Generation Behind?

Is DeepSeek Already a Generation Behind?

March 27, 2025

DeepSeek, once seen as a formidable AI contender, is now facing a serious hurdle: outdated hardware. While its flagship model, DeepSeek-R1, has made waves in the AI community, its reliance on NVIDIA’s H800 chips—already a step behind the industry’s best—could leave it struggling to keep pace. With U.S. restrictions further limiting access to cutting-edge chips, DeepSeek may already be a full generation behind in the AI race.

DeepSeek's Hardware Dilemma

DeepSeek-R1 was trained on NVIDIA’s H800 GPUs, a workaround chip designed specifically to comply with U.S. export controls. While effective at the time, the H800 was never on par with NVIDIA’s best offerings. Western AI firms were already working with the more powerful H100, and now they’ve moved even further ahead with the H200 and NVIDIA’s next-gen Blackwell architecture. Meanwhile, DeepSeek is left relying on a dwindling supply of H800s, unable to access the latest hardware that is driving AI breakthroughs.

Gregory C. Allen, an AI policy expert at CSIS, summed up the situation in his March 7, 2025, report: “China’s ability to leverage pre-ban hardware has given it a temporary edge.” But temporary is the key word. DeepSeek’s ability to innovate is now limited by its aging infrastructure, and AI does not wait for anyone.

Stockpiling Wasn’t a Long-Term Solution

DeepSeek saw the writing on the wall and stockpiled thousands of NVIDIA A100 and H800 GPUs before restrictions fully took effect. But AI training demands ever-increasing computing power, and a warehouse full of last-gen chips is not the same as having access to the newest and most efficient hardware. DeepSeek’s models, while competitive today, are fundamentally limited by the tools used to build them.

Allen’s report highlights this growing disadvantage: “The pace of AI innovation is outstripping our ability to govern it effectively.” That applies to DeepSeek itself. While it may still produce strong AI models, the lack of cutting-edge GPUs means they will take longer to train, be more expensive to run, and ultimately struggle to match the efficiency and scale of Western AI competitors.

Performance Limitations

DeepSeek R1 has shown strong performance in benchmarks like coding and math, but it reportedly operates with 141 billion parameters—a fraction of the estimated 1 trillion-plus parameters in models like GPT-4 and future iterations from OpenAI and xAI.

More importantly, next-gen AI models require next-gen chips. DeepSeek is already hitting hardware-imposed limitations, and as AI models grow larger and more complex, those constraints will become more pronounced. Without access to higher memory bandwidth and greater compute density, DeepSeek’s models will simply fall behind in speed, efficiency, and overall capability.

AI Investment Remains Strong

DeepSeek’s successe scared the market, now their struggles may buoy the market.  It appears that the broader AI industry remains on solid footing. If there were any lingering fears that AI capital investment would slow, they may be put to rest. The world’s largest tech companies—Microsoft, Google, Amazon, and Meta—have all reaffirmed their long-term AI spending plans. NVIDIA, at the center of this AI boom, continues to secure massive orders from cloud providers, autonomous vehicle developers, and enterprise AI solutions, reinforcing confidence that investment in AI infrastructure isn’t slowing down.

Even BlackRock and Microsoft’s recent $30 billion AI infrastructure initiative signals that the demand for high-performance computing will only grow. AI spending is becoming a long-term strategic priority, not just for tech companies but for entire industries ranging from healthcare to finance.

At this point, the only force that could derail AI investment would be a severe economic downturn. But absent a global recession, the momentum behind AI capital spending looks unstoppable. The technology is still in its early innings, and with each new breakthrough, the appetite for faster chips, more efficient models, and greater computational power only intensifies. The future of AI remains bright, and for companies that can secure the right hardware and talent, the best is yet to come.