Monsieur Say Lays Down the Law: DeepSeek Is Great for Nvidia

Everyone’s heard of Moore’s Law of semiconductors, but the law that has truly governed the semiconductor industry since the beginning was laid down by Jean Baptiste Say more than 200 years before the first transistor: “Supply creates its own demand.” A useful product produced in increasing abundance at lower costs will find ever more uses. Transistors are everywhere and in everything only because their price has dropped by a billionfold since the 1950s.

Though people complain about the price of Nvidia’s GPUs, the real story is that Nvidia became the world’s most important company by making artificial intelligence affordable. DeepSeek’s algorithms make AI even more affordable; that will make Nvidia and Nvidia investors even richer.

For AI, processing speed is no longer the relevant limit. The slowdowns and bottlenecks come from moving data around, in and out of memory, and back and forth as the machine learns and updates. DeepSeek uses several innovations to overcome these bottlenecks.

Multi-Head Latent Attention (MLA) allows the model to access different categories of information simultaneously, rather than serially.

Mixture of Experts (MoE) divides an AI model into separate sub-networks (or “experts”), each specializing in a subset of the input data. Then the model activates only the experts it needs for a task, economizing on compute and communication.

DualPipe, by passing data forward and backward simultaneously, reduces intervals wherein GPUs remain idle while data is exchanged or synchronized. Once again, the impact is to keep data moving so as not to waste the GPUs computational power.

This all adds up to doing AI more economically, which is how Nvidia became a $3 trillion dollar company in the first place. DeepSeek’s model optimizations reduce per-model training costs, but those models still depend on high-performance GPUs. The more accessible AI is, the more it will be used, and the more GPUs NVDA will sell.

Table 1 highlights how DeepSeek’s AI optimizations affect GPU demand and benefit key industry players. Ultimately, all these innovations reinforce Nvidia’s dominance in AI hardware by making AI more accessible while still relying on GPUs. The primary beneficiaries include Nvidia, AMD, and key memory suppliers like SK Hynix and Micron.

Mr. Huang knows all about M. Say

If you have any doubts that reducing the cost of AI is to NVDA’s advantage, check in with CEO Jensen Huang. Ask him why his company developed the Blackwell-class GPUs that make AI training more efficient, i.e., more widely affordable. The B200 and B300 support FP8 and FP4 mixed-precision computing, dynamically adjusting precision levels based on workload needs. By leveraging FP4, Blackwell GPUs deliver up to five times the computational performance of previous-generation accelerators, reducing AI training costs while maintaining high efficiency. Meanwhile, the next Blackwell generation, the B300, will be optimized to take advantage of innovations used by DeepSeek, such as Mixture of Experts, in trillion-parameter models.

DeepSeek is good news for advanced memory makers such as Micron and SK Hynix. MoE requires high VRAM (video random access memory) and memory bandwidth, reinforcing demand for AI accelerators with HBM3E. NVDA’s B200 and B300 will further increase VRAM demand for trillion-parameter models.

Table 2 illustrates that DeepSeek’s innovations reduce the cost per AI model but do not decrease demand for high-performance GPUs with HBM3E. Nvidia’s Blackwell-class GPUs, such as the B200 and B300, remain essential for trillion-parameter models, reinforcing Nvidia’s AI market dominance.

DeepSeek and the ASIC makers

DeepSeek’s AI models reduce the complexity of inference tasks and thereby enable hyperscale data centers to shift some spending from GPUs toward custom AI ASICs. iGoogle’s Trillium TPU series, Amazon’s Trainium, and Broadcom’s Jericho3-AI are key competitors. Nvidia remains protected by its end-to-end ecosystem, which includes GPUs, CUDA, and TensorRT, enabling superior AI model training and inference capabilities that are already widely adopted by model developers. (Table 3)

Table 4 shows how DeepSeek’s efficiency could accelerate hyperscalers’ shift toward in-house AI ASICs, reducing reliance on Nvidia GPUs for inference. DeepSeek could lower Nvidia’s total inference market share, though Nvidia’s stronghold on training remains intact.

What Deepseek clearly has not done is reduce capital spending by NVDA’s leading customers. Microsoft, Meta, Google, and Amazon are increasing AI-related capital expenditures. Even with AI ASICs becoming more prominent, Nvidia’s GPUs remain a core component of these expansions. (Table 5)

Oh no! DeepSeek is spying on us! (Not)

Predictably, DeepSeek is inspiring another round of protectionism disguised as national security precautions, with zero evidence of any such threat, as usual. South Korea’s defense ministry has restricted DeepSeek access on military networks to prevent security breaches. Taiwan has advised government agencies to avoid the platform due to concerns over data security risks. Italy became the first European country to ban DeepSeek, removing it from app stores due to privacy violations. Australia followed with a nationwide ban on government systems, citing security threats. India blocked DeepSeek on government devices, fearing potential foreign surveillance.

In the United States, bipartisan legislation is being introduced to restrict DeepSeek’s use in federal agencies. The U.S. military, including the Department of Defense, NASA, and intelligence agencies, has also issued warnings against using DeepSeek over alleged cybersecurity risks.

This latest offensive in the trade war only continues the breakup of the beneficial global semiconductor ecosystem, which has advanced both the technology and demand for it. These restrictions could limit DeepSeek’s global expansion but are unlikely to impact Nvidia’s AI dominance.

Investor takeaway

DeepSeek’s AI innovations enhance model efficiency but reinforce the necessity of high-performance GPUs, ensuring continued demand for AI accelerators. The rising competition between GPUs and AI ASICs signals a shift in AI infrastructure, but Nvidia’s dominance remains strong due to its broad software and hardware ecosystem.

DeepSeek’s reduction of AI training and inference costs will only accelerate the AI industry’s growth, driving record investments and returns for semiconductor makers including NVDA.

George Gilder is the world’s leading technology futurist and leads the Gilder Technology Report team. GTR colleague Dr. Robert Castellano also writes at his Substack Dr. Robert Castellano’s Semiconductor Deep Dive.