Cloud providers have seen a surge in demand for Nvidia’s H200 chips after Chinese AI company DeepSeek entered the competition for leading foundation models.
Although Nvidia’s stock dropped 16% on Monday, DeepSeek has attracted AI researchers’ attention since its first model, V2, in May 2024. The company’s December release of V3 significantly boosted interest. When DeepSeek R1, its reasoning model competing with OpenAI’s o1, launched in January, demand for H200 chips skyrocketed.
“The launch of DeepSeek R1 has rapidly increased H200 demand,” said Robert Brooks, founding member and VP of revenue at Lambda, a cloud provider. He noted that companies are pre-purchasing large quantities of H200 chips before their public availability.
DeepSeek’s Open-Source Models Drive Industry Disruption
DeepSeek’s models are open source, making them affordable for users. However, running them at scale requires powerful hardware or cloud services.
On Friday, analysts at Semianalysis reported that DeepSeek’s impact was tangible, affecting H100 and H200 chip pricing. Nvidia’s CFO, Colette Kress, confirmed in November that H200 GPU sales had already reached double-digit billions.
Unlike Meta, OpenAI, and Microsoft, which have poured billions into AI infrastructure, DeepSeek trained its models with weaker chips. This unexpected efficiency has worried investors, raising concerns over whether the massive AI infrastructure investments were necessary.
DeepSeek uses fewer chips than competitors, although the exact number remains debated. Despite the lighter training process, inference—using the models for real-world tasks—remains computationally demanding, cloud providers report.
H200 Chips Essential for Running DeepSeek’s Full Model
Running DeepSeek V3 in full form requires H200 chips, the only widely available Nvidia GPU that can handle the model on a single node.
Cloud firms explain that splitting the model across weaker GPUs increases complexity, risk of errors, and slows performance. “Adding that complexity almost always leads to slower speeds,” said Srivastava, an industry expert.
DeepSeek’s most advanced models contain 678 billion parameters, more than Meta’s Llama (405 billion) but fewer than ChatGPT-4 (1.76 trillion). Nvidia’s upcoming Blackwell chips will also support V3 in full, but those chips have only just started shipping.
Finding enough H200 GPUs to run DeepSeek’s models efficiently is challenging unless already allocated. Cloud provider Baseten, which optimizes AI performance, does not own GPUs but rents capacity from data centers. Its clients prioritize inference speed, critical for real-time AI applications like conversational AI.
DeepSeek’s low-cost open-source models have been a game-changer for businesses seeking high-performance AI at lower costs. With demand for H200 chips soaring, cloud providers are racing to secure enough hardware to meet growing market needs. The rapid advancements in AI are also driving new breakthroughs in computer science, shaping the future of technology.
