Rudolph Angler Cloud providers are seeing a surge in demand for Nvidia’s H200 chips after Chinese AI firm DeepSeek entered the foundation model race. The stock market reacted strongly on Monday, sending Nvidia’s shares down 16% as investors processed the impact.
DeepSeek first gained attention in the AI research community after launching its V2 model in May 2024. However, the release of V3 in December turned heads, showcasing impressive efficiency. In January, DeepSeek introduced R1, its reasoning model, which competes directly with OpenAI’s o1.
Since then, demand for H200 GPUs has skyrocketed. “Enterprises are now pre-purchasing large blocks of Lambda’s H200 capacity before public availability,” said Robert Brooks, vice president of revenue at cloud provider Lambda.
Open-Source Model Spurs Market Disruptions
DeepSeek’s models are open source, allowing users to access them at minimal cost. However, running these models at scale still requires powerful hardware or cloud computing services.
On Friday, semiconductor analysts at Semianalysis reported “tangible effects” on H100 and H200 pricing due to DeepSeek’s increasing popularity. According to Nvidia CFO Colette Kress, total sales of H200 GPUs have already surpassed double-digit billions as of the November earnings call.
DeepSeek’s approach differs from Western AI companies. The company trained its models on less powerful hardware, as noted in its research paper. This efficiency has rattled investors, who now question the necessity of massive AI infrastructure investments by Meta, OpenAI, and Microsoft.
While DeepSeek’s training process used fewer, weaker chips, the inference process remains compute-intensive, cloud providers confirm. “It is not light and easy to run,” said Srivastava, an industry expert. Many firms avoid Meta’s 405 billion parameter Llama model due to its high processing demands.
DeepSeek provides smaller versions, but even its most powerful model remains cheaper to operate than competing alternatives. This cost advantage has excited firms looking to deploy full-scale AI models without excessive computational costs.
H200 Becomes Critical for Running DeepSeek Models
Nvidia’s H200 chips are currently the only widely available option capable of running DeepSeek’s V3 model in full form on a single node (8 interconnected chips).
Alternative approaches, such as spreading workloads across weaker GPUs, require specialized expertise and introduce performance risks. This complexity slows down operations, noted Srivastava.
The most powerful DeepSeek model contains 678 billion parameters, more than Meta’s Llama model (405 billion) but fewer than OpenAI’s GPT-4 (1.76 trillion). Nvidia’s next-gen Blackwell chips, expected to fully support V3, have only just started shipping.
The spike in demand for H200 GPUs has made securing high-speed compute capacity difficult unless pre-allocated.
Baseten, a cloud infrastructure company, doesn’t own GPUs but optimizes performance for firms using their own hardware. Its clients prioritize inference speed, essential for real-time AI applications like voice conversations.
DeepSeek’s cost-efficient, open-source AI models are proving to be a game-changer in the industry, forcing cloud providers and enterprises to rethink their AI strategies.