DeepSeek has released an open-source reasoning model called R1, which has significantly outperformed the best models from U.S. companies such as OpenAI. This breakthrough comes at a self-reported training cost of less than $6 million, a stark contrast to the billions spent by Silicon Valley companies. Nvidia and OpenAI CEOs, Jensen Huang and Sam Altman respectively, have been focusing on a new concept in AI development known as "test-time scaling." Despite Nvidia's stock plummeting by 17% following DeepSeek's emergence, the company described R1 as "an excellent AI advancement."
The concept of test-time scaling introduces a new dimension to the "scaling law" proposed by OpenAI researchers in 2020. This law suggests that enhancing AI systems involves expanding computation and data, which necessitates more chips. Nvidia's GPUs remain integral to AI development, yet the rise of alternative models like DeepSeek's R1 has impacted its stock value. Notably, DeepSeek utilized special versions of Nvidia's GPUs designed for the Chinese market, ensuring full compliance with export controls.
Nvidia remains optimistic about DeepSeek's breakthrough, viewing it as an opportunity to generate more demand for its graphics processing units (GPUs). A spokesperson from Nvidia asserted the importance of their technology in AI development:
"Inference requires significant numbers of NVIDIA GPUs and high-performance networking," – Nvidia spokesperson
Moreover, tech giants like Microsoft and Meta are heavily investing in AI infrastructure. Microsoft plans to allocate $80 billion in 2025 alone, while Meta intends to invest between $60 to $65 billion in capital expenditures as part of its AI strategy. These substantial investments highlight the industry's commitment to advancing AI capabilities.
DeepSeek's utilization of test-time scaling exemplifies how new models can be developed with greater efficiency. BofA Securities analyst Justin Post noted the potential implications of this development:
"If model training costs prove to be significantly lower, we would expect a near-term cost benefit for advertising, travel, and other consumer app companies that use cloud AI services, while long-term hyperscaler AI-related revenues and costs would likely be lower," – BofA Securities analyst Justin Post
OpenAI's models, such as o1, also integrate forms of the test-time scaling law, demonstrating its growing influence across the sector. The emergence of DeepSeek’s R1 model has consequently raised questions about the effectiveness of multi-billion dollar capital investments in Nvidia-based AI infrastructure.