Nvidia CEO praises DeepSeek’s impact on open-source AI

Nvidia CEO Jensen Huang highlighted the impact of Chinese startup DeepSeek on the open-source AI movement during his keynote at CES in Las Vegas on January 5.

Huang said DeepSeek’s models, including the R1, have contributed to rapid growth in open-source AI and narrowed the performance gap between open and top proprietary models.

He cited China-developed models DeepSeek-V3.2, Kimi K2 from Moonshot, and Qwen from Alibaba as examples of this trend.

Nvidia, meanwhile, introduced its next-generation Rubin hardware platform, which is expected to launch later this year and promises 5x the AI computing power of previous products.

Huang noted continued strong demand for Nvidia’s H200 chips in China, despite ongoing US export controls restricting sales of its most advanced chips to the Chinese market.

🔗 Source: South China Morning Post

🧠 Food for thought

Implications, context, and why it matters.

DeepSeek performance claims need independent, enterprise-grade checks

Huang says DeepSeek R1, DeepSeek‑V3.2, Kimi K2, and Qwen are closing the gap with paid models. Many matchups use vendor numbers or summaries, not like‑for‑like tests on real enterprise work. 1 2 3
DeepSeek‑R1 can match GPT‑4 on some math and coding tests. It has a 128K‑token context window (a token is a small chunk of text; the context window is how much text a model can consider at once). Its open‑source license changes support and lock‑in choices for large buyers. 2 1
DocsBot pegs GPT‑4 at about 32.8x the per‑token price of DeepSeek‑R1. DeepSeek‑R1 does not handle audio or fine‑tuning (fine‑tuning is tailoring a model to a specific dataset), which curbs multimodal use cases (combining inputs like text, images, and audio). 2 1
Open‑source buzz does not equal production use. Enterprise IT and compliance teams still need strong security, clear frameworks, or support contracts that paid models often bundle. Many open tools require teams to build these in‑house.

Data center operators can win Rubin‑era cooling retrofit budgets

NVIDIA’s Vera Rubin NVL72 packs 72 GPUs (an NVL72 is a Rubin‑series server configuration with 72 GPUs). Rubin‑class racks can hit about 400 kW. Air cooling tops out near 30 kW, so liquid cooling is required in many sites (circulating chilled liquid to remove heat). A rack here refers to a cabinet that houses servers and networking gear. 4 5
Liquid cooling runs about $55,000 per rack. It can lift efficiency by roughly 30% by pushing Power Usage Effectiveness (PUE) below about 1.1 versus near 1.5 for air. That creates a solid return on investment (ROI) for infrastructure providers. 5
Microchannel liquid plates can lift heat transfer by about 30% (high‑surface‑area cold plates). They bond to GPU dies (the silicon chips inside GPU packages). Materials cost 3 to 5 times more, which opens room for cooling makers or system integrators (firms that assemble and deliver complete systems). 5
New Chinese data‑center rules set PUE at 1.3 or lower. That level often needs liquid cooling for next‑gen builds. Retrofits will rise across Asia‑Pacific colocation markets (data centers that rent space, power, and cooling to tenants). 5