NVIDIA Rubin AI server racks consume 362kW per rack; market concerns about power/cooling feasibility and data center OPEX.
Nvidia shares dropped 3.6% to around $201.10 late Tuesday, pushing the chipmaker's valuation under $5 trillion. A product detail from Monday is likely more important for investors, though. Rolling out Nvidia's next systems could hinge not just on processor demand but on power and cooling requirements.
362 kilowatts is the critical figure here. Super Micro says its Nvidia Vera Rubin NVL4 liquid-cooled rack operates at that power level. Eight racks draw 3.2 megawatts, and the system can scale to a gigawatt. "The institutions that accelerate infrastructure deployment will lead the next generation of breakthroughs," said Super Micro CEO Charles Liang.
Nvidia can keep making faster chips, but customers cannot use them for AI profits until data centers have the power, cooling, and finished spaces ready. Morgan Stanley has warned developers are looking at possible power crunches in 2027 and 2028, blaming underinvestment in grids and supply chain strain.
Nvidia remains "the most important company in the AI buildout," according to D.A. Davidson's Gil Luria, but investors "see more leverage" in other names as new bottlenecks emerge. Nvidia has gained 14% this year through Monday, placing it at the bottom of the 30-member PHLX Semiconductor Index, which more than doubled over the same stretch. Broadcom was second from last.
Super Micro shares soared 15.7% Monday after the company unveiled its end-to-end Rubin system, covering power distribution, liquid cooling, site preparation, installation, and testing. Investors responded positively to Super Micro taking Nvidia hardware and delivering a complete working data center.
Nvidia announced that Bull, Dell Technologies, Gigabyte, Hewlett Packard Enterprise, and Super Micro will build Rubin systems, which can house up to 144 graphics processors per rack. Nvidia expects the systems to hit the market in the fourth quarter. CEO Jensen Huang called Vera Rubin "a new instrument for science."
Nvidia said 35 AI supercomputers using its chips are being built in 23 European countries, aiming to give more than 3 million researchers access. The company is extending its reach from U.S. cloud giants to include government science and national computing projects.
Nvidia also rolled out the BioNeMo Agent Toolkit on Tuesday, targeting life-sciences research. CEO Jensen Huang called it "the scientific toolbox" for AI agents. Over 50 companies are already using the new technology.
Chip stocks took a hit amid a tough market. AMD dropped about 5.7%, Broadcom lost 2.5%. The Philadelphia semiconductor index sank over 7% as traders grew wary of heavy debt-fueled AI bets and priced in higher rates. "Recent AI developments have brought up a lot of questions about spending and ramping of semiconductor capacity," said Thomas Martin, senior portfolio manager at Globalt.
AI stocks are "highly concentrated and flow-driven," according to Ross Mayfield, investment strategy analyst at Baird, leaving the trade exposed to quick shifts in sentiment. Another wave of product launches did not shield Nvidia. At Nvidia's scale, Mayfield said, technical wins need to turn into quicker cash flow, not just bigger performance figures.
The risks are substantial. Power hookups could be delayed. Facility upgrades might cost more than expected. Cooling equipment could be in short supply, stalling customer installations even when demand for chips remains steady. Josh Parker, Nvidia's head of sustainability, said its warm-water cooling system could cut cooling-tower water use from about 2.6 million gallons per megawatt a year to "near zero." Nvidia did not share construction costs. The cooling design also does not address electricity generation needs.
Nvidia's annual shareholder meeting begins Wednesday at 9 a.m. Pacific. Investors want details on whether customer sites, power and cooling setups can handle Rubin when it ships in the fourth quarter—a less visible but essential step for turning plans into sales.