Why trust Greenbot
Nvidia CEO Jensen Huang announced new artificial intelligence (AI) chips on Tuesday that will need five times more power than current models, raising questions about whether data centers can handle such high energy demands.
At the GPU Technology Conference (GTC) 2025, Huang revealed Nvidia’s chip plans through 2028, including the Rubin Ultra chips that will use 600 kilowatts per server rack when they arrive in late 2027. Current Blackwell B200 racks use about 120 kilowatts each.
The Rubin Ultra chips will have four graphics processing units per package with 1TB of HBM4e memory each. A full NVL576 rack will contain 576 GPU dies and deliver 15 exaflops of FP4 inference performance – about 14 times faster than this year’s Blackwell Ultra systems.
Nvidia’s roadmap starts with Blackwell Ultra in the second half of 2025, which adds more memory to current Blackwell GPUs (288GB, up from 192GB). Next comes Vera Rubin in late 2026, which pairs a new GPU with Nvidia’s first custom CPU called Vera.
Source: Nvidia
The Vera CPU marks a big change for Nvidia, with 88 custom ARM cores and 176 threads. This processor will be twice as fast as the CPU in current Grace Blackwell chips.
Some experts worry about cooling these power-intensive systems. Mark Wade, CEO of Ayar Labs, questioned whether traditional cooling methods can handle such high power levels.
“Just look at the power consumption going up and up on racks with electrical connections,” Wade told Reuters. “Optics is the only technology that gets you off of that train.”
However, Huang said Nvidia had looked at using light-based technology in its main GPU chips but found copper wires still work better. “Copper is far better,” Huang said, pointing out that copper connections break down less often than the current optical options. For now, Nvidia plans to push for more performance instead.
Huang also said Nvidia’s next architecture after Rubin will be named “Feynman” after physicist Richard Feynman and should arrive in 2028.
These announcements come as Nvidia maintains its lead in the AI chip market. The company has already shipped $11 billion worth of Blackwell chips in 2025, with cloud providers buying 1.8 million units.
For users who want to try these new chips, Nvidia announced a desktop computer called the DGX Station with a single GB300 Blackwell Ultra chip. Companies like Asus, Dell, and HP will sell versions of this desktop.
Nvidia’s aggressive roadmap aims to stay ahead of competitors like AMD and Intel while addressing the growing computational demands for AI training, inference, and reasoning tasks that power applications from chatbots to humanoid robots.
