Nvidia opened the year's biggest tech show by admitting it had rebuilt basically its entire product line at once, and framed that as a feature rather than a headache. At CES 2026, CEO Jensen Huang put Rubin, the company's next AI computing platform, into full production on stage at the Fontainebleau in Las Vegas: six new chips launching together instead of trickling out one generation at a time the way GPUs used to.
Huang named the platform for Vera Rubin, the astronomer whose work on galaxy rotation helped point scientists toward dark matter, and the naming is not just flattery. Nvidia calls its approach "extreme codesign," a fancier way of saying it designed the chips, the server trays, the racks, the networking and the software all at the same time instead of building each piece separately and hoping they play nice later. Think of it as the difference between commissioning an orchestra to write one symphony together versus mailing sheet music to five composers who have never met.
The lineup includes Rubin GPUs rated at 50 petaflops of inference performance using a compact number format called NVFP4, Vera CPUs built to shuffle data and manage AI agents rather than crunch numbers themselves, and a new generation of networking gear: NVLink 6 for wiring chips together inside a rack, Spectrum-X Ethernet Photonics for linking racks to each other, plus ConnectX-9 network cards and BlueField-4 chips that handle data traffic so the GPUs do not have to stop and do it themselves. NVFP4 represents numbers using just 4 bits instead of the far more precise 32-bit format most conventional software still relies on, and it works because AI math tolerates rounding a lot better than a bank statement does.
Economics, not raw speed, was how Huang framed the whole platform. "Computing has been fundamentally reshaped as a result of accelerated computing, as a result of artificial intelligence," he said. "What that means is some $10 trillion or so of the last decade of computing is now being modernized to this new way of doing computing." Nvidia says Rubin cuts the cost of generating each AI token to roughly a tenth of what the Blackwell platform charged, though the company has not published the workload or methodology behind that comparison, so the figure is a company estimate, not an audited benchmark.
"The faster you train AI models, the faster you can get the next frontier out to the world," Huang told the crowd. "This is your time to market. This is technology leadership." Nvidia also introduced a companion storage system, called the Inference Context Memory Storage Platform, meant to keep a growing pile of AI conversation history close enough to the chips that they are not left waiting for it. The company claims five times the throughput and five times the performance per dollar of its previous approach, another set of figures that arrived without a published methodology.
Leave a Reply