Nvidia now makes $2,300 in profit every second on the back of the AI revolution. Its data center business is so gigantic, even its networking hardware now rakes in more money than its gaming GPUs. Now, the company is announcing the AI GPUs that it hopes will extend its commanding lead: the Blackwell Ultra GB300, which will ship in the second half of this year, the Vera Rubin for second half of next year, and the Rubin Ultra that will arrive in the second half of 2027.
This year’s Blackwell Ultra isn’t we originally expected, when Nvidia said last year that it would begin producing new AI chips on a yearly cadence, faster than ever before. But Nvidia quickly moved on from Blackwell Ultra during today’s GDC keynote to reveal its next architecture, Vera Rubin, whose full rack should offer 3.3x the performance of a comparable Blackwell Ultra one.
Nvidia isn’t making it easy to tell how much better Blackwell Ultra is than the original Blackwell. In a prebriefing with journalists, Nvidia revealed a single Ultra chip will offer the same 20 petaflops of AI performance as Blackwell, but now with 288GB of HBM3e memory rather than 192GB of the same. Meanwhile, a Blackwell Ultra DGX GB300 “Superpod” cluster will offer the same 288 CPUs, 576 GPUs and 11.5 exaflops of FP4 computing as the Blackwell version, but with 300TB of memory rather than 240TB.
Image: Nvidia
Mostly, Nvidia compared its new Blackwell Ultra to the H100, the 2022 chip that originally built Nvidia’s AI fortunes and from which leading companies might presumably want to upgrade: there, Nvidia says it offers 1.5x the FP4 inference and can dramatically speed up “AI reasoning,” with the NVL72 cluster capable of running an interactive copy of DeepSeek-R1 671B that can provide answers in just ten seconds instead of the H100’s 1.5 minutes. Nvidia says that’s because it can process 1,000 tokens per second, ten times that of Nvidia’s 2022 chips.
But one intriguing difference is that some companies will be able to buy a single Blackwell Ultra chip: Nvidia announced a desktop computer called the DGX Station with a single GB300 Blackwell Ultra on board, 784GB of unified system memory, built-in 800Gbps Nvidia networking, and the promised 20 petaflops of AI performance. Asus, Dell, and HP will join Boxx, Lambda, and Supermicro in selling versions of the desktop.
Nvidia will also offer a single rack called the GB300 NVL72 that offers 1.1 exaflops of FP4, 20TB of HBM memory, 40TB of “fast memory,” 130TB/sec of NVLink bandwidth and 14.4 TB/sec networking.
But Vera Rubin and Rubin Ultra may dramatically improve on that performance when they arrive in 2026 and 2027. Rubin has 50 petaflops of FP4, up from 20 petaflops in Blackwell. Rubin Ultra will feature a chip that’s effectively contains two Rubin GPUs connected together, with twice the performance at 100 petaflops of FP4, and nearly quadruple the memory at 1TB.
A full NVL576 rack of Rubin Ultra claims to offer 15 exaflops of FP4 inference and 5 exaflops of FP8 training, for what Nvidia says is 14x the performance of the Blackwell Ultra rack it’s shipping this year. Find other specs by blowing up the images below:
Nvidia says it has already shipped $11 billion worth of Blackwell revenue; the top four buyers alone have purchased 1.8 million Blackwell chips so far in 2025.
Nvidia’s pushing these new chips — and all its AI chips — as essential to the future of computing, and is trying to argue today that companies will need more and more computing power, not less as some assumed after DeepSeek shook up investor assumptions and sent Nvidia’s stock price tumbling. At the Nvidia GPU Technology Conference today, founder and CEO Jensen Huang says the industry needs “100 times more than then we thought we needed this time last year” to keep up with demand.
Huang says Nvidia’s next architecture after Vera Rubin, coming 2028, will be named Feynman — presumably after Richard Feynman, the famous theoretical physicist. He said some of pioneering astronomer Vera Rubin’s family was in the audience today.
