NVIDIA now earns $ 2,300 profit every second at the back of the AI revolution. His activity in the data center is so gigantic, even network equipment now earns more money than a graphic processor. Now the company announces GPU AI that it hopes to extend its commander: Blackwell Ultra GB300, which will be sent in the second half of this year, Vera Rubin for the second half of next year, and Rubin Ultra, which will appear in the second half of 2027.
This year’s Blackwell Ultra is not what we originally expected when Nvidia said last year that she would start producing modern AI systems during the annual term, faster than ever before, because Blackwell Ultra is not on modern architecture. But Nvidia quickly moved from Blackwell Ultra during today’s GTC speech to reveal that the next architecture of Vera Rubin, whose full stand should offer 3.3x performance of comparable Blackwell Ultra One.
Nvidia does not make it easier to say how much better Blackwell Ultra is than the original Blackwell. In the prebroker with journalists, Nvidia revealed that a single Ultra Chip would offer the same 20 peafls AI performance as Blackwell, but now with 288 GB of HBM3E memory instead of 192 GB of the same. Meanwhile, the Blackwell Ultra DGX GB300 “Superpod” cluster will offer the same 288 processors, 576 GPU and 11.5 Exaflts FP4 computers as a Blackwell version, but with 300 TB of memory rather than 240TB.
Photo: Nvidia
Mostly Nvidia compared its modern Blackwell Ultra with H100, the 2022 system, which originally built a fortune of AI nvidia and whose leading companies may want to update: NVIDIA claims that it offers 1.5x FP4 inference and can dramatically accelerate “AI reasoning”, from NVL72, which is organized by the depth 671b Just 10 seconds instead of 1.5 minutes H100. Nvidia claims that this is happening because it can process 1000 tokens per second, 10 times larger than from NVIDIA 2022 systems.
But one intriguing difference is that some companies will be To be able to buy one Blackwell Ultra Chip: Nvidia announced a desktop computer called DGX Station with a single GB300 Blackwell Ultra on board, 784 GB of unified system memory, built -in NVIDIA network by 800 GBS Nvidia and promised 20 Peaflts of Ai. ASUS, Dell and HP will join Boxx, Lambda and Supermicro in the sale of a desktop version.
NVIDIA will also offer a single rack called GB300 NVL72, which offers 1.1 ExAFLOPS FP4, 20TB of HBM memory, 40TB “Fast Memory”, I 130TB/s of nvlink bandwidth I 14.4 TB/set Networking.
But Vera Rubin and Rubin Ultra can dramatically improve this performance when they arrive in 2026 and 2027. Rubin has 50 FP4 petaplops, compared to 20 Petaflts in Blackwell. Rubin Ultra will be equipped with a system that effectively contains two Rubin GPS connected to each other, with twice as high efficiency with 100 petaflts FP4 and almost four times memory at 1 TB.
The full stand NVL576 Rubin Ultra claims that it offers 15 exaflts in the FP4 inference and 5 lectures of FP8 training for what, according to NVIDIA, is 14 times Blackwell Ultra Rack performance that shipping this year. Find other specifications by blowing up the images below:
1/2
Nvidia claims that she has already sent Blackwell revenues worth $ 11 billion; Only the four best buyers have bought 1.8 million Blackwell tokens in 2025 so far.
Nvidia pushed these modern tokens – and all its AI systems – as necessary for the future of calculations and is trying to argue today that companies will need more and more computing power, no less, because some assumed after the Deepseek had the assumptions of investors and sent the price of Nvidia shares. At today’s NVIDIA GPU technology conference, the founder and general director of Jensen Huang claims that the industry needs “100 times more than then we thought that we need this time last year” to keep up with demand.
Huang says that the next Nvidia architecture after Vera Rubin, the upcoming 2028, will be called Feynman – probably later Richard FeynmanThe renowned theoretical physicist. Said some of the Vera Rubin pioneering astronomerThe family was in the audience today.