Nvidia now makes $2,300 in revenue each second on the again of the AI revolution. Its knowledge middle enterprise is so gigantic, even its networking {hardware} now rakes in extra money than its gaming GPUs. Now, the corporate is saying the AI GPUs that it hopes will prolong its commanding lead: the Blackwell Extremely GB300, which is able to ship within the second half of this yr, the Vera Rubin for second half of subsequent yr, and the Rubin Extremely that can arrive within the second half of 2027.
This yr’s Blackwell Extremely isn’t we initially anticipated, when Nvidia stated final yr that it might start producing new AI chips on a yearly cadence, sooner than ever earlier than. However Nvidia shortly moved on from Blackwell Extremely throughout immediately’s GDC keynote to disclose its subsequent structure, Vera Rubin, whose full rack ought to provide 3.3x the efficiency of a comparable Blackwell Extremely one.
Nvidia isn’t making it simple to inform how a lot better Blackwell Extremely is than the unique Blackwell. In a prebriefing with journalists, Nvidia revealed a single Extremely chip will provide the identical 20 petaflops of AI efficiency as Blackwell, however now with 288GB of HBM3e reminiscence slightly than 192GB of the identical. In the meantime, a Blackwell Extremely DGX GB300 “Superpod” cluster will provide the identical 288 CPUs, 576 GPUs and 11.5 exaflops of FP4 computing because the Blackwell model, however with 300TB of reminiscence slightly than 240TB.
Principally, Nvidia in contrast its new Blackwell Extremely to the H100, the 2022 chip that initially constructed Nvidia’s AI fortunes and from which main corporations may presumably need to improve: there, Nvidia says it gives 1.5x the FP4 inference and may dramatically pace up “AI reasoning,” with the NVL72 cluster able to operating an interactive copy of DeepSeek-R1 671B that may present solutions in simply ten seconds as an alternative of the H100’s 1.5 minutes. Nvidia says that’s as a result of it might probably course of 1,000 tokens per second, ten instances that of Nvidia’s 2022 chips.
However one intriguing distinction is that some corporations will have the ability to purchase a single Blackwell Extremely chip: Nvidia introduced a desktop laptop referred to as the DGX Station with a single GB300 Blackwell Extremely on board, 784GB of unified system reminiscence, built-in 800Gbps Nvidia networking, and the promised 20 petaflops of AI efficiency. Asus, Dell, and HP will be part of Boxx, Lambda, and Supermicro in promoting variations of the desktop.
Nvidia may even provide a single rack referred to as the GB300 NVL72 that provides 1.1 exaflops of FP4, 20TB of HBM reminiscence, 40TB of “quick reminiscence,” 130TB/sec of NVLink bandwidth and 14.4 TB/sec networking.
However Vera Rubin and Rubin Extremely might dramatically enhance on that efficiency once they arrive in 2026 and 2027. Rubin has 50 petaflops of FP4, up from 20 petaflops in Blackwell. Rubin Extremely will characteristic a chip that’s successfully accommodates two Rubin GPUs linked collectively, with twice the efficiency at 100 petaflops of FP4, and practically quadruple the reminiscence at 1TB.
A full NVL576 rack of Rubin Extremely claims to supply 15 exaflops of FP4 inference and 5 exaflops of FP8 coaching, for what Nvidia says is 14x the efficiency of the Blackwell Extremely rack it’s transport this yr. Discover different specs beneath:
Nvidia says it has already shipped $11 billion price of Blackwell income; the highest 4 consumers alone have bought 1.8 million Blackwell chips up to now in 2025.
Nvidia’s pushing these new chips — and all its AI chips — as important to the way forward for computing, and is attempting to argue immediately that corporations will want increasingly computing energy, not much less as some assumed after DeepSeek shook up investor assumptions and despatched Nvidia’s inventory value tumbling. On the Nvidia GPU Expertise Convention immediately, founder and CEO Jensen Huang says the business wants “100 instances greater than then we thought we wanted this time final yr” to maintain up with demand.