Nvidia unveils Vera Rubin platform at GTC 2025

At its GTC 2025 keynote, Nvidia chief executive Jensen Huang unveiled the company’s next data-centre GPU architecture, named Vera Rubin after the astronomer whose work provided evidence for dark matter. Paired with Nvidia’s custom Vera CPU, the platform was claimed to reach up to 50 petaflops of inference throughput — more than double the roughly 20 petaflops Nvidia attributed to the current Blackwell generation — with the Vera CPU itself roughly twice as fast as Blackwell’s Grace CPU. Vera Rubin was guided for availability in the second half of 2026.

The announcement doubled as confirmation of an annual product cadence: Nvidia laid out a roadmap running from Blackwell Ultra, launching later in 2025, through Vera Rubin in H2 2026 to Rubin Ultra — a four-GPU package rated at up to 100 petaflops — in H2 2027, with a further architecture named Feynman slated for 2028. Committing publicly to a yearly release schedule was itself notable: it signalled Nvidia’s confidence that it could keep outpacing rivals such as AMD on a cycle competitors would have to match to stay relevant, and gave cloud providers and hyperscalers multi-year visibility to plan capital spending against.

As with all vendor-supplied performance figures, the petaflop and throughput claims came from Nvidia itself, ahead of independent benchmarking on shipping hardware. The announcement nonetheless set the terms for a compute buildout in which Nvidia’s roadmap, rather than any single product launch, had become the industry’s de facto capacity-planning calendar.