Nvidia’s Vera Rubin Platform Explained: What Comes After Blackwell

Nvidia’s annual GPU cadence moved from Hopper to Blackwell in 2024, and the successor to Blackwell — announced at Computex 2024 and detailed in full at CES in January 2026 — is Vera Rubin, named for astrophysicist Vera Rubin, whose work provided key evidence for dark matter. It’s not just a new GPU: Nvidia is shipping it as a platform of multiple co-designed chips, the same way Blackwell shipped alongside Grace CPUs and NVLink switches.

Six Chips, One Platform

Per Nvidia’s own developer blog, the Vera Rubin platform launched with six new chips: the Vera CPU, the Rubin GPU, an NVLink 6 switch, a ConnectX-9 SuperNIC, a BlueField-4 DPU, and a Spectrum-6 Ethernet switch. A March 2026 update added a seventh chip to the platform. This is Nvidia’s now-familiar playbook: rather than selling a GPU alone, it sells the CPU, the GPU, the networking silicon, and the switches as one designed-together system, which is also how it captures value across the entire rack rather than just the chip.

The Numbers That Matter

According to Nvidia’s own published specs:

• Compute: the Rubin GPU delivers 50 petaflops of NVFP4 inference and 35 petaflops of NVFP4 training performance, versus 10 petaflops for each on Blackwell — roughly a 3.5x and 5x jump respectively on Nvidia’s own per-chip figures (note this uses NVFP4, a 4-bit floating-point format, so it isn’t directly comparable to performance figures quoted in other precisions).
• Memory: up to 288 GB of HBM4 per GPU, with up to 22 TB/s of aggregate memory bandwidth — roughly 2.8x Blackwell’s HBM3e bandwidth, by Nvidia’s account.
• Transistors: the Rubin GPU packs 336 billion transistors across 224 streaming multiprocessors.
• Networking: NVLink 6 provides 3.6 TB/s of bidirectional bandwidth per GPU — double the prior generation — and in a full NVL72 rack, all 72 GPUs connect in an all-to-all topology through NVLink 6 switch trays.
• Rack-scale numbers: a single NVL72 compute tray delivers 200 petaflops of NVFP4, and a “Vera Rubin Superchip” (two Rubin GPUs plus one Vera CPU) delivers 100 petaflops of NVFP4.
• Vera CPU: up to 1.5 TB of LPDDR5X system memory at up to 1.2 TB/s.

Nvidia’s own marketing also cites up to 10x higher inference throughput, 10x lower cost per token, and a quarter as many GPUs needed for training versus the prior generation, though the company’s blog post doesn’t spell out the exact test conditions behind those multiples, so they should be read as Nvidia’s own framing rather than independently verified benchmarks.

Timeline and What’s Next

Nvidia announced Vera Rubin at Computex 2024 and gave full technical detail at CES in January 2026. Independent reporting indicates the hardware entered full production around May 2026 and deployment around September 2026, consistent with Nvidia’s stated annual-cadence roadmap. Nvidia has also already announced two chips beyond Rubin: Rubin Ultra (described as effectively two Rubin dies connected together, roughly doubling Rubin’s performance to 100 sparse petaflops NVFP4), targeted for 2027, and Feynman as the architecture after that. As of late October 2026, Nvidia and Oracle also announced plans to expand Vera Rubin deployments to multi-gigawatt scale and bring the Vera CPU specifically to Oracle Cloud Infrastructure, underscoring how quickly hyperscalers are committing to the platform even as it’s still rolling out.

Why the Platform Framing Matters

The six-to-seven-chip framing isn’t just marketing color. NVLink 6’s doubled bandwidth and the all-to-all NVL72 topology are what let 72 physical GPUs behave, for many workloads, like one much larger logical accelerator — which is precisely the property large language model training and inference at frontier scale depends on. Nvidia’s moat has never really been the GPU silicon alone; it’s the combination of that silicon with CUDA, NVLink, and now an entire rack-scale system that competitors have struggled to match piece for piece, let alone as an integrated whole.

FAQ

Is Rubin a single chip or a platform?

Rubin is the name of the GPU specifically; “Vera Rubin” refers to the full platform of co-designed chips, which per Nvidia’s developer blog launched with six chips and gained a seventh in March 2026.

How much faster is Rubin than Blackwell?

By Nvidia’s own published specs, Rubin’s per-GPU NVFP4 figures are 50 petaflops inference and 35 petaflops training, versus 10 petaflops each for Blackwell — though these are vendor-reported numbers in a specific low-precision format, not independently benchmarked comparisons.

What comes after Rubin?

Nvidia has already announced Rubin Ultra, targeted for 2027 and described as roughly double Rubin’s performance, followed by an architecture Nvidia calls Feynman.

Sources

• Nvidia Developer Blog — Inside the Nvidia Rubin Platform
• Nvidia Newsroom — Blackwell platform announcement (incl. Vera Rubin / Oracle update)
• Wikipedia — Rubin (microarchitecture)