Vera Rubin Release Date: What Changes From GB300 to Rubin

Vera Rubin's release date stopped being the interesting question weeks ago. NVIDIA has held fall 2026 as the delivery window for its first cloud cohort across its public roadmap for most of the year, and nothing since has moved that date. What's actually unresolved, if you're a team running GB300 (Blackwell Ultra) in production right now, is what breaks underneath you when Rubin capacity starts landing, and whether that's a reason to reserve ahead of it or keep running GB300 another cycle.

TL;DR: Vera Rubin Release Date and What Changes From GB300 to Rubin

• Release date: Vera Rubin NVL72 ships to its first cloud cohort in fall 2026, a target NVIDIA has held across its 2026 roadmap and not revised.
• Not affected by the Kyber delay: that slip to 2028 applies to Rubin Ultra's NVL144/NVL576 rack, not to base Vera Rubin NVL72.
• Software target: CUDA Toolkit 13.4 (9 September 2026) added preview support for Rubin at compute capability SM_107, explicitly marked not for production.
• Driver: Rubin (sm_107) kernels need the R615 driver branch; the R610 driver bundled with the CUDA 13.4 Developer Preview isn't enough on its own.
• Memory default: on coherent platforms including Vera Rubin, CUDA 13.4 flips the driver's default from NUMA to CDMM, so code assuming NUMA topology needs an explicit override before upgrading.
• On Spheron: B200 runs $11.12/hr on-demand and B300 runs $5.84/hr spot, both as of 27 Oct 2026; GB300 and Vera Rubin are reservation-only. Pre-order Vera Rubin capacity.

Pricing fluctuates based on GPU availability. Spheron rates above are live as of 27 Oct 2026; GB300 and Vera Rubin carry no published hourly rate anywhere as of this writing. Check current GPU pricing for live rates.

Why the Vera Rubin Release Date Question Is Surfacing Now

Three things landed close together. CUDA Toolkit 13.4 shipped on 9 September 2026 with the first compiler target for Rubin silicon. NVIDIA's Kyber rack, the system Rubin Ultra depends on, was reported slipping to 2028 in July, which set off a round of "wait, which Rubin is delayed?" confusion that's still circulating. And GB300 has been in production long enough that teams who reserved it early in the year are now asking what the next hardware refresh costs them in migration work, not just dollars.

That last point is what most coverage skips. Specs tables tell you what Rubin can do. They don't tell you what has to change in a GB300 environment that already has a working driver stack, a pinned CUDA version, and memory-management assumptions baked into its code. That's the gap this post fills, and it's the same gap our Vera Rubin NVL72 chip guide leaves open, since it was written before CUDA 13.4 existed.

Architecture at a Glance: Blackwell Ultra (GB300) vs Rubin

The headline numbers are already well covered on this blog, so here's the condensed version. GB300 pairs B300 GPUs (288GB HBM3e, 8 TB/s, NVLink 5 at 1.8 TB/s per GPU) in an NVL72 rack. Rubin moves to HBM4 at up to 22 TB/s, roughly 2.75x the bandwidth at the same 288GB capacity, and NVLink 6 doubles per-GPU interconnect to 3.6 TB/s. For decode-heavy serving and long-context inference, that bandwidth jump matters more day to day than the headline FLOPS figure.

For the full rack-level breakdown, including the FP64 and INT8 trade-offs Blackwell Ultra made against GB200, see the GB300 vs GB200 NVL72 pricing and availability guide. For every published Rubin spec straight from NVIDIA's own table, the Vera Rubin NVL72 guide has the complete breakdown so this post doesn't need to repeat it.

What Actually Changes for Builders: CUDA 13.4, Drivers and NVLink 6

This is the part a spec sheet doesn't cover, and it's where a GB300 team's actual migration cost lives.

Compiler target. CUDA Toolkit 13.4, released 9 September 2026, added preview functional support for Rubin at compute capability 107, commonly written SM_107. That's the first time Rubin has had a real compiler target rather than a roadmap slide. NVIDIA is explicit that this support is not for production or benchmarking, so the practical move for a GB300 team is to start compiling and testing against SM_107 now, not to deploy a production Rubin workload on it yet.

Driver branch. Running sm_107 kernels requires the R615 driver branch; the R610 driver that shipped with the CUDA 13.4 Developer Preview is not sufficient on its own. If your GB300 fleet is pinned to whatever driver branch shipped with your current CUDA version, plan the R615 upgrade as its own change, separate from any toolkit bump, and test it against your existing Blackwell workloads before you need it for Rubin.

Memory management default. On NVIDIA's coherent platforms, including Grace Hopper, Grace Blackwell and Vera Rubin, CUDA 13.4 flips the driver's default memory management mode from NUMA to CDMM (Coherent Driver-based Memory Management). Code that assumes NUMA topology for memory placement or affinity breaks silently under the new default. Select NUMA mode explicitly before you upgrade the driver, rather than discovering the change after a workload's latency profile shifts.

None of these three are GPU-specific bugs. They're toolkit and driver facts that apply the moment you touch CUDA 13.4, whether or not Rubin hardware is near your rack yet, which is exactly why they're worth fixing in your GB300 environment now.

Availability Today vs What's Still Pre-Order

GB300 NVL72 is shipping and quote-gated at most providers today. Vera Rubin NVL72's fall 2026 target for its first cohort of cloud customers hasn't moved, a status our Kyber rack delay post confirms is unaffected by the separate Rubin Ultra slip. That first cohort runs through the largest hyperscalers and neoclouds; broader marketplace access is expected to follow into 2027. A preview compiler target doesn't change any of that: it means NVIDIA and its first-cohort partners are bringing up software ahead of silicon reaching broader hands, which is normal for a generation transition.

On Spheron: B200 runs $11.12/hr on-demand and B300 runs $5.84/hr spot, both as of 27 Oct 2026. GB300 and Vera Rubin are both reservation items rather than hourly SKUs, since neither has a published hourly rate from any provider yet. A reservation secures a place in line through Spheron's data center partner network once capacity clears; it doesn't guarantee a launch-day price or date, and NVIDIA's named first-cohort partners for Vera Rubin get allocations ahead of specialist marketplaces. A hard 2026 Rubin deadline belongs with that cohort first; a team planning past that window can reserve on Spheron and keep its B200/B300 fleet running meanwhile.

Pricing fluctuates based on GPU availability. Spheron rates above are live as of 27 Oct 2026; GB300 and Vera Rubin carry no published hourly rate anywhere as of this writing. Check current GPU pricing for live rates.

Should You Reserve GB300 Now or Wait for Rubin?

Reserve GB300 now if you need rack-scale capacity running this quarter, or your workload is bandwidth-sensitive enough that waiting six-plus months for Rubin's 22 TB/s costs you real revenue. GB300 is available today, and the driver and CUDA groundwork above (R615, CDMM default) is worth doing on your existing fleet regardless of which generation you reserve next.

Reserve Vera Rubin ahead of availability only if your timeline already points past fall 2026 and you want a confirmed place in queue once first-cohort capacity clears to specialist marketplaces. Don't treat a Rubin reservation as a reason to delay a GB300 decision you need resolved this year. Our Rubin NVL72 cloud availability and cost-per-token post walks through that allocation sequencing in more detail.

Where to Go Deeper

This post stays scoped to the GB300-to-Rubin jump specifically. If you want the three-generation view instead, with Hopper included alongside Blackwell and Rubin, the NVIDIA Rubin vs Blackwell vs Hopper comparison lays out memory, compute and NVLink progression across all three side by side, plus workload-by-workload guidance on which generation to rent for what.

While your Rubin reservation clears its first cohort, B200 and B300 are both billed on Spheron today, on-demand or spot, with no multi-year contract.

Reserve GB300 or Vera Rubin capacity →