When Does NVIDIA Vera Rubin Ship? Buy Blackwell Now or Wait?

NVIDIA's Vera Rubin platform has moved from roadmap slide to shipping hardware in the second half of 2026, but early volume is going to hyperscalers and large cloud providers. For most enterprise buyers, the practical question is not whether Rubin is faster than Blackwell (it is, by NVIDIA's figures) but when you could realistically receive and power it, and what you lose by waiting. Our view: buy Blackwell for anything you need in production before mid-2027, and plan Rubin for later projects. Here is the evidence.

What is NVIDIA Vera Rubin?

Vera Rubin is the successor to Blackwell, announced in detail at CES in January 2026 as a six-chip platform: the Vera CPU, Rubin GPU, NVLink 6 switch, ConnectX-9 SuperNIC, BlueField-4 DPU and Spectrum-6 Ethernet switch (NVIDIA). Each Rubin GPU carries 288GB of HBM4 at 22 TB/s, with 50 petaFLOPS of NVFP4 for inference and 35 petaFLOPS for training, while the Vera CPU has 88 custom Arm cores (The Register). The flagship Vera Rubin NVL72 rack combines 72 Rubin GPUs and 36 Vera CPUs with 20.7 TB of HBM4. NVIDIA dropped the planned NVL144 name and kept counting packages rather than dies.

• Memory bandwidth: 22 TB/s per GPU, which The Register reports as 2.8 times Blackwell, the main driver of faster token generation.
• NVIDIA's claims versus Blackwell: up to a 10x reduction in inference token cost and a 4x reduction in GPUs needed to train mixture-of-experts models (NVIDIA). Treat these as vendor best cases.
• An 8-GPU format, HGX Rubin NVL8, for x86 servers, with Dell, HPE, Lenovo and Supermicro expected to build systems on it.
• Liquid cooling becomes mandatory for HGX systems this generation, according to The Register.

When does Vera Rubin ship?

NVIDIA said in January that Rubin-based products would be available from partners in the second half of 2026. On August 26, 2026, its quarterly results described Vera Rubin as ramping into full production, with racks running at partners including CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure and Nebius (NVIDIA). Supermicro announced on September 23, 2026 that it is shipping Vera Rubin NVL72 racks (Supermicro), and CoreWeave announced Cognition as the first customer running production workloads on Vera Rubin NVL72, describing the service as in limited availability for select customers (CoreWeave).

So Rubin is real and shipping, but limited availability is the operative phrase. NVIDIA's supply commitments jumped to $279 billion, mostly memory, according to Jon Peddie Research, and the largest cloud buyers have first call on that output. Enterprises ordering a handful of servers should not expect Rubin deliveries at scale in 2026.

Has Vera Rubin been delayed?

The standard Rubin platform has not been publicly delayed. The delay reports concern the next step. At GTC in March 2026, NVIDIA previewed Kyber, a rack with 144 GPUs in vertical compute trays for Vera Rubin Ultra, targeted for 2027, followed by the Feynman generation in 2028 (Jon Peddie Research). On July 6, 2026, SemiAnalysis reported that Kyber had slipped by more than 12 months to 2028 because of manufacturing problems with the rack's printed circuit board midplane (CNBC). NVIDIA disputed it, telling Bloomberg that its road map is intact (DCD). Treat Rubin Ultra timing as uncertain and do not base 2027 plans on it.

What does Rubin need from your data center?

Much more than most enterprise rooms provide. Supermicro lists its Vera Rubin NVL72 rack at 227kW, with direct liquid cooling capturing more than 98% of heat (Supermicro). That is well above the roughly 135kW that Lenovo quotes for a GB300 NVL72 rack. Supermicro's HGX Rubin NVL8 system is a 2U liquid-cooled design, with up to nine systems in a rack. If your site cannot provide facility water and high-density power today, the build-out alone may take longer than the wait for hardware. Our guide to power and cooling planning for dense AI deployments lists the work involved.

Will enterprises be able to buy HGX Rubin NVL8 servers?

Yes, eventually. NVIDIA describes HGX Rubin NVL8 as a board linking eight Rubin GPUs through NVLink for x86-based platforms, and names Dell, HPE, Lenovo and Supermicro as expected server partners (NVIDIA). The first cloud providers NVIDIA listed for 2026 were AWS, Google Cloud, Microsoft and OCI, plus CoreWeave, Lambda, Nebius and Nscale. That ordering matters for planning: cloud providers and rack-scale NVL72 buyers are first in line, and 8-GPU enterprise servers typically follow once OEMs have validated their designs and supply broadens. Until your preferred OEM publishes ship dates and configurations for its HGX Rubin systems, treat any Rubin delivery date in an enterprise quote as provisional, and ask the supplier what it is based on.

Should you buy Blackwell now or wait for Rubin?

Buy Blackwell now if any of these apply:

• You need production capacity in the next two or three quarters. Blackwell servers are shipping through OEMs today, while Rubin volume is concentrated at cloud providers.
• Your site is air-cooled or limited to around 10 to 40kW per rack. HGX B200 and HGX B300 are available in air-cooled servers; Rubin HGX is liquid-cooled.
• Your workload is fine-tuning, retrieval-augmented generation or moderate inference that does not need Rubin's bandwidth.
• You can depreciate the hardware over three years or more. Blackwell will run current models productively long after Rubin ships. Our article on upgrade timing from Hopper to Blackwell explains the cycle.

Wait for Rubin, or reserve it in the cloud first, if your first deployment is mid-2027 or later, you are building a new liquid-cooled facility anyway, and your workload is large-scale inference or mixture-of-experts training where cost per token decides the economics.

The fastest GPU on paper is the one you can receive, power and cool before your project deadline.

Will Blackwell prices fall when Rubin ships?

Do not count on it in the near term. GPU rental prices rose in early 2026, with SemiAnalysis reporting H100 one-year contract pricing up almost 40% between October 2025 and March 2026 (SemiAnalysis), and memory costs for servers have climbed sharply. A new generation has historically softened prices for the previous one over time, but in a supply-limited market, Blackwell demand from buyers who cannot get Rubin may keep prices firm. Our posts on H100 and H200 pricing and why server prices went up in 2026 show how this has played out.

How do you plan a purchase across two GPU generations?

• Buy for the workload you have now, not the model you might train in two years.
• Size power and cooling for the next generation even if you buy the current one, so a later Rubin deployment does not require another facility project.
• Keep the network forward compatible: 800G optics and switches serve both B300 and Rubin.
• Use cloud reservations to test Rubin on your own models before committing capital.
• Write a refresh point into your roadmap, as described in our multi-year AI infrastructure roadmap.

Nexus Compute supplies new Blackwell GPU servers and components through authorized distribution, with the full manufacturer warranty, and checks each configuration against your power, cooling and platform before dispatch. Tell us your timeline and site limits and we will quote realistic options within 48 business hours, with purchase orders and net terms for qualified organizations.