NVIDIA Rubin
Vera Rubin Release Date, Specs, Rack Power, and What It Means for GPU Terms Signed Today
Vera Rubin is shipping to the largest clouds now, with broad access expected in 2027. Most teams will rent or host Blackwell first, so the question is how to sign today without being stuck when Rubin arrives. We find the capacity and write the upgrade path into the terms.
Price My GPU RequirementUpdated October 22, 2026. Figures checked against NVIDIA’s product pages and earnings statements; analyst estimates are labeled as estimates.
NVIDIA has moved to a new data center GPU architecture every year, and Rubin keeps that pace. Each Rubin GPU pairs with NVIDIA's new Vera CPU, built on 88 custom Arm cores, in the Vera Rubin superchip (often called the VR200), uses 288 GB of HBM4 memory per GPU, and ships in a liquid-cooled rack that NVIDIA says delivers up to five times Blackwell's inference throughput at up to ten times lower cost per token. Those are vendor claims for the newest workloads, and they explain why the first allocations went to the hyperscalers. Until Rubin is broadly available, most teams fill the gap with GPU rental on Blackwell, and teams that own racks need AI and GPU colocation built for Rubin’s density, which very few halls offer today. This page covers the NVIDIA Rubin release date, the specs, what a Rubin rack needs from a data center, and how to plan GPU capacity around it.
Sources: NVIDIA, May 2026; NVIDIA second quarter fiscal 2027 results, August 2026; NVIDIA on CoreWeave, September 2026; rack power from analyst Ming-Chi Kuo, January 2026; Rubin Ultra timing from DCD, July 2026, citing SemiAnalysis. Future dates are NVIDIA targets and can move.
NVIDIA GPU Roadmap: Blackwell to Feynman
Today's rentable flagship is Blackwell Ultra, the B300 and GB300 NVL72. Prices and availability for current GPUs are on B200 and B300 rental and GB200 and GB300 NVL72.
Renting or owning, the constraint is the same: very few facilities can power and cool a 200 kW rack today, and the ones that can are being reserved now. Send your GPU count, workload, term and start date. You'll hear back within 24 hours with what is available now, what Rubin changes, and the terms that keep you flexible.
Vera Rubin Specs vs Blackwell
Per-GPU and per-rack figures side by side. NVIDIA has revised several Rubin numbers since CES, so where its current product page differs from the January launch figures, the current page is shown and the earlier figure noted.
| Per GPU | B200 (Blackwell) | B300 (Blackwell Ultra) | Rubin |
|---|---|---|---|
| Memory | 192 GB HBM3e | 288 GB HBM3e | 288 GB HBM4 |
| Memory bandwidth | 8 TB/s | 8 TB/s | 19.2 TB/s on NVIDIA's current page (22 TB/s at launch) |
| NVFP4 compute | 10 PF dense | 15 PF dense | 35 PF dense (training), 50 PF sparse (inference) |
| NVLink per GPU | 1.8 TB/s | 1.8 TB/s | 3 TB/s on NVIDIA's current page (3.6 TB/s at launch), NVLink 6 |
| Power | Up to 1,200 W | Up to 1,400 W | About 1.8 to 2.3 kW (analyst estimate; NVIDIA has not published a TDP) |
| Transistors | 208 billion | 208 billion | 336 billion, two compute dies |
| Per rack | GB200 NVL72 | GB300 NVL72 | Vera Rubin NVL72 |
|---|---|---|---|
| GPUs and CPUs | 72 GPUs, 36 Grace CPUs | 72 GPUs, 36 Grace CPUs | 72 GPUs, 36 Vera CPUs (3,168 Arm cores) |
| GPU memory | 13.4 TB HBM3e | 20 TB HBM3e | 20.7 TB HBM4 |
| Memory bandwidth | 576 TB/s | 576 TB/s | 1,400 TB/s |
| NVFP4 inference | 1,440 PF sparse | 1,440 PF sparse | 3,600 PF sparse |
| NVLink | 130 TB/s | 130 TB/s | 216 TB/s |
| Rack power | About 132 kW nominal | About 132 kW nominal, 155 kW peak | About 190 to 230 kW (analyst estimate) |
| Cooling | Liquid, direct to chip | Liquid, direct to chip | 100 percent liquid, 45°C supply water |
| Status | Shipping | Shipping | Shipping since August 2026, largest clouds first |
Sources: NVIDIA product pages for Vera Rubin NVL72, GB300 NVL72, GB200 NVL72 and HGX, read October 22, 2026; NVIDIA’s Rubin platform overview and Blackwell Ultra overview; Blackwell rack power from HPE QuickSpecs, September 2026; Rubin power from analyst Ming-Chi Kuo. NVIDIA’s 2025 roadmap called this rack the Vera Rubin NVL144, counting GPU dies; it is the same rack, now marketed as NVL72. Performance figures are vendor claims. Free to cite with a link to this page.
What a Rubin Rack Needs From a Data Center
Each generation draws more power per rack, and Rubin is a big step. A Vera Rubin NVL72 rack draws about 190 to 230 kW by analyst estimates, roughly half again what a GB300 NVL72 draws, and NVIDIA’s next rack design, Kyber, was planned at about 600 kW. Very few existing colocation halls can take either. Even the eight-GPU HGX Rubin NVL8 server is liquid-cooled only.
Blackwell from HPE QuickSpecs, September 2026; Vera Rubin from analyst Ming-Chi Kuo, January 2026; Kyber from NVIDIA’s GTC 2025 plan, as reported by DCD. NVIDIA has not published rack power for Rubin. Free to cite with a link to this page.
Liquid cooling, sized for the next rack
Rubin runs entirely on liquid, with 45°C supply water. A hall built for 130 kW racks will not serve 230 kW without more plant, flow and floor. Ask what density the loop is designed for, not what it runs today. See direct to chip cooling.
Power per rack position
Megawatts available to the building are not the same as 200 kW delivered to one rack. Get the per-rack figure and the energization date in writing.
A lease that outlives the hardware
A colocation lease runs five to ten years; a GPU generation, about one. Sign for the density path, including 800 V DC readiness, not today’s rack. See AI and GPU colocation.
Send the rack count, density and date. We find the liquid-cooled space that can actually deliver 200 kW per rack on your timeline, before it is reserved. The operator you choose pays us.
What Rubin Costs
NVIDIA does not publish prices, and no cloud has published a Rubin rate per GPU-hour. These are the most credible estimates, and they are estimates.
| Item | Estimate | Source |
|---|---|---|
| Vera Rubin NVL72 rack, bill of materials | About $7.8 million, including about $4 million of GPUs and about $2 million of memory | Morgan Stanley, May 2026, via Wccftech |
| Vera Rubin NVL72 rack, high case | About $9.1 million if HBM4 prices keep rising in 2027 | Bernstein, July 2026, as reported |
| Rubin GPU | About $55,000 each, implied by the rack estimate | Morgan Stanley, May 2026, as reported |
| GB300 NVL72 rack, for comparison | About $4 million | Morgan Stanley, as reported September 2026 |
| Rubin rental, per GPU-hour | No published rate; early capacity sold by reservation and custom quote | CoreWeave pricing, checked October 2026 |
Rubin CPX and Rubin Ultra
Rubin CPX, off the roadmap
Announced in September 2025 as a specialized GPU for the context stage of inference, with a Vera Rubin NVL144 CPX platform targeted for the end of 2026. It was missing from NVIDIA’s GTC 2026 roadmap, where Groq 3 LPX inference racks took its place. NVIDIA has not formally cancelled it.
Rubin Ultra, 2027 target
Four compute chiplets and about 1 TB of HBM4e per package, in the Kyber rack, which NVIDIA restated at GTC 2026 as Kyber NVL144. NVIDIA’s target is the second half of 2027. In July 2026, SemiAnalysis reported Kyber slipping to 2028; NVIDIA said its roadmap is intact.
Feynman, 2028
The architecture after Rubin, paired with a new CPU called Rosa, on NVIDIA’s public roadmap for 2028. Detailed specifications have not been published.
Rubin or Blackwell: Planning GPU Capacity Now
Rubin will not be broadly rentable for most teams until 2027, so the near-term decision is how to buy Blackwell without regretting it.
| When you need capacity | What to do | Terms to negotiate |
|---|---|---|
| Before 2027 | Rent or reserve Blackwell: B200, B300 or the GB300 NVL72 | An upgrade path to Rubin, a shorter term, or a renewal option at a capped rate |
| Through 2027 | Reserve Blackwell now and plan a move to Rubin as it opens up | A price review or step-down as Rubin capacity arrives |
| 2027 or later | Plan for Rubin, and reserve early: allocations are claimed in order | Delivery date in writing, with a remedy if it slips |
| Owning racks | Buy Blackwell for near-term use, or wait for Rubin, but secure the hall now | Liquid-cooled space sized for 230 kW racks, with a path to 800 V DC and beyond |
Allocation goes to the largest buyers first
Like the GB200 NVL72 before it, early Rubin capacity is going to hyperscalers and the biggest neocloud contracts. Smaller teams get access later, and reservations matter.
A long Blackwell term needs an exit
A 36-month reservation signed today runs well into the Rubin years. The rate is only half the deal; the upgrade path and exit are the other half.
Owning means planning the building
Owned racks need a hall designed for the next density step, and that space is scarcer than the GPUs. Developers planning for it can start with a build to suit data center scoped through our data center consulting.
Send your GPU count, workload, term and start date. You'll hear back within 24 hours, and the provider you choose pays us.
How We Help
You send the requirement
GPU model, count, workload, term and start date, or for owned racks, the density and timeline. You'll hear back within 24 hours.
We find what exists
GPU providers that can deliver Blackwell now or Rubin when it opens up, or liquid-cooled colocation for owned racks, confirmed before you commit.
We negotiate the terms
Rate, delivery date, upgrade path and exit, so the capacity you buy today does not trap you when Rubin arrives. The provider you choose pays us.
NVIDIA Rubin: Frequently Asked Questions
When does NVIDIA Rubin come out?
It is already shipping to the largest buyers. NVIDIA launched the Vera Rubin platform at CES in January 2026 and began production shipments of Vera Rubin NVL72 racks in August 2026. By the end of August, racks were running at CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud and Nebius, and CoreWeave made Rubin generally available in September 2026. Broad rental access for most companies is expected in 2027.
Can I rent a Vera Rubin GPU yet?
Only in limited ways. CoreWeave made Rubin generally available in September 2026, and Google Cloud, Microsoft Azure, Oracle Cloud and Nebius have racks running, but early capacity is going mostly to large contracts and reservations, and no cloud has published a Rubin rate per GPU-hour. For most teams the practical window to rent Rubin is 2027.
What are the NVIDIA Rubin GPU specs?
Each Rubin GPU has 336 billion transistors across two compute dies, 288 GB of HBM4, 19.2 TB/s of memory bandwidth on NVIDIA's current product page (22 TB/s was quoted at launch), 35 petaflops of dense NVFP4 for training and 50 petaflops of sparse NVFP4 for inference, and NVLink 6. NVIDIA has not published its power draw; analysts put it at about 1.8 to 2.3 kW.
What is Vera Rubin?
NVIDIA's platform for the Rubin generation: the Rubin GPU paired with NVIDIA's new Vera CPU, which has 88 custom Arm cores, plus HBM4 memory and NVLink 6. It ships as the liquid-cooled Vera Rubin NVL72 rack and as the HGX Rubin NVL8, an eight-GPU server board. It is named after the astronomer Vera Rubin.
What is the VR200?
The common name for the Vera Rubin superchip, a Vera CPU paired with Rubin GPUs, which is the building block of the Vera Rubin NVL72 rack.
What is the Vera Rubin NVL72?
NVIDIA's rack-scale Rubin system: 72 Rubin GPUs and 36 Vera CPUs connected by NVLink 6 in one fully liquid-cooled rack, with 20.7 TB of HBM4, 1,400 TB/s of memory bandwidth and 3,600 petaflops of sparse NVFP4 inference, per NVIDIA's product page. NVIDIA's 2025 roadmap called it the Vera Rubin NVL144, counting GPU dies rather than packages.
How much faster is Rubin than Blackwell?
NVIDIA claims up to 5 times Blackwell's inference throughput, 3.5 times its training performance and up to 10 times lower cost per token on mixture-of-experts models. At rack level, NVIDIA's own figures show the Vera Rubin NVL72 with 3,600 petaflops of sparse NVFP4 inference against 1,440 for the GB300 NVL72, and 1,400 TB/s of memory bandwidth against 576. These are vendor figures, not independent tests.
How much power does a Vera Rubin rack use?
NVIDIA has not published a figure. Analyst estimates put a Vera Rubin NVL72 rack at about 190 to 230 kW, compared with about 132 kW nominal and 155 kW peak for a GB300 NVL72. At those estimates, 1 MW of IT load holds about 310 to 380 Rubin GPUs; NVIDIA's own page cites 40,000 GPUs in 100 MW.
Does Vera Rubin need liquid cooling?
Yes. The Vera Rubin NVL72 is 100 percent liquid-cooled with 45°C supply water, and even the eight-GPU HGX Rubin NVL8 server is liquid-cooled only. At about 190 to 230 kW per rack, it needs a facility loop designed for that density.
What happened to Rubin CPX?
NVIDIA announced Rubin CPX in September 2025 as a GPU for the context stage of inference, with a Vera Rubin NVL144 CPX platform targeted for the end of 2026. It was missing from NVIDIA's GTC 2026 roadmap in March 2026, where Groq 3 LPX inference racks took its place. NVIDIA has not formally cancelled it.
When does Rubin Ultra ship?
NVIDIA's target is the second half of 2027, with four compute chiplets and about 1 TB of HBM4e per package in the Kyber rack. In July 2026, SemiAnalysis reported the Kyber rack slipping to 2028 and a smaller two-die Rubin Ultra; NVIDIA said its roadmap is intact.
What is the Kyber rack?
NVIDIA's next rack design for Rubin Ultra. It was announced at GTC 2025 as the Kyber NVL576 at about 600 kW, and restated at GTC 2026 as Kyber NVL144, counting 144 Rubin Ultra packages. It is a new rack format, not an upgrade to today's NVL72 racks, and is tied to NVIDIA's 800 V DC power architecture.
What comes after Rubin?
Feynman, on NVIDIA's public roadmap for 2028, paired with a new CPU called Rosa. Detailed specifications have not been published.
How much does NVIDIA Rubin cost?
NVIDIA has not published prices. Morgan Stanley estimated the bill of materials for a Vera Rubin NVL72 rack at about $7.8 million in May 2026, including about $4 million of GPUs, roughly twice its estimate of about $4 million for a GB300 NVL72. Bernstein put a high case at about $9.1 million if memory prices keep rising. No cloud has published a Rubin rate per GPU-hour.
Should I wait for Rubin or rent Blackwell now?
If you need capacity before 2027, rent Blackwell now: Rubin allocations go to the largest buyers first. Negotiate an upgrade path, a shorter term or a renewal option, so a long Blackwell commitment does not lock you out when Rubin is available to you.
Will Blackwell rental prices fall when Rubin ships?
Past transitions have pushed older-generation rental rates down as new capacity arrives, though demand has kept prices firm through 2026 and NVIDIA says it is supply-constrained. Build that possibility into reserved terms with a price review or a shorter commitment.
Can Metro Colo Advisory help with Rubin capacity?
Yes, in two ways: we find rented GPU capacity on terms that keep you flexible, and we find liquid-cooled colocation for owned racks at Rubin densities. The provider you choose pays us, so it costs you nothing.
Price Your GPU Requirement
Tell us the GPU, how many, the term and the start date. You'll hear back within 24 hours from the person who will run your search. If waiting for Rubin is the better answer, we will tell you. No cost, no obligation, and the provider you choose pays us.
Once you have contacted a provider directly, it usually will not work through an advisor on your deal.
Renting GPUs now: the GPU rental hub, B200 and B300 rental and GB200 and GB300 NVL72.
Hosting Rubin racks: AI and GPU colocation, direct to chip cooling and wholesale colocation.