In this note
Vera Rubin NVL144 and Vera Rubin NVL72 are the same rack. NVIDIA used the NVL144 name on its 2025 roadmap and now sells the product as Vera Rubin NVL72: 72 Rubin GPUs and 36 Vera CPUs, drawing 187 kW in the Max-Q power profile or 227 kW in Max-P, all of it cooled by liquid at a 45°C inlet.
There's no list price. Reports during 2026 put a rack at roughly US$5 million to US$7.8 million. First revenue shipments went out in August 2026, so for most buyers this is a 2027 delivery, into a site that needs 1.4 to 1.7 times the power of a GB300 NVL72 rack.
Why Vera Rubin NVL144 is now called NVL72
At GTC in March 2025, Jensen Huang said Blackwell "was named wrong." Each Blackwell GPU package holds two dies, so by die count GB200 NVL72 is a 144-GPU system. NVIDIA said it would count dies from Rubin onward, and the Rubin rack went on the roadmap as Vera Rubin NVL144, rated at 3.6 exaflops of FP4 inference (Tom's Hardware (opens in a new tab)).
That didn't last. By the CES launch in January 2026 NVIDIA was counting packages again, and both the product page (opens in a new tab) and the October 2026 datasheet (opens in a new tab) say Vera Rubin NVL72. The hardware didn't change: 72 packages, 144 dies, 36 Vera CPUs and one NVLink 6 domain, in the same Oberon footprint as GB200 and GB300 (NVIDIA calls it the third-generation MGX rack). Even the headline number survived. The datasheet lists 3,600 PFLOPS of NVFP4 inference, a sparse figure.
VR200 is the name you'll see in analyst and supply-chain reporting, including the Morgan Stanley estimates further down. NVIDIA's own pages don't use it, but it refers to the same rack.
NVL144 is also back in NVIDIA's vocabulary for a different product: its March 2026 Vera Rubin POD post (opens in a new tab) says the next Kyber rack will first appear "with Vera Rubin Ultra as a standalone NVL144 system." A quote for an "NVL144" with a 2027 or 2028 date is probably Rubin Ultra.
Vera Rubin NVL144 CPX is something else again. Announced in September 2025, it paired Rubin GPUs with Rubin CPX (opens in a new tab) prefill chips, and at GTC 2026 NVIDIA's Ian Buck told press "we've pulled CPX" to focus on Groq LPU decode (transcript (opens in a new tab)).
Vera Rubin NVL72 specs vs GB300 NVL72
The compute and memory figures are NVIDIA's, from the Vera Rubin datasheet and the GB300 NVL72 page (opens in a new tab). NVIDIA publishes GB300 FP8 and BF16 with sparsity, so those two are halved here to dense; for NVFP4 it gives a dense number directly. GB300 power and cooling come from an OEM product guide, since NVIDIA's page doesn't list them.
| Per rack | GB300 NVL72 | Vera Rubin NVL72 |
|---|---|---|
| GPUs / CPUs | 72 Blackwell Ultra / 36 Grace | 72 Rubin / 36 Vera |
| NVFP4 dense | 1,080 PFLOPS | 2,520 PFLOPS |
| FP8 dense | 360 PFLOPS | 1,260 PFLOPS |
| BF16 dense | 180 PFLOPS | 288 PFLOPS |
| GPU memory | 20 TB HBM3e | 20.7 TB HBM4 |
| GPU memory bandwidth | 576 TB/s | 1,400 TB/s |
| NVLink bandwidth | 130 TB/s | 216 TB/s |
| CPU memory | 17 TB LPDDR5X | up to 54 TB LPDDR5X |
| Rack power | 135 kW TDP, 155 kW peak | 187 kW Max-Q, 227 kW Max-P |
| Cooling | about 90% liquid, 10% air | 100% liquid, fanless trays |
Memory and NVLink bandwidth came down during the year. NVIDIA's January platform blog (opens in a new tab) quoted "up to 22 TB/s" of HBM4 bandwidth and 3.6 TB/s of NVLink per GPU, and the March POD post still had 3.6 TB/s and 260 TB/s per rack. The October datasheet gives 1,400 TB/s of HBM4 bandwidth per rack (about 19.4 TB/s per GPU) and 3 TB/s of NVLink per GPU. SemiAnalysis had written in February that memory suppliers were struggling to meet NVIDIA's requirements and that early shipments would land closer to 20 TB/s. If someone quotes you 22 TB/s, ask which figure the delivered systems actually meet.
Vera Rubin NVL72 power consumption
The datasheet gives three rack figures.
| Profile | Rack power | Used for |
|---|---|---|
| Max-P | 227 kW | Highest absolute throughput |
| Max-Q | 187 kW | Better performance per watt |
| MaxLPS | 156 kW | More racks inside a fixed facility power budget |
Per GPU, SemiAnalysis (opens in a new tab) puts Max-Q at 1,800 W and Max-P at 2,300 W, against 1,000 to 1,400 W for Blackwell. NVIDIA's own datasheet assumes 1,800 W per GPU in its NVL4 benchmarks, which lines up. Max-Q and Max-P are default profiles on the same hardware, not separate SKUs, and Max-P adds about 20% at the rack.
MaxLPS belongs to NVIDIA DSX, which the datasheet describes as a set of chip, thermal, system and software features for getting the most out of a fixed power budget, partly by reallocating GPU power that would otherwise sit unused. It's aimed at buyers whose limit is megawatts, not floor space. NVIDIA's reference 100 MW AI factory holds 40,000 Rubin GPUs with MaxLPS. That works out to about 555 racks, or roughly 180 kW of facility power per rack once cooling and networking are counted.
The power train is different too. SemiAnalysis describes the reference rack as four 110 kW power shelves in an N+1 arrangement, each fed three-phase 415 to 480 VAC from two 100 A whips and stepping down to a 50 V busbar rated above 5,000 A, which is liquid-cooled. Sites built for 800 VDC can feed the rack from a separate power rack instead, with DC-DC shelves in the GPU rack.
The rack also carries six times the energy storage of GB300 (400 J per GPU, per NVIDIA) to absorb the load swings of synchronized training. NVIDIA says that cuts peak current demand by up to 25%.
For comparison, an OEM's GB300 NVL72 product guide (opens in a new tab) lists 135 kW TDP and up to 155 kW peak, on six or eight 33 kW power shelves with one 60 A whip each.
Vera Rubin NVL72 cooling
It's all liquid. NVIDIA specifies 100% liquid cooling at a 45°C warm-water inlet, and the compute and NVLink switch trays are cable-free, hose-free and fanless. There's no air-cooled version. The OEM guide puts GB300 at roughly 90% liquid and 10% air, so a GB300 room still needs air handling for about a tenth of each rack's heat. A Vera Rubin room needs very little.
The inlet temperature is the same 45°C GB300 is specified for, which helps operating cost: water that warm can usually be cooled with dry coolers rather than chillers.
Volume is the harder problem. SemiAnalysis estimates a Vera Rubin rack at roughly double the power and heat of a Blackwell rack (180 to 220 kW against 120 to 140 kW). A CDU that serves about ten GB200 racks today will serve far fewer of these, and in-row CDUs currently top out around 2 MW. Plan for bigger units or more of them.
Vera Rubin NVL72 price
NVIDIA has never published list prices for NVL72 racks. Everything below is reported by third parties, not quoted by NVIDIA, and the numbers move with the buyer, the configuration and what's counted.
| Reported by | Vera Rubin NVL72 | GB300 NVL72 | Notes |
|---|---|---|---|
| Tom's Hardware, citing an unnamed source (opens in a new tab), March 2026 | US$5M to US$7M | US$6M to US$6.5M | Vera Rubin quotes include about US$1M of flash storage; the source says many prices circulating online are ODM prices without a proper warranty |
| Morgan Stanley, via Tom's Hardware (opens in a new tab), May 2026 | about US$7.8M | about US$4M | Estimated cost to major cloud providers; memory about 25% of it, around US$2M |
| Morgan Stanley, via Tom's Hardware, May 2026 | about US$55,000 per Rubin GPU | n/a | What NVIDIA reportedly plans to charge hyperscalers per GPU in volume |
The two GB300 figures are US$2 million or more apart, which says a lot about how far to trust any single number here.
You may also have seen US$8.8 million in a headline. In that March Tom's Hardware piece, US$7 million to US$8.8 million is the reported range for Rubin Ultra NVL144 (VR300), the next generation, not for this rack.
Prices for 2027 delivery look more likely to rise than fall. Each Vera Rubin rack carries 54 TB of LPDDR5X and about US$1 million of NAND, and DRAM contract prices climbed at record rates through 2026. In August, Bloomberg reported that NVIDIA had told its largest customers server prices would rise by more than 15% in many cases, on Grace Blackwell and Vera Rubin systems shipping in early 2027 (Tom's Hardware (opens in a new tab)).
The rack is also only part of the cost of owning one. Networking outside the rack, the site's power and cooling, spares, staff and financing all come on top. Our note on GPU depreciation and useful life covers how owners account for the hardware over a term.
When Vera Rubin ships
| Date | Event |
|---|---|
| March 2025 | GTC: Vera Rubin NVL144 on the roadmap for the second half of 2026, Rubin Ultra NVL576 for the second half of 2027 |
| September 2025 | Rubin CPX and the Vera Rubin NVL144 CPX rack announced for the end of 2026 |
| January 2026 | CES: launched as Vera Rubin NVL72, "in full production" (opens in a new tab), with partner products in the second half of 2026 |
| March 2026 | GTC: CPX pulled, Groq 3 LPX racks added to the platform, Kyber NVL144 shown for Rubin Ultra |
| May 31, 2026 | NVIDIA says Vera Rubin is ramping into full production (opens in a new tab), with production shipments from the fall |
| July 2026 | SemiAnalysis reports the Kyber rack slipping to 2028 (opens in a new tab); NVIDIA says "our roadmap is intact" |
| August 2026 | First revenue shipments of VR200 NVL72 racks, with a hyperscaler first to deploy; NVIDIA guides about US$20 billion of Vera Rubin revenue in its fiscal third quarter |
| September 2026 | A GPU cloud receives its first production racks (opens in a new tab) and has a customer running production workloads on them |
On the August earnings call, NVIDIA CFO Colette Kress said the company had "purchase orders from every major hyperscaler, AI cloud, and system OEM." Those buyers take the first quarters of output. Tom's Hardware's rough math (opens in a new tab) is that US$20 billion in a quarter comes to about 2,550 to 3,975 racks, if it were all NVL72.
For an enterprise or a research program, a realistic delivery is some time in 2027. Your place in a supplier's queue will matter more than NVIDIA's announcement dates.
A rack delivered in 2027 also arrives close to its successor. Rubin Ultra is still on NVIDIA's roadmap for 2027 despite the reported Kyber delay. Our note on GB300 or waiting for Vera Rubin covers what that means for a multi-year term.
What changes for a site built for GB300
When NVIDIA announced the Rubin rack in 2025 it called it drop-in compatible with Blackwell NVL72 infrastructure, and the footprint is the same. What feeds the rack isn't.
| Site item | GB300 NVL72 | Vera Rubin NVL72 |
|---|---|---|
| IT power per rack | 135 kW TDP, 155 kW peak | 187 kW Max-Q, 227 kW Max-P |
| Heat to air | about 10% | close to none from the trays |
| Liquid inlet | up to 45°C | 45°C |
| Rack power shelves | six or eight 33 kW shelves | four 110 kW shelves, N+1 |
| Feed per power shelf | one 60 A whip, 346 to 480 VAC wye | two 100 A whips, three-phase 415 to 480 VAC, or 800 VDC from a separate power rack |
| CDU sizing | about 10 GB200-class racks per CDU is common | roughly twice the heat per rack |
If we were designing a site now for a 2027 install, we'd do four things:
- Size busways, whips and the cooling loop for 227 kW per rack. Then running Max-Q is an operating choice, not a limit you built in.
- Recount racks per hall. On the same power, a hall laid out for GB300 at 135 kW holds about 40% fewer Vera Rubin racks at 227 kW.
- Size CDUs by heat rather than rack count, which means fewer racks per CDU or larger units.
- Talk to the utility early. The rack's energy storage smooths its own swings, but a multi-megawatt training cluster still changes load fast.
Electricity cost per rack
Power is the easiest line to put in a memo. This is an estimate that assumes the rack runs at its rated profile all year, so treat it as a ceiling for that profile.
| Assumption | Low | High |
|---|---|---|
| Rack IT power | 187 kW (Max-Q) | 227 kW (Max-P) |
| Hours per year | 8,760 | 8,760 |
| PUE | 1.2 | 1.2 |
| Electricity price | US$0.08 / kWh | US$0.12 / kWh |
| Energy per rack-year | about 1.97 GWh | about 2.39 GWh |
| Electricity per rack-year | about US$157,000 | about US$286,000 |
Against a reported rack price of US$5 million to US$7.8 million, that's about 2% to 6% of the hardware cost a year. The bigger unknowns in a 2027 Vera Rubin budget are the rack price, which is still moving with memory, and the delivery date.
Planning a Vera Rubin cluster
We design, source, build and operate NVIDIA GPU clusters, including the site work a 227 kW rack needs: power distribution, liquid cooling and network, in a colocation facility, your own data center or a retrofit. Vera Rubin NVL72 is on our waitlist with first allocation and can't be reserved yet; GB300 NVL72 can. See Vera Rubin NVL72, GB300 NVL72, cluster deployment, reserved clusters or talk to us.
Questions this note answers
Yes. NVIDIA called the rack NVL144 on its 2025 roadmap, counting 144 GPU dies, and now sells it as Vera Rubin NVL72, counting 72 GPU packages. It's the same hardware: 72 Rubin GPUs and 36 Vera CPUs in one NVLink domain.
NVIDIA's datasheet lists 187 kW per rack in the Max-Q profile, 227 kW in Max-P and 156 kW with MaxLPS. A GB300 NVL72 is about 135 kW TDP, so size the electrical and cooling design for 227 kW.
NVIDIA doesn't publish a list price. Reports in 2026 ranged from about US$5 million to US$7.8 million per rack, the top of that range being a Morgan Stanley estimate of what hyperscalers pay.
First revenue shipments of Vera Rubin NVL72 racks went out in August 2026, and the first production racks were running at a hyperscaler and a GPU cloud by September. Most buyers outside the largest cloud providers should plan on 2027 delivery.
Yes. NVIDIA specifies 100% liquid cooling at a 45°C warm-water inlet, and the compute and switch trays have no fans. There's no air-cooled version of the rack.
Rubin Ultra is the GPU generation after Rubin, on NVIDIA's roadmap for 2027 in NVL72, NVL144 (the new Kyber rack) and NVL576 configurations, though SemiAnalysis reported in July 2026 that Kyber slips to 2028. Vera Rubin NVL72 is the current generation, shipping since August 2026.
Sources
- NVIDIA Vera Rubin datasheet (October 2026) (opens in a new tab)
- NVIDIA Vera Rubin NVL72 product page (opens in a new tab)
- NVIDIA Newsroom: Rubin platform launch at CES (January 2026) (opens in a new tab)
- NVIDIA Technical Blog: Inside the NVIDIA Rubin Platform (January 5, 2026) (opens in a new tab)
- NVIDIA Technical Blog: NVIDIA Vera Rubin POD (March 16, 2026) (opens in a new tab)
- NVIDIA Newsroom: Vera Rubin ramps into full production (May 31, 2026) (opens in a new tab)
- NVIDIA Blog: first production Vera Rubin cluster (September 30, 2026) (opens in a new tab)
- NVIDIA Newsroom: NVIDIA unveils Rubin CPX (September 2025) (opens in a new tab)
- NVIDIA GB300 NVL72 product page (opens in a new tab)
- OEM product guide: GB300 NVL72 rack-scale AI (opens in a new tab)
- SemiAnalysis: Vera Rubin, Extreme Co-Design (February 25, 2026) (opens in a new tab)
- Tom's Hardware: Nvidia announces Rubin GPUs in 2026, Rubin Ultra in 2027 (March 18, 2025) (opens in a new tab)
- Tom's Hardware: GTC 2026 Ian Buck press Q&A transcript (March 23, 2026) (opens in a new tab)
- Tom's Hardware: Price of Vera Rubin NVL72 racks (March 24, 2026) (opens in a new tab)
- Tom's Hardware: Morgan Stanley VR200 NVL72 cost estimate (May 21, 2026) (opens in a new tab)
- Tom's Hardware: Nvidia's Kyber rack for Rubin Ultra reportedly slips to 2028 (July 6, 2026) (opens in a new tab)
- Tom's Hardware: Nvidia reportedly warns of 15% price hikes on AI servers (August 23, 2026) (opens in a new tab)
- Tom's Hardware: Nvidia expects US$20 billion of Vera Rubin sales in Q3 FY2027 (August 27, 2026) (opens in a new tab)