Vera Rubin explained: what changes and when it matters to buyers

Vera Rubin is NVIDIA's next GPU platform after Blackwell, and it is shipping: NVIDIA says production shipments began in August 2026. For most buyers it matters from 2027, because early volume is limited and goes to the largest buyers first. When it does arrive, it brings about 2.4 times Blackwell's memory bandwidth per GPU in NVIDIA's NVL72 rack (about 2 times on NVIDIA's figure for its 8-GPU servers). In the first peer-reviewed benchmark, run on racks, it did about twice the work per GPU. Its rack systems use compute trays with almost no internal cables, and prices are higher.

You will hear a lot of big numbers about Rubin. This guide sorts them into published specs, NVIDIA's own claims and analysts' estimates, and answers the questions buyers ask us. Figures are as of 24 October 2026. NVIDIA has revised several Rubin specs since launch, so we use its current product pages and say where earlier figures differ.

The short answers

What is NVIDIA Vera Rubin?

Vera Rubin is the name of NVIDIA's whole platform for 2026 and 2027, after the American astronomer Vera Rubin. At CES in January 2026 NVIDIA presented it as six new chips designed to work together:

• Rubin GPU: the AI processor, with 288 GB of HBM4, the newest generation of high-bandwidth memory.
• Vera CPU: NVIDIA's own processor, which replaces the Grace CPU in rack systems.
• NVLink 6 switch: links GPUs so they can share memory and work as one.
• ConnectX-9 SuperNIC: the network card that joins racks into clusters, at up to 1.6 Tb/s (terabits per second) per GPU in NVL72 racks. NVIDIA's 8-GPU DGX Rubin NVL8 stays at 800 Gb/s per GPU, as on the DGX B300.
• BlueField-4 DPU: a data processing unit that takes networking, storage and security work off the CPUs.
• Spectrum-6: an Ethernet switch chip with the optics built into the package (co-packaged optics), at 102.4 Tb/s.

In March 2026 NVIDIA added a seventh chip, the Groq 3 LPU (language processing unit), for separate racks aimed at very fast responses.

You will meet Rubin in two main forms:

• Vera Rubin NVL72: a liquid-cooled rack with 72 Rubin GPUs and 36 Vera CPUs on one NVLink network. NVIDIA first called it "NVL144", when it counted each GPU's two dies separately. It is the same hardware.
• HGX Rubin NVL8: an 8-GPU server built by server makers, with x86 host processors the maker chooses, like today's HGX B300 servers. NVIDIA's own DGX Rubin NVL8 uses two Intel Xeon processors, as the DGX B300 does. A version hosted by one Vera CPU also exists.

Our guide to GPU form factors explains HGX, DGX and NVL, and our NVL72 explainer covers the Blackwell racks that Vera Rubin replaces. To register early interest in Rubin nodes, see our Vera Rubin page.

Vera Rubin vs Blackwell: what changes in the hardware?

Four things change for buyers. Memory bandwidth per GPU rises about 2.4 times in the NVL72 rack, and about 2 times on NVIDIA's figure for 8-GPU servers. Dense low-precision compute (the FP4 and FP8 number formats used to run models) rises about 2.3 times in FP4 and 3.5 times in FP8, against a GB300 GPU. NVLink rises about 1.7 times, and so does power per 8-GPU server. Memory capacity stays at 288 GB per GPU, the same as the B300.

An 8-GPU server: HGX B300 against HGX Rubin NVL8

This is the format we source for customers. Figures are NVIDIA's published peak specs for each 8-GPU system. FP4, FP8 and BF16 are number formats for AI maths: the fewer the bits, the faster the GPU runs. FP4 and FP8 are compact formats that run models fastest; BF16 is a common format for training. FP64 is the high-precision format used in science. "Dense" figures count no shortcuts. NVIDIA's headline figures are often "sparse": they count a shortcut that only helps when a model has been pruned in a particular pattern.

Source: NVIDIA's HGX and DGX pages, read 24 October 2026. DGX is NVIDIA's own-brand version of the 8-GPU server. PF is petaflops and TF teraflops. TB/s is terabytes per second; network speeds are in Gb/s or Tb/s (gigabits or terabits per second), and 1 Tb/s is one eighth of 1 TB/s. Ratios are our arithmetic.

Two things stand out. First, NVIDIA's own table lists 2 TB of HBM4 for 8 Rubin GPUs, slightly less than the 2.1 TB it lists for the HGX B300. Server makers list 2.3 TB (8 times 288 GB). If your model does not fit on a B300 node today, Rubin will not change that; see our GPU memory guide for sizing. Second, power rises by about 70%, and every 8-GPU Rubin server design we found cools its GPUs with liquid. That narrows the data centre halls that can take them.

A rack: GB300 NVL72 against Vera Rubin NVL72

Source: NVIDIA's Vera Rubin NVL72 and GB300 NVL72 pages and the Vera Rubin NVL72 datasheet (October 2026). GB300 rack power is from press and analyst estimates, as NVIDIA does not list it; the cooling share is Huang's figure. Ratios are our arithmetic.

NVIDIA's headline 50 petaflops per Rubin GPU is the sparse figure; the dense figure is 35. We compare dense with dense throughout.

The Vera CPU

Vera has 88 NVIDIA-designed Olympus cores and 176 threads, against 72 cores for Grace. NVIDIA lists up to 1.5 TB of memory per CPU at 1.2 TB/s (Grace: 480 GB at 512 GB/s). Its link to the GPUs runs at 1.8 TB/s, twice Grace's 900 GB/s. That memory matters for AI serving, because the GPUs can park the working memory of long conversations (the KV cache) in it. One caution: TrendForce and SemiAnalysis report that early systems ship with about half that CPU memory, because of the memory shortage.

Specs NVIDIA has revised

Read the current product page, not the launch slides. NVIDIA has changed two headline figures without explaining why:

• Memory bandwidth per GPU: 13 TB/s when Rubin was first shown in March 2025, up to 22 TB/s at CES in January 2026, and 19.2 TB/s on NVIDIA's Vera Rubin NVL72 page and October 2026 datasheet today. NVIDIA's HGX page still shows 22 TB/s in its per-GPU table, while its 8-GPU system figure, 130 TB/s, works out to about 16 TB/s per GPU, about twice Blackwell's 8 TB/s (our arithmetic). Even the DGX Rubin NVL8 page disagrees with itself: its text says 176 TB/s for 8 GPUs (22 each), its spec table 130 TB/s.
• NVLink 6 per GPU: 3.6 TB/s at CES (260 TB/s per rack), and 3 TB/s (216 TB/s per rack) on NVIDIA's current US NVLink and Vera Rubin NVL72 pages and its October 2026 datasheet. Only the NVLink page marks the figure as preliminary. NVIDIA's HGX page (in its per-GPU table) and some of its regional NVLink pages, including India's, still show 3.6 TB/s.

One product is off the 2026 plan: Rubin CPX, a lower-cost GPU meant for reading very long inputs, was pulled at GTC in March.

How much does Vera Rubin cost, and how does pricing change?

Up, roughly double per rack. NVIDIA publishes no prices, but analysts estimate a Vera Rubin NVL72 rack at about US$7.8 million (Morgan Stanley) to US$9.1 million (Bernstein), against Morgan Stanley's US$4 million for a GB300 NVL72. By 24 October 2026 nobody had reported a price for an 8-GPU Rubin server. No hyperscaler (the largest cloud companies, such as AWS, Microsoft and Google) listed a Rubin rental price either. Estimates vary widely because they measure different things: the cost of parts, a server maker's price or the price to a hyperscaler.

On Morgan Stanley's figures, the GPUs alone cost about 58% more per rack: nearly US$4 million against about US$2.5 million. Compare only figures from the same analyst: estimates for a GB300 NVL72 run from US$4 million (Morgan Stanley) to US$8.6 million, depending on the source and what it measures. Bernstein says the widely quoted figure of about US$8 million for a Rubin rack used outdated memory prices.

Why it costs more: memory

Memory is now the biggest driver. Morgan Stanley estimates memory at about a quarter of a Vera Rubin rack's cost, up from about 9% in a GB300 rack. Bernstein puts memory and storage together at about 35%. And memory is getting dearer. TrendForce forecasts the average HBM price to rise 121% in 2027. Industry estimates put a 36 GB HBM4 stack at about US$600 in 2026 and about US$1,300 in 2027. Michael Dell said in October 2026 that memory, not demand, is the industry's bottleneck, and that it gets worse in 2027.

NVIDIA appears to be passing this on. In August 2026 Bloomberg, citing unnamed sources, reported that NVIDIA had told some of its biggest customers of price rises, above 15% in many cases. The rises cover Grace Blackwell and Vera Rubin systems shipping from early 2027; NVIDIA did not comment. On its results call the same month, NVIDIA's finance chief said memory prices were above expectations. Price increases NVIDIA had already made take effect in the first quarter of its 2028 financial year (about February to April 2027); products and amounts were not named.

Cheaper per unit of compute, not per box

A Rubin rack costs more but does more. On Morgan Stanley's figures, a Vera Rubin rack costs about 1.95 times as much as a GB300 rack. It delivers about 2.3 times the dense FP4 compute, so it costs about 16% less per unit of dense FP4 compute (our arithmetic; the result moves a lot with the GB300 price used). Against the measured MLPerf results further down, the same price gap works out to anywhere from about 12% more to about 24% less hardware cost per token, depending on the test (our arithmetic, hardware price only).

NVIDIA's own claim is far bigger: one tenth the cost per million tokens (the word pieces a model reads and writes). But that is a projection for one large reasoning model, Kimi-K2-Thinking, compared with the older GB200 NVL72. Real savings depend on your model and on how fast your users need answers.

Blackwell is not getting cheaper

Normally a new generation pushes the old one's prices down. H100 rental prices at specialist GPU clouds fell from peaks of about US$8 an hour in 2023 to US$1.50 to US$3.00 by late 2025, over the period in which Blackwell arrived (Silicon Data). In 2026 that pattern broke. AWS's published price for reserved 8-GPU B300 capacity blocks in Oregon rose about 38% between December 2025 and October 2026. Its B200 price in Ohio rose about 52%. One public price tracker shows the same direction: on-demand list prices for 17 of the 21 GPU models it follows, including the H100, were higher in October 2026 than a year earlier.

SemiAnalysis estimated in September 2026 that three-year rental rates in July 2026 were over US$8.50 per GPU-hour for Rubin. That compares with about US$5 per GPU-hour in a GB300 NVL72. But SemiAnalysis estimates Rubin gives about 62% more output per rental dollar, at a typical speed of 80 tokens per second per user. In short: expect to pay more per GPU-hour for Rubin, and judge it on cost per result, not on the hourly rate.

How is Vera Rubin built differently from Blackwell?

The biggest change is inside the NVL72 compute tray, the flat drawer that holds the chips. Its fans and almost all its internal cables are gone, and its liquid tubes drop from six to two. NVIDIA says robots now build one in about five minutes, against about two hours by hand for Grace Blackwell. The rack around it changes much less.

Inside the tray: almost no cables, fewer hoses, no fans

Each Vera Rubin NVL72 compute tray holds two Vera CPUs and four Rubin GPUs. On a Grace Blackwell tray, dozens of internal cables (43 by Huang's count; Fast Company, which saw both trays side by side, counted 50 or 60) and six liquid tubes had to be fitted by hand. On Vera Rubin, a circuit board across the middle of the tray (a midplane) and board-to-board connectors replace the internal signal cables. Quick-release coolant connectors cut the tubes from six to two. The fans are gone too, because everything on the tray is liquid-cooled, and going fanless is what freed the space for the midplane.

NVIDIA's figures for the gain:

• Build time: by robots, not by hand. NVIDIA's own figures vary: from over 1.5 hours to 3 hours for Grace Blackwell, and from one to five minutes for Vera Rubin.
• Right first time: Andrew Bell, NVIDIA's senior vice president of hardware engineering, told Fast Company that only about 20% of early Grace Blackwell trays were right first time. For Vera Rubin, Bell put the figure at about 95%.
• Servicing: NVIDIA's launch release said assembly and servicing are up to 18 times faster; later NVIDIA documents say 20 times and, in its October 2026 datasheet, 90 times. The German tech site heise, citing NVIDIA, reported that swapping an NVLink switch tray takes about six minutes, against 100 minutes on Grace Blackwell.

These are vendor figures that nobody has measured independently. The direction still matters: cables were the main failure point when Grace Blackwell trays were built, and they slowed that rollout (SemiAnalysis, Tom's Hardware). A tray that is quicker to build and repair should mean faster ramps and shorter downtime when a part fails.

The rack: an evolution, not a revolution

Vera Rubin NVL72 uses the same rack footprint as the GB300 NVL72. Both follow NVIDIA MGX, the standard rack design that partners build to. So a data centre that took Grace Blackwell racks can plan for Rubin in a similar way, with more power and more coolant flow (SemiAnalysis estimates about 2 to 2.5 times Blackwell's). Each rack holds:

• 18 compute trays and 9 NVLink switch trays;
• a copper NVLink "spine" at the back: about 5,000 cables in four cartridges, more than 3 km of cable in total;
• four 110 kW power shelves fed with AC power, which feed a 54 V DC busbar (a solid power rail) up the rack;
• liquid cooling with coolant at up to 45°C.

Power is the real change: up to 227 kW per rack at full power, or 156 kW in NVIDIA's lowest-power profile, against an estimated 135 to 150 kW for a GB300 NVL72.

The rack's copper NVLink spine goes away only with Kyber, the rack NVIDIA designed for Rubin Ultra, the higher-end follow-on to Rubin. Kyber uses vertical blades, a circuit-board midplane instead of the spine, and 800 V DC power. NVIDIA's roadmap shows it in 2027. SemiAnalysis reported in July 2026 that it has slipped into 2028; NVIDIA has not confirmed that and says its roadmap is intact.

Who builds it

More of the building moves to NVIDIA and a few large manufacturers. Server makers have traditionally bought NVIDIA's GPU boards and built their own trays. For Vera Rubin, J.P. Morgan reported in November 2025 that NVIDIA would ship fully built and tested compute trays (what the industry calls "L10"). Partners would then assemble and test whole racks ("L11"). NVIDIA has not confirmed it. SemiAnalysis says only three manufacturers, Foxconn, Quanta and Wistron, have automated tray lines. NVIDIA's partner list for Vera Rubin systems includes Dell, HPE, Lenovo, Supermicro, Foxconn, Inventec, QCT, Wistron, Wiwynn, Pegatron and Compal, across more than 350 factories in 30 countries.

If the L10 reports are right, Rubin racks from different brands will be more alike than before. Server makers will then compete more on delivery, integration and support than on design. NVIDIA has also announced DSX POD, a design for prefabricated AI factories in containers that partners build and test before shipping. It says deployment at a prepared site can take as little as 14 days instead of six months, with first deliveries planned for early 2027.

8-GPU Rubin servers

The 8-GPU HGX Rubin NVL8 servers that we would source are built by server makers, not NVIDIA, and designs differ. All five we found, from Supermicro, ASUS, Compal, Wiwynn and Gigabyte, are 2U or 3U servers (two or three standard rack units tall) with liquid-cooled GPUs. Compal lists about 24 kW per server, in line with NVIDIA's figure for its DGX Rubin NVL8. Some are cable-free inside, such as an ASUS design with board-to-board connections; others, such as a Gigabyte server shown at GTC, still use cables for the drives. One Supermicro-built system is sold only as a complete, integrated rack, so not every design can be bought as a single node.

Is Vera Rubin the "wireless" server Jensen Huang talked about?

Yes, almost certainly, but "wireless" here means cable-free inside the tray, not radio. Huang has shown the Vera Rubin compute tray at keynote after keynote since October 2025, and stressed the missing cables each time:

Quotes are from third-party transcripts of the keynotes. We found no record of Huang calling the server "wireless". The word most likely comes from Huang's "no wires" line in Berlin. It also appeared in coverage of the CES talk: Taiwan's Economic Daily News used a Chinese term that translates as "going wireless".

What is still cabled or plumbed:

• The NVLink spine: about 5,000 copper cables at the back of each rack, linking all 72 GPUs.
• Power: AC power feeds into each power shelf, whose power supplies are still air-cooled.
• Tray connections: NVIDIA's Andrew Bell told Fast Company the "cableless" tray still has two cables, one to power and one to the network. Each tray is also cabled to the cluster network and the management network.
• Liquid: quick-release couplings from each tray to the rack's coolant manifolds.

The benefit is faster building and servicing, not speed: a cable-free tray does not make the GPUs compute any faster. The claim applies only to the NVL72 compute and switch trays. NVIDIA's HGX page makes no cable-free claim for 8-GPU servers, and those vary by maker, as above.

Do not confuse it with NVIDIA's actual wireless work: AI-RAN, computers for 6G mobile networks. NVIDIA is building these with Nokia, and announced a US$1 billion investment in the company in October 2025. Huang covered that in some of the same keynotes. At GTC Washington, Huang described that telecom computer as able to "communicate wirelessly", so if you heard "wireless" there, that may be what was meant.

How much faster is Vera Rubin than Blackwell, and why?

About twice the work per GPU in the first peer-reviewed benchmark. In MLPerf Inference v6.1 (September 2026), a Vera Rubin NVL72 rack did 1.7 to 2.6 times the DeepSeek-R1 work per GPU of a GB300 NVL72. No result for an 8-GPU Rubin server has been published yet. The gain comes from four places, roughly in order of importance for serving AI models:

1. Memory bandwidth: writing each word (token) of an answer means reading the model's weights from memory, so speed is often limited by memory bandwidth, not compute. Rubin's is about 2.4 times Blackwell's in the rack.
2. Low-precision compute: about 2.3 times the dense FP4 compute of a GB300 GPU, and about 3.5 times the FP8. Gains in BF16, a common format for training, are smaller: about 1.8 times per 8-GPU server.
3. NVLink 6: about 1.7 times the GPU-to-GPU bandwidth. This matters most for mixture-of-experts (MoE) models, which split each layer into many "experts" spread across GPUs that constantly exchange data.
4. Rack-scale serving: the larger CPU memory holds the KV cache of long conversations, and BlueField-4 adds a fast storage tier for it. NVIDIA's Dynamo software also splits the work, so some GPUs read the prompt while others write the answer.

NVIDIA's claims, and what they compare against

The first benchmark results

The first peer-reviewed numbers came on 16 September 2026, in MLPerf Inference v6.1, the industry benchmark run by MLCommons. Vera Rubin NVL72 was entered in the preview category, for systems not yet on general sale, by NVIDIA and by one GPU cloud. The submitters run the tests themselves, and the results are then peer-reviewed.

So a Vera Rubin rack did about twice the work per GPU of a GB300 rack, and the gap was widest when users need fast answers. On Qwen3-VL-235B, a large vision model run in BF16 on both systems, the gains were 1.8 to 3.7 times. Three cautions:

• Power: each Rubin GPU is rated at up to 2,300 W, against 1,400 W for a GB300 GPU. Per rated watt, the DeepSeek-R1 gain shrinks to about 1.05 to 1.6 times (our arithmetic). MLPerf did not measure power in these runs.
• Blackwell keeps improving: software alone made GB300 NVL72 1.6 to 2.9 times faster on DeepSeek-R1 between MLPerf v5.1 and v6.1, a year apart. Rule changes may explain part of that.
• Claims shrink in practice: NVIDIA claimed 1.5 times the dense FP4 compute for GB300 over GB200; in MLPerf, throughput rose 1.45 times offline and only about 1.25 times in the server test.

Nobody has yet published a training benchmark for Rubin, or any result for an 8-GPU Rubin server. In MLPerf v6.1 the Rubin rack did about 1.9 to 2.2 times the work per GPU of NVIDIA's 8-GPU DGX B300. That compares a 72-GPU rack with an 8-GPU node, at about twice the rated power per GPU (our arithmetic). And the best 8-GPU B300 server beat the GB300 rack per GPU in the server test: 8,847 against 8,291 tokens per second. So for steady, high-volume serving that fits in eight GPUs, a B300 node keeps pace with a GB300 rack.

Who gains the most

The biggest gains are for large MoE reasoning models with long inputs, served at fast response speeds across a full NVL72 rack. Expect less on an 8-GPU server, on smaller models, on batch work and on BF16 training. FP4 gains need your model converted to NVIDIA's 4-bit format, NVFP4. NVIDIA reports an accuracy loss of 1% or less on DeepSeek-R1, but test your own model. For double-precision science (FP64), Rubin brings back what the B300 cut. Its native FP64 is slightly below the B200's (33 against 37 TF per GPU), but NVIDIA lists 200 TF per GPU for FP64 matrix maths (DGEMM) using emulation on the Tensor Cores. Test your own code before you count on that. Our Hopper vs Blackwell guide covers how to match a GPU to a workload.

When is Vera Rubin available, and when does it matter to buyers?

For most buyers, from 2027. NVIDIA says production shipments began in August 2026, and Google Cloud plans general availability of its Rubin instances in November 2026, but early volume is limited and goes to the largest buyers first.

Supply stays tight

NVIDIA's finance chief said in August 2026 that supply will remain a bottleneck at least until the end of NVIDIA's 2028 financial year, in late January 2028. TrendForce expects Blackwell to make up about 70% of NVIDIA's high-end GPU shipments in 2026. It sees the GB300 as the main driver until mid-2027, and Vera Rubin as the largest share during 2027. Most of the large Rubin deployments announced outside the hyperscalers land between 2027 and mid-2028. An anonymous claim on X in October 2026 that Rubin shipments had been cut by half has not been confirmed by any reputable source.

For Blackwell, about 10 to 14 months passed between NVIDIA's announcement in March 2024 and instances at a US GPU cloud (February 2025) and at AWS (May 2025). Rubin looks similar if you count from its CES launch in January 2026: limited, by-request access came about nine months later, and Google Cloud's planned general availability about ten months later (our arithmetic). TrendForce's forecasts point to broader supply during 2027. We expect wider availability through our partner network in 2027. That is our estimate, not an NVIDIA date.

Your software

Existing CUDA code (software written for NVIDIA's GPU programming platform) should run. NVIDIA promises full backward compatibility, and its CUDA 13.4 release notes (October 2026) list Rubin support. Support in popular serving frameworks arrives as the platform ships, so check your stack's release notes. The real work is FP4: to get the headline gains, models need converting to NVFP4 and checking for accuracy.

Your data centre

Rubin needs liquid cooling. Plan for about 24 kW per 8-GPU server, or 156 to 227 kW per NVL72 rack, with coolant at up to 45°C. NVIDIA says the racks can run on dry coolers, without chillers, in normal conditions up to about 38°C outside. It also says chillers stay important in hot weather, and Schneider Electric says chillers usually remain necessary in hot, humid parts of India. And the bar is high. In May 2025, one Indian operator said its sites in NVIDIA's DGX-Ready programme for liquid cooling support up to 200 kW per rack. That is below a Vera Rubin rack at full power. Our India data centre power guide and guide to hosting GPUs in India cover what to check.

Should you wait for Vera Rubin or rent Blackwell now?

If you need GPUs in the next 6 to 12 months, rent Blackwell now. Rubin volume is limited and goes to the largest buyers first, and prices are rising for both generations, so waiting does not guarantee cheaper compute.

One more point on contract length: NVIDIA now releases a new generation about every year, so Rubin bought in 2027 will be a generation behind within about a year. Match the term to how long your workload needs the hardware. Our guides on renting, buying or using an API and questions to ask a GPU provider help with that decision.

How FlexiCloud can help

We source dedicated 8-GPU B300 and B200 nodes, as single-tenant bare metal, across our partner network in India, the EU and the US; see GPU server prices. For Rubin, register early interest. It costs nothing and commits you to nothing, and we will come back with pricing and availability as our partners confirm allocation. We do not rent NVL72 racks. If you are not sure whether to wait, book a call, tell us your workload and timeline, and we will tell you honestly whether Blackwell fits now.

Where these figures come from

Specs: NVIDIA's Vera Rubin NVL72, HGX, NVLink, Vera CPU, DGX Rubin NVL8 and GB300 NVL72 pages (read 24 October 2026), the Vera Rubin NVL72 datasheet (October 2026), NVIDIA's DGX Vera Rubin user guide and NVIDIA's technical blog posts of January, March and July 2026. Benchmarks: MLCommons MLPerf Inference v6.1 (16 September 2026) and SemiAnalysis. Prices: Morgan Stanley and Bernstein estimates as reported by Tom's Hardware and others, Reuters, Bloomberg, TrendForce, Silicon Data, a US GPU appraisal firm and AWS's published Capacity Blocks prices. Timeline and supply: NVIDIA's results call of 26 August 2026, NVIDIA, Microsoft and Google Cloud announcements, and TrendForce. Assembly: NVIDIA, Fast Company, heise, SemiAnalysis and server makers' announcements. Keynote quotes are from third-party transcripts. Ratios marked "our arithmetic" are calculated from these published figures.