Technology · Analysis, October 2026
NVIDIA Vera Rubin, explained.
The biggest change in AI hardware since the H100. What it is, what it really does, what it costs to run in Australia, and how to run one in your own building.
- HBM4 memory per GPU
- 288 GB
- Memory speed per GPU
- 22 TB/s
- Inference per GPU, NVFP4
- 50 PF
- GPUs acting as one, per rack
- 72
- Fast memory per rack
- 20.7 TB
- Power per rack
- 190–230 kW
The data centre is the computer.
For twenty years the unit of computing was the server. With Rubin, NVIDIA makes it the whole room.
Vera Rubin is five kinds of rack, designed to work as one machine. The NVL72 joins 72 Rubin GPUs and 36 Vera CPUs so tightly they behave like a single giant GPU. Around it sit racks for running an agent’s tools, for writing answers fast, for storage, and for the network that ties them together.
The reason is agents. An agent doesn’t answer one question. It reads, plans, calls tools, checks its work and goes again, for minutes or hours. That needs every part of the room working in step.
Memory, not maths.
The headline number is compute: 50 petaflops per GPU, five times Blackwell. It is the least important number.
Writing an answer means reading the whole model from memory for every word. So the speed limit is memory, not maths. Each Rubin GPU reads its memory at 22 terabytes a second, nearly three times Blackwell and more than six times the H100.
Agents make this worse. They carry long histories, documents and tool results, all held in memory while they work. Rubin is built for that load.
Memory speed per GPU
Two chips for one answer.
Every answer has two halves. First the model reads the whole prompt at once. That is heavy computing. Then it writes the answer one word at a time. That is a race to fetch from memory.
Rubin splits them. Rubin GPUs do the reading. New Groq 3 chips, from NVIDIA’s US$20 billion deal with Groq, do the writing from very fast memory built into the chip. It is NVIDIA admitting that no single chip is best at both.
What’s claimed, and what’s been shown.
Rubin has been shipping for a few months. Most numbers are still NVIDIA’s. Here is how they stand up.
Claimed by NVIDIA
Ten times lower cost per token than Blackwell
Early independent analysis finds about 1.5 to 5 times lower cost than GB200, depending on how fast each user is served. The tenfold figure compares against Blackwell’s year-old software.
Claimed by NVIDIA
Five times the inference compute per GPU
A specification, so it holds on paper. Real work is usually limited by memory, which is why the bandwidth gain matters more.
Claimed by CoreWeave, analysed by SemiAnalysis
About 5.4 times the tokens per megawatt of GB200
Measured on a pre-production rack running DeepSeek R1, one prompt and one answer at a time. Long-context agent work, where Rubin should do best, has not been measured yet.
Claimed by Jensen Huang
No chillers needed
Plausible. The rack takes 45°C water, warm enough to cool with outside air for most of the year in southern Australia.
What it takes to run one.
The chip is ready. Most buildings are not.
190 to 230 kW a rack
A normal enterprise rack draws 5 to 10 kW. One Rubin rack draws as much as 40 homes.
All liquid-cooled
There is no air-cooled version. Water at 45°C goes straight to every chip.
About 1.8 tonnes
Most office floors and many older data halls cannot take the load.
800-volt DC power
NVIDIA is moving racks to high-voltage DC. Few Australian sites have it yet.
Rubin Ultra in 2027
The next rack, called Kyber, is expected to draw about 600 kW. Build for that now.
A grid queue
Data centres have asked for 66.9 GW of connections on the east coast, against about 2.2 GW today.
Australia has an advantage. Cool southern climates suit 45°C water cooling. NVIDIA has named Australia a 2 GW sovereign AI program with eight local operators. The constraint is the grid, so the sites with their own power will get Rubin first.
What Rubin costs to run in Australia.
Nobody publishes this, so we worked it out. Move the sliders to test your own assumptions. Every input says whether it is a published figure or our assumption.
Three things stand out.
The rack is most of the cost. About four-fifths, at our starting point. Electricity is under a tenth. Cheap power matters less than people think.
Utilisation is everything. A busy rack costs half as much per token as a half-idle one. Moving from 60% to 90% use cuts the cost by a third.
So most organisations should buy Rubin by the token. Owning a rack only pays if you can keep it busy for four years.
Cost to serve a million output tokens
A$10.53
345 billion tokens a year, at 18,240 tokens a second flat out.
- The rack itself
- A$2,825k · 78%
- Electricity
- A$240k · 7%
- Space and support
- A$570k · 16%
- Each year
- A$3.63m
One NVL72 rack, excluding GST, financing and software. Throughput is from one early benchmark on one model, so treat the result as a range, not a quote.
Rubin in your own building.
A full NVL72 rack belongs in a data centre. The 8-GPU Rubin server does not have to. Here are the three ways to run Rubin in Australia, under your control.
A Rubin server in your building
DGX or HGX Rubin NVL8. 8 GPUs, 2.3 TB of GPU memory, about 24 kW, liquid-cooled.
For organisations whose data must stay on their own premises.
It needs: A room that can take liquid cooling and 24 kW, or a small retrofit.
A private cage in an Australian data centre
Rubin servers or full NVL72 racks, in space only you can reach.
For organisations that want their own hardware without building for it.
It needs: A contract with the data centre. We handle the rest.
Rubin by the token
Shared capacity through our API, in Australia, billed per token.
For everyone else. No capital, no facility, no idle hardware.
It needs: An API key.
What we do
We check your site, size the system to your workload, buy through the server makers, install it, and run the models and software on it. You get private AI that never leaves your control, without building an AI team.
Our view.
Rubin is a real step forward, and the biggest gains are where they matter most: memory, and the cost of long agent work.
The tenfold claim is marketing. Two to five times is what the early evidence supports, and that is still a big number.
For most Australian organisations the question is not whether to buy Rubin. It is how to get its economics without its building.
For the few who must own their hardware, an 8-GPU Rubin server in a liquid-ready room is now a realistic project, not a data centre build.
Sources
- NVIDIA Vera Rubin platform
- Vera Rubin NVL72 specifications (Hashrate Index)
- Vera Rubin NVL72 vs GB200 NVL72 inference TCO (SemiAnalysis)
- Vera Rubin in full production (The Next Web)
- The Groq 3 LPU and the US$20 billion deal (Tom’s Hardware)
- No chillers needed for Vera Rubin (Facilities Dive)
- The data centre isn’t ready (Teradata)
- Rubin Ultra NVL576 at 600 kW (DatacenterDynamics)
- Vera Rubin rack price, Morgan Stanley estimate (Digital Citizen)
- DGX Rubin NVL8 specifications (Flopper)
- Australia as a 2 GW sovereign AI program (TechTimes)
Independent analysis by Continuum Labs, October 2026. NVIDIA, Vera Rubin, DGX, HGX and NVLink are trademarks of NVIDIA Corporation. Illustrations are our own. Figures are as published at the time of writing and will change.