GPU cloud · NVIDIA GB300 NVL72
Dedicated NVIDIA GB300 NVL72 racks, in the Middle East and Asia
72 Blackwell Ultra GPUs and 36 Grace CPUs on one NVLink fabric, as a whole rack reserved for you. FlexiCloud sources it, delivers it and keeps the hardware running; you run your own stack.
Pricing on request. Minimum one full rack of 72 GPUs on a one-year commitment.
The numbers that matter
- Blackwell Ultra GPUs on one NVLink fabric
- 72
- HBM3e GPU memory per rack
- 20 TB
- NVLink bandwidth across the rack
- 130 TB/s
- Minimum order, one-year term
- 1 rack
Built for
Where a whole rack earns its place
An NVL72 rack makes 72 GPUs behave like one very large GPU. That matters for the biggest models and the heaviest reasoning workloads, and is overkill for jobs that fit on eight.
Frontier training
Train models too big for an 8-GPU server
Every GPU in the rack talks to every other at full NVLink speed, so tensor and expert parallelism can span 72 GPUs without dropping to the slower network between servers.
- Trillion-parameter and mixture-of-experts training
- Long-context pretraining and post-training
- One NVLink domain instead of many small ones
Reasoning inference
Serve the largest models at scale
Blackwell Ultra adds faster attention and more FP4 compute than Blackwell. NVIDIA reports up to 50 times the AI factory output of Hopper on reasoning models such as DeepSeek-R1.
- Very large mixture-of-experts serving
- Long-context, multi-step reasoning
- High throughput per megawatt
When to pick it
Choose the rack when eight GPUs are not enough
If your model fits on an 8-GPU server, a B300 node uses the same Blackwell Ultra GPU at a far smaller commitment, in India, the EU and the US. Pick the NVL72 when the model needs more than eight GPUs to act as one.
- Same GPU as the B300, 72 to a fabric
- Grace CPUs share memory with the GPUs
- Compare with a B300 node before you commit
What we run, and what you run
The rack is yours. The hardware is our job.
A liquid-cooled rack drawing over 100 kW needs a data centre built for it. Our partners have that; you get the rack, without building the facility.
Full control of the stack
Root access to all 18 compute trays to run your own OS image, CUDA version, container runtime and scheduler. The software is entirely yours.
Watched around the clock
GPU health, thermals, coolant, memory errors and the NVLink fabric are monitored continuously, so a failing tray is caught before it stalls a run.
An engineer, not a queue
Support is a person who runs GPU infrastructure. When something is wrong you talk to someone who can fix it.
Middle East and Asia
Racks in partner data centres across the Middle East and Asia, built for liquid cooling and rack-scale power. Tell us where your data has to live.
Yours alone
A dedicated, single-tenant rack. No shared GPUs, no noisy neighbours, no scheduler deciding when your job runs.
Predictable cost
A fixed term at a fixed rate, confirmed in writing. No spot-market surprises and no bill that doubles because a job ran long.
Terms
How it is sold
This is a dedicated rack on a term, not an hourly instance. The terms are simple and the same for everyone.
Minimum order
One rack: 72 × Blackwell Ultra
The unit is a full GB300 NVL72 rack. All 72 GPUs share one NVLink fabric, which is why the rack, not a slice of it, is the unit.
- 72 Blackwell Ultra GPUs, 36 Grace CPUs
- 20 TB HBM3e and 17 TB LPDDR5X per rack
- 130 TB/s NVLink across the rack
- Dedicated, single tenant, liquid-cooled
Pricing and term
Pricing on request
Rack pricing depends on region, term length and how many racks, so we quote each one. One-year minimum; longer terms cost less per GPU-hour. Billed upfront for the term, with the finer details agreed case by case.
- Firm quote in writing before you commit
- One-year minimum commitment
- Multi-rack clusters priced on request
Availability
The Middle East and Asia
Racks are allocated per region on a first-committed basis. Tell us where the data lives and where the team sits, and we will confirm availability and lead time before you commit to anything.
- Middle East
- Asia
- Lead time confirmed per region before order
- Need India, the EU or the US? See the B300
How it works
From first conversation to a running rack
- 01
Tell us the workload
Training or inference, the model size, how many racks, and which region. Fifteen minutes with an engineer tells us whether a rack is the right fit or a B300 cluster would do.
- 02
We confirm capacity and price
A firm quote for your region and term, with the lead time to a live rack. No commitment until you have the number in writing.
- 03
We hand over the rack
A dedicated rack, networked and ready to log into with root access. You bring your own OS image, drivers and scheduler; the hardware underneath stays ours to keep running.
- 04
You run. We keep it healthy.
Your team owns the stack and the workload. Ours watches the hardware and cooling, handles failures, and is a phone call away for the whole term.
Questions
What people ask before they commit
A liquid-cooled rack that joins 72 NVIDIA Blackwell Ultra GPUs and 36 Grace CPUs into one system. Each GB300 superchip pairs one Grace CPU with two Blackwell Ultra GPUs, and NVLink connects all 72 GPUs so they work as one very large GPU for training and inference.
Per rack, NVIDIA publishes: 72 Blackwell Ultra GPUs and 36 Grace CPUs with 2,592 Arm Neoverse V2 cores; 20 TB of HBM3e GPU memory at up to 576 TB/s, up to 288 GB per GPU; 17 TB of LPDDR5X CPU memory, 37 TB of fast memory in all; 130 TB/s of fifth-generation NVLink; 1,440 petaflops of FP4 with sparsity (1,080 without); and ConnectX-8 networking at 800 Gb/s per GPU. NVIDIA notes the figures may be revised.
One full rack of 72 GPUs. The GPUs share one NVLink fabric, so the rack is the unit of allocation. More capacity is added in whole racks, joined over InfiniBand or Ethernet.
One year. This is a dedicated rack reserved for you, not an hourly instance, and longer terms cost less per GPU-hour. The term is billed upfront, with the finer details agreed case by case and confirmed in writing before you commit.
We quote each rack, because the price depends on region, term length and how many racks you need. Tell us the region and term and we will come back with a firm figure in writing before you commit.
NVIDIA does not publish a price. Analysts at Loop Capital estimated a GB300 NVL72 rack at US$3.7 million to US$4.0 million in 2025, roughly US$51,000 to US$56,000 per GPU (our arithmetic). Morgan Stanley put the liquid-cooling parts alone at about US$50,000 a rack. On top come networking, installation and a data centre that can power and cool a rack of over 100 kW.
Both use the same Blackwell Ultra GPU. An HGX B300 server has 8 GPUs on one NVLink fabric; the GB300 NVL72 puts 72 on one fabric and adds Grace CPUs that share memory with the GPUs. Pick the rack when your model needs more than eight GPUs to act as one, such as the largest mixture-of-experts models; otherwise B300 nodes are a far smaller commitment, and we offer them in India, the EU and the US.
In the Middle East and Asia, in partner data centres built for liquid-cooled, rack-scale power. Allocation is tight, so tell us your region early and we will confirm availability and lead time before you order.
It is fully liquid-cooled and draws in the region of 120 to 140 kW per rack; NVIDIA describes it as a 120 kW rack. You do not need that facility yourself: the rack runs in our partner’s data centre, and we keep the hardware and cooling healthy.
We deliver the rack as dedicated bare metal with network and root access to its compute trays, and keep the hardware running: continuous monitoring of GPUs, cooling and the NVLink fabric, hardware failure handling, and an engineer you can reach directly. The software stack on top, from the OS and drivers to the scheduler, is yours.
No. We work with partners who own the racks and hold the capacity in each region. FlexiCloud sources the rack, delivers it to you dedicated, and keeps the hardware running for the whole term. The software you run on top is yours.
Go deeper
More on rack-scale GPUs
What NVL72 means, how it compares with an 8-GPU server, and the interconnect that makes 72 GPUs one machine.
Hopper vs Blackwell: NVIDIA’s GPU generations compared
NVIDIA’s Hopper (H100, H200) and Blackwell (B200, B300) generations compared: memory, architecture, interconnect, power and price, and a practical guide to which one actually fits your AI workload.
Read itGB200 vs B200, GB300 vs B300: what Grace Blackwell and NVL72 mean
B200 and B300 are GPUs; GB200 and GB300 pair them with NVIDIA’s Grace CPU and usually come as 72-GPU NVL72 racks. What that means, side by side, and who needs which.
Read itPCIe, SXM, HGX, DGX and NVL72: NVIDIA GPU form factors explained
PCIe, SXM, HGX, DGX, NVL and NVL72 explained in plain language, with a table of which form the H100, H200, B200, B300, GB200, GB300, RTX PRO 6000 and Vera Rubin come in.
Read itNVLink, NVSwitch and InfiniBand: the fabric that makes eight GPUs act as one
Buying GPUs is easy; making eight of them act as one is a question of fabric. How NVLink and NVSwitch scale up inside a node, how InfiniBand scales out between nodes, why a GPU that waits is money burning, and why a FlexiCloud B300 node is a single NVLink-5 domain.
Read it
Tell us the model and the region
Fifteen minutes with an engineer gets you a firm price, a lead time, and a straight answer on whether a whole rack is the right fit. No commitment until you have the number in writing.