NVIDIA in the UAE, designed and delivered by 10ⁿ Tech
The GPU is the easy part of a GPU cluster; the rack, the fabric and the six teams sharing sixteen cards are the design
NVIDIA presents itself as a full-stack data-centre platform for what it calls the era of AI reasoning: GPUs (Hopper, Blackwell and the Rubin generation), Vera CPUs, BlueField DPUs, Spectrum-X Ethernet and Quantum InfiniBand networking, rack-scale GB200 and GB300 NVL72 systems, HGX and DGX servers, NVIDIA AI Enterprise software with NIM microservices, Omniverse for simulation and digital twins, RTX PRO workstations for the desk, and Jetson for the edge. 10ⁿ Tech designs and delivers the parts of that stack a UAE enterprise or government platform actually orders: H200 and B300 nodes from Dell, Lenovo, HPE and Supermicro, the fabrics and storage around them, the Kubernetes layer with the GPU operator and Multi-Instance GPU, and the RTX PRO workstations our own digital twin studio runs on.
A word on scale. NVL72 racks, DGX SuperPODs and sovereign AI factories are sold through NVIDIA's largest partners and the hyperscalers; we are not that, and we tell customers who need that to go there. What we design and deliver is the two-to-sixteen-GPU estate: a federal platform on four HGX nodes, a training-centre cluster on PCIe H200 NVL nodes, a GPU-cloud pod for a regional operator, a BIM studio on RTX PRO 6000 workstations. The UAE is now on the list of countries with the most accessible US GPU exports, so buying is not the hard part. What the quote does not tell you is that the node draws 15 kW, that a rack built for 6 kW will not take two of them, that the lead time is ten to sixteen weeks, and that a sixteen-GPU estate has to serve development, testing, staging and production without the teams fighting over cards. That is our work, and we have published the numbers.
What we deliver with NVIDIA
| Product line | Typical use | How we deliver it |
|---|---|---|
| H200 HGX nodes (4 and 8 GPU) | Sovereign AI platforms, inference and fine-tuning on-premises | Sized to the model and the concurrency; power and cooling checked per rack before the node is quoted |
| H200 NVL, L40S and RTX PRO 6000 Server in PCIe servers | AI inference alongside standard workloads, GPU virtual desktops, rendering | Dell, Lenovo, HPE or Supermicro hosts with the GPU the software licence expects |
| B300 HGX systems | GPU-cloud and large training builds | Eight-GPU 8U nodes at about 16 kW operating; dual-mode ConnectX-8 NICs for RoCE at 2x400G per port |
| Spectrum-X Ethernet, Quantum InfiniBand, BlueField DPUs | Multi-node training, high-throughput inference, tenant isolation | Leaf-spine designs laid out per GPU rail; the leaf count follows the number of rails, not nodes, which changes the switch bill |
| NVIDIA AI Enterprise, NIM microservices, GPU operator, Multi-Instance GPU | Running and sharing the estate | Licensed per GPU; MIG partitions one Hopper GPU into up to seven isolated instances; we used it to run six environments on two non-production nodes |
| RTX PRO workstations and Omniverse | Digital twins, Unreal Engine, BIM, simulation | The workstations our own twin team builds on; Omniverse where a customer's twin has to take live data from more than one tool |
| Jetson edge modules | Vision and robotics at the edge, on-site inference | Specified inside industrial designs with the integrator building the machine; we handle the compute and the network |
| Storage for GPU estates | Model weights, datasets, checkpoints | NVMe tiers close to the GPUs, parallel file systems for training, object storage for the archive |
| Azure ND H200 v5 as the interim | Programmes that cannot wait for hardware | Reserved capacity in UAE North with the same Kubernetes deployment, moved on-premises when the racks are commissioned |
How we work with NVIDIA in Dubai
For a federal entity's sovereign AI platform we specified sixteen H200s across four four-GPU HGX nodes: two for production, one for UAT and pre-production, one carrying development, integration and a sandbox through MIG. Thirty-two kilowatts of usable IT power over two 48U racks on A and B feeds, air cooled, a 25 and 100 GbE fabric, and a hardened backup repository as the disaster-recovery answer. The article that came out of it is the most read thing on our site about AI.
For a GPU-cloud operator we designed the B300 build: eight-GPU HGX nodes, a per-rail RoCE fabric on 800G switching, BlueField DPUs for the tenant network, and the storage tiers behind it. The lesson worth repeating from that design is that the ConnectX-8 ports run as two 400G Ethernet lanes per 800G port, which decides which transceivers go on the switch bill.
For programmes that cannot wait for hardware, we put the first wave on a reserved ND H200 v5 instance in Azure UAE North, eight H200s with NVLink between them, and moved to the on-premises estate when the racks were commissioned. Same models, same Kubernetes deployment, two homes.
On the NVIDIA relationship itself: 10ⁿ Tech is a member of NVIDIA Inception, the programme for companies building on NVIDIA, and buys GPUs inside OEM configurations rather than as a direct NVIDIA reseller. We say that plainly because customers sometimes assume a badge that is not there.
Where it has gone
- A sixteen-H200 sovereign AI platform design for a federal entity: four HGX nodes, two racks, 32 kW, MIG for non-production
- A B300 HGX GPU-cloud build with a per-rail RoCE fabric and BlueField DPUs for a regional GPU-as-a-service operator
- Cloud-first delivery on Azure ND H200 v5 in UAE North ahead of an on-premises estate
- Dell R770 nodes with H200 NVL for a government AI training platform
- RTX PRO workstations running our own Unreal Engine digital twin and pixel-streaming studio
References are described, not named; most of these customers do not permit logos. Ask us for a call with one of them.
How to buy NVIDIA through 10ⁿ Tech
GPU nodes are quoted through NVIDIA's OEM partners and UAE distribution with the lead time on the quote, currently ten to sixteen weeks for HGX systems; we propose a cloud or PCIe interim where a date is fixed. Every quote comes with a per-rack power and cooling statement and a fabric bill that includes the transceivers. Design and integration are priced as a project with a week-one capacity and burn-in test as the first milestone.
Contact us with the part numbers or the workload and the date; we reply within a working day.
Frequently asked questions
Is 10ⁿ Tech an NVIDIA reseller?
10ⁿ Tech is an NVIDIA Inception member and buys NVIDIA GPUs inside Dell, Lenovo, HPE and Supermicro configurations through their UAE distribution, with NVIDIA AI Enterprise licences alongside. We are not an NVIDIA Partner Network reseller of bare GPUs, and rack-scale NVL72 systems go through NVIDIA's largest partners.
Can we buy NVIDIA H200 or B300 servers in the UAE?
Yes. The UAE is among the countries with the most accessible US GPU exports, and HGX systems from several OEMs are quoted through regional distribution. Lead time is the constraint: ten to sixteen weeks from order for HGX nodes, so the delivery date belongs on the quote.
How much power does a GPU server need?
An eight-GPU H200 HGX node draws up to about 10 kW; an eight-GPU B300 HGX node about 16 kW operating. A rack built for 6 kW takes neither. We size the rack, the feeds and the cooling before we size the GPU count, and we say so on the design.
What is Multi-Instance GPU and why does it matter for a small estate?
MIG partitions one NVIDIA Hopper GPU into up to seven isolated instances with their own memory and compute. It lets a sixteen-GPU estate run development, testing, staging and a sandbox on two nodes while production keeps the rest, which is the difference between a four-node and a six-node design.
Do you use NVIDIA for digital twins as well as AI?
Yes. Our Unreal Engine twins and pixel-streaming rooms run on RTX PRO workstations and servers, and Omniverse comes in where a twin has to take live data from several tools at once. The same GPU knowledge runs both practices.
