NVIDIA DGX Spark in Dubai: What It Does and Who Needs One

NVIDIA DGX Spark is a 1.2 kg desktop machine built to hold a large AI model in memory rather than to run one at maximum speed. It carries 128 GB of coherent unified memory, which is more than any single workstation graphics card offers, at 273 GB/s, which is far less bandwidth than one. For a Dubai team deciding where to put its AI work, that single trade is the whole decision.

NVIDIA announced general availability on 13 October 2025, with orders opening on 15 October 2025 through NVIDIA directly, through channel partners worldwide, and through partner-built systems from Acer, ASUS, Dell, GIGABYTE, HP, Lenovo and MSI. A year on, the question in most UAE offices is no longer whether the machine exists. It is whether it answers a problem the business actually has.

What is actually in the box

The published specification from NVIDIA's DGX Spark hardware documentation is specific enough to plan around. The GB10 Grace Blackwell Superchip pairs a 20-core Arm CPU, built from 10 Cortex-X925 and 10 Cortex-A725 cores, with a Blackwell-architecture GPU carrying 6,144 CUDA cores and fifth-generation Tensor Cores. Memory is 128 GB of LPDDR5X on a 256-bit interface running at 4266 MHz, giving 273 GB/s. Storage is 1 TB or 4 TB of self-encrypting NVMe M.2.

Networking matters more than people expect on a machine this size. Alongside a single 10 GbE RJ-45 port, Wi-Fi 7 and Bluetooth 5.4, the rear panel carries two QSFP ports driven by a ConnectX-7 NIC at 200 Gbps. Those QSFP ports exist for one job: linking a second unit. NVIDIA rates a single DGX Spark for models up to 200 billion parameters and a two-unit pair for up to 405 billion.

Power is the detail that changes where the machine can live. The external supply is rated 240 W and the GB10 package itself carries a 140 W thermal design power. The chassis measures 150 mm by 150 mm by 50.5 mm. It runs DGX OS. A box that size and that power draw sits on a desk in a Business Bay or JLT office on an ordinary socket, which is not true of a four-GPU server.

The number that decides whether it suits you

Compare 273 GB/s against the alternative most UAE design and engineering teams already understand. NVIDIA publishes 1,792 GB/s of memory bandwidth for the RTX PRO 6000 Blackwell Workstation Edition, on a 512-bit interface with 96 GB of GDDR7. That card has roughly 6.5 times the bandwidth of DGX Spark and about three quarters of its memory capacity.

For token generation, bandwidth is usually the binding constraint, so a workstation card will produce output faster on any model that fits inside its 96 GB. DGX Spark wins when the model does not fit at all. Run the arithmetic: 120 billion parameters at four bits each is roughly 60 GB of weights, and context, activations and overhead push the working figure higher still, which is uncomfortable on 96 GB and comfortable on 128 GB. That is the real line between the two products.

Teams that buy DGX Spark expecting a fast inference server are the ones who end up disappointed. Teams that buy it to develop, fine-tune and test against a model they will later deploy on larger hardware tend to get what they paid for. If your current bottleneck is a mobile workstation that runs out of VRAM, the capacity argument applies. If it is throughput for live users, it does not.

A 64 GB version arrives on 23 October 2026

On 2 October 2026, NVIDIA announced a 64 GB version of DGX Spark, launching on 23 October 2026. It keeps the same GB10 Superchip, the same DGX OS and the same software stack, and NVIDIA rates it for models up to 100 billion parameters on one unit or 200 billion across a clustered pair. In NVIDIA's own published testing, two clustered 64 GB units delivered up to 1.7 times the performance of a single unit, because clustering pools the memory to 128 GB and doubles the available bandwidth.

That announcement is four days old at the time of writing, and it changes the buying maths for anyone who was about to commit. A UAE business planning a purchase this month now has two entry points in the line rather than one, and the 64 GB unit is the smaller of them. What any of it costs landed in the UAE depends on duty, VAT treatment and the channel you buy through, so a figure quoted in another market is not a number to plan against.

Why a Dubai business would run this on a desk

The usual answer is cost, and it is usually the wrong one. The defensible reason in this market is where the data is allowed to go. A firm licensed in the Dubai International Financial Centre sits under DIFC Data Protection Law No. 5 of 2020, which commenced on 1 July 2020 and restricts transferring personal data to a third country unless an adequate level of protection applies or the conditions in Article 27 are met. A team that cannot route client records through a foreign cloud region has a genuine reason to keep inference inside the office.

The second reason is practical. Dubai summer temperatures put real load on comms rooms, and adding a dense GPU server to a cupboard that is already struggling is how outages start. A 240 W desktop unit avoids that problem entirely, provided it is not tucked into an unventilated cabinet. The same housekeeping rules apply as anywhere else in the room, and the ones that prevent outages are mostly about airflow and labelling.

Where it does not belong

• Production inference for many concurrent users. 273 GB/s and a 140 W package are development-scale numbers. Serving a customer-facing workload belongs on server hardware with the cooling and redundancy to match.
• General virtualisation. DGX Spark runs DGX OS on Arm. It is not a replacement for an x86 host running a mixed estate, and sizing that estate is a different exercise with different numbers.
• Anywhere software is x86-only. The 20-core CPU is Arm. Check every dependency in your stack before the purchase order, not after delivery.
• Clusters larger than two. NVIDIA rates the paired configuration at 405 billion parameters. Beyond that the answer is different hardware, not more Sparks.

What to settle before you raise a purchase order

Four questions decide whether this purchase ages well. First, name the model you intend to run and work out its memory footprint at your chosen quantisation, then compare that against 64 GB and 128 GB rather than against a marketing figure. Second, decide now whether you will ever need a second unit, because the QSFP ports and the 200 Gbps ConnectX-7 NIC only pay off if you plan the pair. Third, choose between 1 TB and 4 TB of NVMe on the basis of how many checkpoints and datasets you keep locally, since the storage is not something you will want to revisit.

Fourth, buy it the way you would buy any other piece of enterprise hardware. Warranty entitlement, support eligibility and software licensing all follow the channel the unit came through, which is why source matters more than the headline price. Read the quotation line by line before you sign it, because the lines that cause arguments later are usually the ones nobody read; the same line-by-line discipline applies here as to a server or a switch.

Codeeo Tech carries NVIDIA hardware alongside Cisco, HPE, Dell EMC and Supermicro from Building A1, IFZA Business Park in Dubai Silicon Oasis. You can browse the NVIDIA range or send the model name and memory figure you arrived at above through the contact form, and the quote will come back against that requirement rather than against a guess.

Cover photo: Dubai skyline unsplash, via Wikimedia Commons (CC0).