Summary: A GB300 desktop and a small DGX Spark cluster do different jobs. The GB300 has 252 GB of memory at 7.1 TB/s and serves large models to a whole team, split into up to seven slices. A Spark has 128 GB at 273 GB/s: enough to fine-tune a 70-billion-parameter model, too slow to serve one to many people. Put serving on the GB300 and development, fine-tuning and tests on the Sparks. The graph, rules and search software around the models runs on CPUs and needs its own tier. Prices and export rules for the UAE are at the end, checked on 16 October 2026.
This month we designed desk-side AI setups in the UAE: single DGX Spark units, small Spark clusters, and GB300 workstations next to them. The question buyers ask first is price. The question that decides whether the money is well spent is which box does which job.
The number that matters is memory bandwidth
Both machines are sold on memory size. Speed comes from memory bandwidth, because a language model reads its weights from memory for every token it writes.
| DGX Spark (GB10) | GB300 desktop | |
|---|---|---|
| Fast memory | 128 GB LPDDR5x, 273 GB/s | 252 GB HBM3e, 7.1 TB/s |
| Second tier | none | 496 GB LPDDR5X, 396 GB/s |
| Total coherent memory | 128 GB | 748 GB |
| AI compute (FP4) | up to 1 petaFLOP | up to 20 petaFLOPS |
| Network | ConnectX-7, 200 Gb/s | ConnectX-8, up to 800 Gb/s |
| Partitioning | none | MIG, up to 7 instances |
Figures from NVIDIA’s product pages and DGX Spark hardware documentation, accessed 18 October 2026.
The fast memory on a GB300 is 26 times quicker than a Spark’s. A published test shows what that means in practice. StorageReview ran the open-weight gpt-oss-120b model under vLLM on an MSI XpertStation WS300, a GB300 desktop, and on a DGX Spark, with up to 128 requests at once. Peak output was 9,294 tokens a second on the GB300 and 500 on the Spark (StorageReview, 24 August 2026).
What a GB300 desktop can run
NVIDIA says the GB300 desktop supports models of up to 1 trillion parameters, without stating the precision. The arithmetic helps: at 4-bit precision each parameter takes half a byte, so a trillion-parameter model needs about 500 GB for its weights. That does not fit in the 252 GB of fast memory. It runs by spilling into the 496 GB second tier, which is 18 times slower.
So plan in two bands:
- Models whose 4-bit weights fit in 252 GB with room for the cache each user adds. These run at full speed. In StorageReview’s tests a 125 GB model sat entirely in fast memory.
- Larger models that spill into the second tier. They run, but slowly. GLM-5.2 at 4-bit needs about 434 GB, so part of it sits in the slower tier: StorageReview measured 35 tokens a second for one user and 139 across 32 users on a single GB300 desktop.
The GB300 is built to be shared. NVIDIA describes it as a desktop for one user or “a centralized compute node for multiple team members”, and MIG splits the GPU into as many as seven isolated instances. StorageReview ran three MIG slices, each serving its own vLLM engine under NVIDIA Dynamo on a small Kubernetes cluster. That is the pattern we use: one slice for the main model, others for embeddings, reranking and guard models, each with its own endpoint.
Two GB300 desktops can be linked over two 400 Gb/s cables. In StorageReview’s test, a second unit took a large model from 188 to 853 tokens a second at 32 users.
The GB300 ships with Ubuntu and NVIDIA’s AI developer tools. NVIDIA AI Enterprise, with enterprise support, is a separate licence: budget for it.
What is left for a DGX Spark cluster
Plenty, but not serving. NVIDIA’s own figures for one Spark are fine-tuning of models up to 70 billion parameters and inference up to 200 billion. Fine-tuning is where a Spark earns its price: it frees the GB300 to keep serving while a developer adapts a model on their own desk.
Inference on a Spark is fine for one person testing. LMSYS measured Llama 3.1 70B at 2.7 tokens a second on a Spark, slower than most people read (LMSYS, 13 October 2025). Serve that model to a department from a Spark and the queue becomes the product.
Clustering raises capacity, not speed. NVIDIA’s page says ConnectX can link up to four Sparks, giving 512 GB for models of up to 700 billion parameters, and its playbook covers larger switch-based clusters. Each Spark’s network port sits behind PCIe links that limit useful throughput to about 200 Gb/s (StorageReview, 5 October 2026). A four-Spark cluster is a good place to test multi-node code and very large models before they go anywhere near production. It is not a cheap GB300.
How we split the work
| Job | Where it runs | Why |
|---|---|---|
| The main model for chat, agents and document answers | GB300, one MIG slice | 7.1 TB/s memory, many users at once |
| Embedding, reranking and guard models | GB300, smaller MIG slices | small models, own endpoints, isolated |
| Fine-tuning up to 70B | Spark | NVIDIA’s stated limit; keeps the GB300 serving |
| Prototypes, evaluations, agent development | One Spark per developer or team | a private box, no queue |
| Multi-node code and very large model tests | Spark cluster | 512 GB across four units, NCCL over ConnectX-7 |
| Knowledge graph, audit-trail database, Bayesian models | CPU servers or virtual machines | no GPU needed; licensed per core or per seat |
Two more decisions belong in the design before anyone buys. Where the documents live, because retrieval needs storage and a vector index next to the GB300. And who may reach which slice, because MIG isolates the hardware and the access rules still need writing. We set out the rack-scale version of the same split in what a sovereign AI platform needs on-premises.
The software around the GPUs runs on CPUs
A desk-side AI platform is more than models. The agents on it usually need four other kinds of software, and none of them uses the GPU:
- A knowledge graph with a reasoner, holding the organisation’s definitions and rules, so an agent looks a rule up instead of guessing it.
- A graph database that records which source, tool and step produced each answer, so a reviewer can trace a decision afterwards.
- Bayesian network models for risk questions, which give a probability and show the factors behind it.
- A web search API that brings in current outside information.
All four run on CPU servers and are licensed per CPU core, per seat or per credit. Budget a CPU tier next to the GB300 and the Sparks: a few servers or virtual machines, with a cluster for the knowledge graph if it must stay up.
Do not install them on the GB300 itself. Its Grace processor has 72 cores, and a per-core licence may count every core on the host unless the publisher agrees otherwise. They would also compete for memory with the model being served. Ask each publisher in writing how cores are counted on a large host and on each node of a cluster: one knowledge-graph vendor sells its enterprise edition from twelve cores up, which matters when a cluster has three nodes.
Read the licence before you embed. One widely used Bayesian toolkit needs a separate deployment licence when its library runs inside a server program, on top of the modelling seats. The same toolkit compiles its key into the program, with no online activation, which suits an air-gapped network. Ask every other publisher how a licence activates with no internet access.
Search is the one deliberate hole in the air gap. The search APIs built for AI agents are cloud services. Route them through one outbound gateway that allows only the provider’s addresses, so no agent reaches the open web directly. Give production, staging and development their own keys, set quotas with an alert below the paid allowance, and copy the usage log into your own logging.
DGX Spark price in the UAE
The 128 GB DGX Spark launched at USD 3,999, rose to USD 4,699 in February and to USD 6,950 on 2 October, in the middle of a memory shortage. A 64 GB version at USD 4,999 is sold from 23 October through OEM partners only. Check which one a quote is for.
UAE retail on 16 October, including VAT:
| Listing | Price (AED) | Stock |
|---|---|---|
| NVIDIA Founders Edition, Amazon.ae | 26,999.90 | “Only 2 left” |
| NVIDIA Founders Edition, Microless | 28,888 | listed in stock, page also says unavailable |
| Founders Edition two-pack, UAEGamers | 59,400 (29,700 each) | not stated |
| MSI EdgeXpert, Amazon.ae and GCC Gamers | 29,199 to 31,500 | 2 left / in stock |
| Dell Pro Max GB10, Gear-up | 35,849 | 3-5 units |
No single listing could supply five units, and no UAE distributor had published an allocation. For a cluster, source the 200 Gb/s switch and cables at the same time: the compact switch most often used with Sparks was out of stock in the UAE, and the NVIDIA switches listed locally were the 100 Gb/s variant.
GB300 desktop prices
Distributors often do not publish GB300 prices. Most quote on request, after the end-user checks that NVIDIA and the US export rules require. The figures below are US list prices and one European reseller’s table, not UAE quotes. Use them for a budget range, then ask for a written quote.
| Vendor | Model | Price (USD) | Source and stated lead time |
|---|---|---|---|
| Dell | Pro Precision (formerly Pro Max) with GB300 | 159,283 | Dell US list; about 8 weeks on Dell’s page |
| HP | ZGX Fury G1n | about 148,500 | European reseller, 4-5 weeks |
| ASUS | ExpertCenter Pro ET900N G3 | about 123,500 | European reseller, 6-12 weeks |
| Gigabyte | W775-V10-L01 | about 112,000 | European reseller, 8-10 weeks |
| MSI | XpertStation WS300 | about 104,500 | European reseller, 8-12 weeks |
| Supermicro | Super AI Station ARS-511GD | 92,887 | Supermicro US eStore list, 12-16 weeks |
Reseller figures are from a table dated 13 October. In the UAE we found one retail listing, an ASUS unit at AED 525,000, on a page that also said “no longer available”. Add one to two weeks to any lead time for export paperwork, freight and customs.
Export rules in one paragraph
On 10 July 2026 the US moved the UAE into a more favourable export group (A:5). UAE government agencies can now receive advanced computing items licence-free. Other buyers still need a US export licence unless they are on the approved list in Supplement No. 8 to Part 740, and end-use rules apply to everyone. The OEM’s trade-compliance team makes the call. A private company should get the licence position in writing before it signs.
Frequently asked questions
What can a GB300 desktop run?
Models whose 4-bit weights fit in 252 GB of fast memory run at full speed, shared across up to seven MIG instances. NVIDIA says it supports models of up to 1 trillion parameters, which at 4-bit means spilling into the slower 496 GB tier.
Is a cluster of DGX Sparks a substitute for a GB300?
No. Four Sparks give 512 GB of capacity but each runs at 273 GB/s against 7.1 TB/s on the GB300. In a published test the GB300 produced 9,294 tokens a second on gpt-oss-120b against 500 on a Spark. Use Sparks for development, fine-tuning and tests.
Does a knowledge graph or graph database need the GPU?
No. Graph databases, knowledge graphs with reasoners and Bayesian models run on CPUs and are licensed per core or per seat. Put them on separate CPU servers, not on the GB300’s own processor.
How much does a DGX Spark cost in the UAE?
On 16 October 2026 the 128 GB Founders Edition was listed at about AED 27,000 to 29,000 including VAT, and OEM GB10 systems at AED 29,000 to 36,000. NVIDIA’s list price is USD 6,950.
Why are GB300 desktop prices hard to find?
Most distributors quote GB300 desktops on request, after NVIDIA’s end-user check and the export review. Published figures are US list prices and European reseller tables, so a UAE buyer needs a written quote.
Next step
If you are planning desk-side or on-premises AI, we can design the split, size it, source it and set it up. See our on-premises AI work and our NVIDIA partnership, or get in touch.
Cover photo by Nikolay on Unsplash.
What we build: on-premise AI in the UAE.