Hardware · AI
NVIDIA DGX Spark and ASUS GX10 in Bulgaria: what is true
NVIDIA's miniature AI computer attracts companies that cannot or do not want to send data to the cloud. But the first version of this article contained figures and claims that we cannot back up with a source. Here is what can be checked, and what remains unconfirmed, including whether it can be bought in Bulgaria and at what price.
- The original (12.05.2026) told of "our first 4 deployments in April" and said the machine "already has local customers in Bulgaria".
"already has local customers in Bulgaria"→ removed: we cannot back it up with a source, and availability in Bulgaria could not be confirmed. [5][6] - The original gave speeds: Llama 3.3 70B 12–18 tokens/s, Qwen 2.5 72B ~14, BGE-M3 ~800 texts/s.
"12–18 tok/s"→ removed: for a dense 43 GB model and a memory bandwidth of 273 GB/s the upper bound is about 6 tokens/s; there are no measurements. [1][7] - The original claimed a "real 3-year TCO of 18–22K BGN".
"18–22K BGN"→ no price is given: there is no primary source or Bulgarian shop with a price; since 01.01.2026 we have been in euro. What remains is a list of costs without amounts. [4][5][6] - Corrected: "GX10 Spark" mixed up two products, the NVIDIA DGX Spark and the ASUS Ascent GX10. Added: specifications from NVIDIA and ASUS, a section on availability in Bulgaria, and sources. [1][2]
01 · THE NAMESTwo names, one platform
The NVIDIA DGX Spark is NVIDIA's own system. The ASUS Ascent GX10 is ASUS's system on the same platform, the NVIDIA GB10 Grace Blackwell chip. NVIDIA lists ASUS among its partners that offer GB10 systems. [1][2] That is why "GX10 Spark" from the first version is not a product name: we are talking about two products with similar capabilities.
02 · SPECIFICATIONSWhat the manufacturers say
The key point is the unified memory: the processor and the graphics chip share the same 128 GB, rather than the graphics chip having a small video memory of its own. [1][2]
| Parameter | NVIDIA DGX Spark [1] | ASUS Ascent GX10 [2] |
|---|---|---|
| Processor | 20-core Arm (10 Cortex-X925 + 10 Cortex-A725) | the same |
| Memory | 128 GB LPDDR5x, unified; 64 GB only through partners, "coming soon" | 128 GB LPDDR5x, unified |
| Memory bandwidth | 273 GB/s | not stated |
| Compute | up to 1 PFLOP at FP4 | 1 PFLOP |
| Networking | 10 GbE; ConnectX-7 (200 Gbps) | 10G Ethernet; ConnectX-7 |
| Power | 240 W | 240 W adapter |
| Storage | up to 4 TB NVMe | variants with 1, 2 or 4 TB |
| Size | 150 × 150 × 50.5 mm | 150 × 150 × 51 mm |
| Weight | 1.2 kg | 1.48 kg |
What NVIDIA promises: inference of models up to 200 billion parameters and fine-tuning up to 70 billion with 128 GB; up to four systems can be connected to each other. [1] Note that the weight differs between the two manufacturers' data.
03 · AVAILABILITYCan you buy it in Bulgaria?
On 13 October 2025 NVIDIA announced that the DGX Spark ships through NVIDIA and its partners. [3] On its page NVIDIA lists Amazon, Arrow, Micro Center, PNY, TD Synnex and Ingram Micro as retail partners; this is the US page, with no list for Bulgaria. [1][4] ASUS states in its specifications that products "may not be available in all markets". [2]
What we checked on 3 October 2026:
- The product address on ASUS Bulgaria redirects to the ASUS Bulgaria home page; we found no official Bulgarian listing. [5]
- A search for "ascent gx10" on desktop.bg returned 0 results. [6]
04 · THE COSTSWhat it really costs
We give no price and no total: as of 3 October 2026 we found no primary source or Bulgarian shop with a price. That is also why we removed the "real 3-year TCO" from the first version. Instead of a number, here is what the cost consists of when you work it out yourself:
- The machine itself, per the supplier's quote, in euro including VAT.
- Protection against power outages and against overheating where it stands.
- Secure remote access (a VPN or similar) and network setup.
- Installation, setting up the models and training the people who use them.
- Backup of the data and of the vector database, if there is one.
- Maintenance work over time.
If the cloud costs you less at your volume, that is also a valid answer. We discuss it with clients before the purchase, not after.
05 · THE SPEEDHonest about speed
The first version gave specific speeds for 70B models. We cannot back them up with a measurement, and the upper bound contradicts them: for a dense model every new word reads the whole set of weights from memory. At 273 GB/s [1] and a 43 GB file, that is about 6 tokens per second at most. The detailed calculation and the model sizes are in a separate article: Ollama on the GX10: 70B models locally, what is true.
Small models (for document search and text recognition) are much lighter and can run much faster, but we will not publish a specific number without our own measurement.
06 · THE MISTAKESCommon mistakes with local AI
This is practical advice, not a report on clients. Three mistakes worth avoiding:
Local is not a replacement for the cloud
For some tasks cloud models remain stronger. A local machine is a complement, mostly when the data must not leave the organisation.
No backup
If you have a vector database with indexed documents, back that up too, not only the originals.
Access that is too wide
Remote access to the machine should be limited to a small, known circle of people and be protected.
This is general information, not a recommendation to buy.
07 · SOURCESSources
- NVIDIA, "NVIDIA DGX Spark" (specifications, partners, 64 GB "coming soon") — nvidia.com/…/dgx-spark
- ASUS, "ASUS Ascent GX10 — Tech Specs" — asus.com/…/asus-ascent-gx10/techspec
- NVIDIA, "NVIDIA DGX Spark Arrives for World’s AI Developers", 13 October 2025 — nvidianews.nvidia.com
- NVIDIA Marketplace, personal AI supercomputers (purchase options) — marketplace.nvidia.com
- ASUS Bulgaria (the product address redirects to the home page, 3 October 2026) — asus.com/bg
- Desktop.bg, search "ascent gx10": 0 results, 3 October 2026 — desktop.bg
- KAGAMI, "Ollama on the GX10: 70B models locally" — 2026-05-ollama-gx10.html
Checked on 3 October 2026. Sources 1–2 are primary; availability and prices change, so see the primary sources and ask a supplier.
08 · RELATEDContinue from here
What fits on GX10: the Memory Math
How to calculate how much memory a model needs.
Analysis · BlogOllama on the GX10: 70B models locally
Sizes, memory bandwidth and a simple calculation of speed.
Step · The LadderSession B1 · 45 min · €99
We discuss whether a local AI makes sense for your company and what machine you need.