ASUS Ascent GX10 vs NVIDIA DGX Spark in 2026: Same GB10 Chip, Same Speed — So What Does the NVIDIA Badge Cost?

By RunAIHome Team · 12 min read
asus-ascent-gx10dgx-sparkgb10nvidialocal-llmmini-pchardwarecomparison

TL;DR: The ASUS Ascent GX10 and the NVIDIA DGX Spark are the same computer: one GB10 Grace Blackwell chip, 128GB at 273 GB/s, and measured-identical inference speed. Your money buys storage tier, chassis, and who ships your OS updates — nothing else. Price the day’s listings for both and take the cheaper 4TB box.

ASUS Ascent GX10 (1TB)ASUS Ascent GX10 (4TB)NVIDIA DGX Spark (4TB)
Best forLowest entry to GB10 — if street price cooperatesSame machine as the Spark, occasionally $500+ cheaperFirst-party updates, 4TB, the reference box
Memory / bandwidth128GB @ 273 GB/s128GB @ 273 GB/s128GB @ 273 GB/s
Price (Sep 2026)$3,999 list; Amazon listings have dipped to $3,099~$4,150 on Amazon when stocked$4,699 (was $3,999 at launch)
The catch1TB fills after ~10 big models; SSD swap voids warrantyStock is erratic across retailersCosts the most for identical silicon

Honest take: Performance is a coin flip, so this is a shopping problem, not a hardware problem. If a 4TB GX10 is in stock near $4,150, buy it and pocket $550. If listings have drifted back toward the Spark’s $4,699 — they do, weekly — take the Spark for the 4TB drive and updates straight from NVIDIA instead of NVIDIA-via-ASUS.

When the GB10 boxes were announced, the pitch wrote itself: ASUS would sell you NVIDIA’s desktop AI supercomputer for $2,999 against the DGX Spark’s $3,999 — a flat $1,000 discount for the same chip. Eleven months of DRAM crisis later, both numbers are dead. NVIDIA moved the Spark to $4,699 in February, and ASUS quietly followed: the GX10’s US list price now reads $3,999 — the Spark’s old launch price — with street listings scattered anywhere from $3,099 to $4,150 depending on the week and the storage inside.

So the question the original pricing made easy (“why pay $1,000 for a badge?”) is now genuinely worth an article: what exactly separates two machines built on the same silicon, and which one should get your money in September 2026?

Same chip, and this time “same performance” is measured, not assumed

Both boxes are built around NVIDIA’s GB10 Grace Blackwell Superchip: a 20-core Arm CPU and a Blackwell GPU on one package, rated up to 1 petaFLOP of sparse FP4 compute, wired to 128GB of LPDDR5X unified memory at 273 GB/s. Same memory, same bus, same GPU, same DGX OS software stack. ASUS doesn’t get to bin the chip differently, and there’s no factory overclock — this isn’t a partner GPU card situation.

StorageReview’s GX10 review ran vLLM serving benchmarks across GPT-OSS, Qwen3-Coder, and Llama 3.1 and found the ASUS box “performed consistently with other Spark systems” — identical within noise. That means every DGX Spark number we’ve published applies to the GX10 unchanged, including the ones from llama.cpp’s official Spark thread:

$ llama-bench -m gpt-oss-120b-MXFP4.gguf -fa 1
| model               | test    |     t/s |
| ------------------- | ------- | ------: |
| gpt-oss 120B MXFP4  | pp2048  | 1723.07 |
| gpt-oss 120B MXFP4  | tg32    |   38.55 |

Call it 38–60 tok/s decode on the flagship 128GB-class MoE (the 60 is LMSYS’s tuned SGLang build), ~1,700–1,950 tok/s prefill, and the same ~2.6 tok/s slog on dense 70B models that every 273 GB/s box is condemned to — we walked through why in the Spark vs Mac Studio M5 Max comparison. If those numbers don’t fit your workload, no amount of choosing between these two boxes fixes it; skip to the last section.

With speed off the table, four real differences remain.

Difference 1: storage — and the non-negotiable SSD

The DGX Spark ships one way: 4TB, $4,699. The GX10 ships in 1TB, 2TB, and 4TB trims — the 2TB oddly steps down to a PCIe Gen4 drive while 1TB and 4TB are Gen5, per ASUS’s own listings.

Storage tier matters more than it sounds like it should, because on both machines the SSD is effectively sealed: ASUS’s support FAQ classifies damage from a customer SSD swap as Customer-Induced Damage outside warranty, and NVIDIA takes the same line on Spark hardware modifications. So the capacity you order is the capacity you own.

Now do the math on 1TB. DGX OS plus the CUDA/NIM stack claims its share out of the box; gpt-oss 120B at MXFP4 is ~60GB; a 70B Q4 GGUF is ~42GB; add a couple of video-generation checkpoints, a fine-tuning dataset, and the Docker cache that NVIDIA’s playbooks assume, and a 1TB drive is doing inventory management inside six months. It’s workable with a NAS for cold weights, but “workable” is doing real lifting. The 4TB trims make the whole question disappear for years — which is exactly why NVIDIA only sells that one.

Difference 2: the update path — NVIDIA direct vs NVIDIA-via-ASUS

Both machines run DGX OS and only DGX OS — ASUS confirms the GX10 supports nothing else. The difference is the release pipeline: Spark updates come straight from NVIDIA, while GX10 builds are “adapted and validated by ASUS” before they reach the DGX Dashboard updater. DGX OS 7.5.0 (April 2026) reached the GX10, but GX10 owners have had to ask on NVIDIA’s forums where their firmware updates live — the answer being ASUS’s support site, on ASUS’s schedule.

For a box whose entire value is the NVIDIA software stack, one extra company in the update chain is a small but honest tax on the discount. It has already mattered once this year, which brings us to the fix section.

The 13 Gbps problem (and the fix that made it 196)

Both boxes carry two QSFP ports off an integrated ConnectX-7 NIC — 200 Gbps links whose whole purpose is pairing two GB10 units into one 256GB pool for models neither runs alone. ASUS sells the GX10 with a stacking cable as an official two-box bundle, and the wiring underneath is genuinely odd: the GB10 attaches the ConnectX-7 over two separate PCIe Gen5 x4 links in multi-host mode, so one physical port surfaces as two interfaces in the OS.

Here’s the problem that bit early adopters on both machines. You cable two units together, run iperf3, and get ~13 Gbps on a link that reports 200:

$ iperf3 -c 192.168.100.2
[  5]   0.00-10.00  sec  15.1 GBytes  13.0 Gbits/sec

Nothing is miscabled. The root cause, per the NVIDIA developer forum thread that ran it down, was ConnectX-7 firmware power-throttling itself off a bogus 27W power report on early builds (kernel 6.17.0-1008-nvidia, driver 580.126.09). The fix is anticlimactic:

$ sudo apt full-upgrade && sudo reboot

Post-update (kernel 6.17.0-1014-nvidia, driver 580.142), the same link runs ~196 Gbps, and independent two-box testing over RDMA/NCCL/vLLM confirms 185–190 Gbps sustained with RoCE and jumbo frames (MTU 9000). The takeaway for this comparison: the fix shipped through the DGX OS channel — Spark owners pulled it the day it landed, GX10 owners waited on the ASUS-validated build. Days, not months, but it’s the update-path difference made concrete.

Difference 3 and 4: the box itself, and the price of admission

The hardware around the chip is where ASUS actually did work. Per IT Pro’s review, the GX10 is a 150 × 150 × 51mm, 1.48kg metal box — lighter than the Spark, which ServeTheHome found packs Mac-Studio-like density — and it puts the power button on the front, which sounds trivial until you’ve reached around the back of a stacked pair for the tenth time. The Spark counters with the collectible factor: NVIDIA’s own champagne-gold chassis, which is either worth $0 or several hundred dollars depending entirely on how you feel about NVIDIA.

Then there’s what you’ll actually pay, which in September 2026 is less a price than a weather report:

  • DGX Spark: $4,699, 4TB, raised from $3,999 on February 23 with memory supply named as the reason. Stable and mostly in stock.
  • Ascent GX10: US list $3,999 as of mid-August (Walmart matched it at $3,999.99 in late August), but Amazon has carried the 1TB trim as low as $3,099.99 and the 4TB at $4,149.99 — with listings cycling in and out of stock week to week. UK pricing sits around £3,640.
  • The rest of the GB10 family — HP ZGX Nano, Lenovo ThinkStation PGX, Dell’s Pro Max, which approaches £6,000 — bracket the same silicon at other prices, which mostly proves the point: you are shopping retailers, not engineering.

The DRAM crisis is the villain of every price line on this site this year, and it has inverted the original story. At launch the GX10 undersold the Spark by $1,000 for a 3TB storage sacrifice. Today a 4TB GX10 at $4,149 versus the 4TB Spark at $4,699 is a clean $550 discount for an identical config — when you can catch it in stock.

Should you buy either one?

A reality check before the decision table, because both machines fail the same buyers. At 273 GB/s, decode speed on big dense models is single-digit — these are the wrong boxes for chatting with a 70B, where a Mac Studio M5 Max runs 2× the decode speed and a $2,959 Strix Halo mini PC matches the Spark’s tok/s for $1,700 less. And if everything you run fits in 24GB, a used RTX 3090’s 936 GB/s embarrasses this entire category. The GB10 case is specific: you want CUDA — fine-tuning (ASUS demonstrated LoRA on Llama 3 8B at 53K peak tokens/sec), prefill-heavy agentic work, NIM/TensorRT pipelines, day-one model support — in one small 240W box, with a petaFLOP of FP4 behind it. Run your model sizes through the VRAM calculator before spending $4,000 to find out you needed bandwidth, not capacity.

What to actually buy

Prices as of September 2026, all verified above; GB10 street prices move weekly, so check both boxes the day you order:

Your situationThe machinePriceWhere
You want the GB10 platform at the lowest real price and 4TB is in stockASUS Ascent GX10 (4TB)~$4,150Check price
GX10 stock is dry, or you want updates and support from NVIDIA with no middlemanNVIDIA DGX Spark (128GB, 4TB)$4,699Check price
You’ll cluster two boxes and keep weights on a NAS — capacity per dollar wins2× ASUS Ascent GX10 (1TB) + stacking cable2× $3,099–3,999Check price
You mostly chat with big MoE models — bandwidth matters more than CUDAMac Studio M5 Max 128GB or MS-S1 Max$2,959–5,099Our comparison
Not sure GB10 fits your workload — test on rented Blackwell firstRented GPU, ~$1/hrpay per hourRunPod

FAQ

Is the ASUS Ascent GX10 slower than the DGX Spark? No. Same GB10 chip, same 128GB at 273 GB/s, and StorageReview’s vLLM testing measured the GX10 as consistent with other Spark-platform systems across GPT-OSS, Qwen3-Coder, and Llama 3.1. Expect 38–60 tok/s on gpt-oss 120B and ~1,700 tok/s prefill on either box.

Why is the DGX Spark more expensive if the hardware is the same? You’re paying for the 4TB drive (GX10 starts at 1TB), NVIDIA’s own chassis, and OS/firmware updates that come straight from NVIDIA rather than through ASUS validation. Whether that’s worth $550–$1,600 depends on the day’s GX10 street price.

Can I upgrade the GX10’s 1TB SSD later? Officially, no — ASUS classifies damage from a customer SSD replacement as Customer-Induced Damage outside warranty, and NVIDIA applies the same policy to the Spark. Buy the storage you’ll need in year three, or plan on network storage for cold model weights.

Can I mix a GX10 and a DGX Spark in a two-box cluster? They’re the same GB10 platform with the same ConnectX-7 QSFP ports, and the software stack is common — but the official pairing guidance and ASUS’s own bundles assume matched units. If you’re buying two anyway, buy two of whichever is cheaper that week and update both to current firmware first, or your 200 Gbps link may run at 13 (see the fix above).

Does the $3,999 GX10 1TB ever make sense over the ~$4,150 4TB? At list-vs-list, no — $150 for 3TB of non-upgradeable Gen5 storage is the easiest money you’ll ever spend. The 1TB trim only makes sense when it’s actually selling near its $3,099 Amazon low, or as the per-node price in a two-box cluster backed by a NAS.

Sources

Last updated September 14, 2026. Prices and specs change; verify current rates before purchasing.

Was this article helpful?