PARTLY

Partly true: the ASUS Ascent GX10 really is a 128 GB GB10 box that runs big models locally, but its 273 GB/s memory makes large dense models crawl and it takes years of real API spend to pay off

Facebook · Oct 2026

The claimUnboxing of the ASUS Ascent GX10: '128 gigs of unified memory, Nvidia's GB10 Grace Blackwell superchip, and enough AI compute to run some seriously large models locally.' On-screen: '1 PFLOP AI performance.' The goal is 'to start moving away from paying for tokens and cloud-based AI,' and 'I don't expect to hit bottlenecks anytime soon.' Ends with 'What should we test first? Comment below,' and the caption tags ASUS USA and ASUS. The creator says he has not turned it on yet.

The hardware facts are right. The GX10 is ASUS's version of NVIDIA's DGX Spark: a GB10 Grace Blackwell chip, 128 GB of LPDDR5x unified memory, ConnectX-7 for pairing two units, and NVIDIA's DGX OS. The '1 PFLOP' line is NVIDIA's sparse FP4 peak, a theoretical number you will not see in chat. The part the reel skips is the bottleneck he says he will not hit. Memory bandwidth is about 273 GB/s, and token generation on a local model is limited by bandwidth, not by the petaflop. Mixture-of-experts models run well (published tests put gpt-oss-120b at roughly 40 to 60 tokens per second), but a dense 70B model lands around 3 to 5 tokens per second and a dense 32B around 10. 'Seriously large' fits in memory; it does not always run fast. Price is moving: the 1 TB unit launched at $2,999 in October 2025, and 2026 listings range from about $3,500 to $5,999 after the DRAM shortage, so check before you buy. On the economics, a box costing several thousand dollars plus power replaces a $20 to $200 a month subscription only after years, and the open models it runs are not frontier cloud models. The unit came out of the box on camera with ASUS tagged and no price shown; we could not confirm whether it was seeded, so read it as possibly a review unit. The 'comment below' close is an engagement hook, not a sales funnel.

What holds up

Reel read frame by frame and transcript read: on-screen text '128GB unified memory, GB10 Grace Blackwell, 1 PFLOP AI performance' matches the box's key-features sleeve; the unit is never powered on in the reel, so no performance is shown.
Specs verified against ASUS's GX10 tech spec pages and reviews (StorageReview, StarryHope roundup): GB10, 128 GB LPDDR5x, 273 GB/s memory bandwidth on a 256-bit bus, 1 PFLOP is sparse FP4 peak, same chip and DGX OS as NVIDIA DGX Spark.
Throughput from published tests: gpt-oss-120b about 44 tok/s (PCMag Middle East, Ollama) to about 60 tok/s (llama.cpp), dense Qwen3 32B about 10 tok/s, dense Llama 70B about 2.7 to 4.4 tok/s depending on quantization; reviewers say dense models make the box 'feel broken'.
Price: launched at $2,999 (1 TB) in Oct 2025; 2026 listings seen at $3,499 (SHI MSRP), $3,999 (cited US spec page) and $5,999 to $7,999 (ASUS store, per a Sep 27, 2026 roundup) amid ASUS's DRAM-driven price hikes. We could not pin a single current US price.
Sponsorship: the caption tags ASUS USA and ASUS but carries no visible #ad or paid-partnership label; we could not verify whether the unit was supplied by ASUS. x-ref #303, #318 for local LLM hardware.

What doesn’t

'I don't expect to hit bottlenecks' is said before the box is switched on; the known bottleneck is 273 GB/s memory bandwidth, which caps dense large-model speed.
'1 PFLOP' is a sparse FP4 marketing peak, not chat speed.
Brand-tagged unboxing with no price and no disclosure line; treat as possibly seeded until the creator says otherwise.
'Stop paying for tokens' ignores that local open models are not the same quality as frontier cloud models, and that the hardware costs years of typical subscription spend.

The catch

The GX10 is a real, well-specced local AI box, and it shines on mixture-of-experts models like gpt-oss-120b. It is not a free pass off the cloud: dense big models are slow on its 273 GB/s memory, the price has risen sharply through 2026, and a person spending $20 to $200 a month on AI would need years to break even, while giving up frontier model quality.

How to actually do it

Add up your actual last three months of AI spend (subscriptions plus API invoices). If it is under a few hundred dollars a month, the box will not pay for itself in its useful life; stay on the cloud and cap spend.
Before buying, run the open models you would use locally (gpt-oss-120b, a Qwen MoE) through a cheap hosted API for a week and see if their quality is good enough for your work.
If they are and you have privacy or volume reasons, price the GX10 against NVIDIA DGX Spark and Mac Studio on the day you buy, and plan to run MoE models with llama.cpp or vLLM; skip dense 70B models unless 3 to 5 tokens per second is acceptable.

Every spec on screen checks out, and the box genuinely runs very large mixture-of-experts models locally at usable speeds. What the reel overstates is the freedom from bottlenecks and the payoff of dropping cloud AI, and it shows an ASUS-tagged unit with no price or disclosure.

Confidence: Medium
Posted by: a homelab and self-hosting creator page (reel ~54K views); reel sent to Buddy Mon Oct 12, 2026, ~7:14 AM

See the original claim →

We test hype for free. We build the real thing for a living.

Thirty minutes, no pitch — and you'll leave with something useful either way.

Book a call with Todd or start with the free Business Checkup →

The Verdict Weekly

Three verdicts every Friday. Free forever, unsubscribe anytime, no spam — that would be ironic.