ASUS Ascent GX10 vs DGX Spark: GB10 Boxes Compared (2026)

Every certified GB10 box on the market, NVIDIA's own DGX Spark included, runs the same silicon at the same power budget. The ASUS Ascent GX10, Dell Pro Max with GB10, HP ZGX Nano, Lenovo ThinkStation PGX, Gigabyte AI TOP ATOM, MSI's GB10 system, and Acer's Veriton GN100 all ship the identical GB10 Grace Blackwell Superchip at 140W, inside a 240W system power budget, as ServeTheHome's GX10 review measured. So "ASUS Ascent GX10 vs DGX Spark" is not really a GPU comparison. It's a storage-generation, warranty, and price comparison wearing a GPU comparison's clothes, and nobody selling you one of these boxes wants you to notice that.

TL;DR: ASUS Ascent GX10 vs DGX Spark, and the Other GB10 Boxes

• Shared chip: every certified GB10 box runs the identical 140W SoC inside a 240W power budget, per ServeTheHome's GX10 review.
• Price moved fast: DGX Spark's 128GB/4TB tier launched at $3,999 and now lists at $6,950 as of Oct. 3, 2026; a new 64GB tier starts at $4,999 on Oct. 23, 2026.
• Storage differs most: ASUS's 1TB drive's write throughput peaks at just 0.72 GiB/s at the 16K block size before falling to 0.62 GiB/s at higher thread counts; Lenovo and Gigabyte offer the same PCIe Gen4 drive class at up to 4TB.
• Throughput is a software choice: GPT-OSS-20B runs around 70 tokens/sec on SGLang versus up to 79.74 tokens/sec on llama.cpp, same GB10 hardware.
• Spheron rents H100 and H200 GPUs by the minute past any GB10 box's 128GB ceiling. H200 GPU rental →

All GB10 Systems Share the Same Chip: What Actually Differs

Before comparing boxes, it helps to be precise about what is and isn't variable. The GB10 Grace Blackwell Superchip pairs a 20-core Arm CPU (10 Cortex-X925 performance cores plus 10 Cortex-A725 efficiency cores) with a Blackwell GPU using 5th-generation Tensor Cores, delivering up to 1 PFLOP of FP4 compute with sparsity and 273 GB/s of unified memory bandwidth, according to NVIDIA's own DGX Spark product page. That page also confirms the chip's 140W TDP, which ServeTheHome's GX10 review independently measured inside a 240W total system power draw on the ASUS box.

Every OEM partner, Acer, ASUS, Dell, Gigabyte, HP, Lenovo, and MSI, builds around that exact same chip, the exact same 128GB LPDDR5X-8533 soldered memory, and the exact same 273 GB/s bandwidth ceiling. NVIDIA's own product page lists all seven as OEM partners alongside its Founders Edition DGX Spark, which puts the certified roster at eight systems total.

That means tokens-per-second on a given model, the number everyone actually cares about, is a property of the chip and the inference stack you run on it, not a property of which vendor's chassis it sits in. A 70B model quantized to fit in 128GB runs at the same speed on a Lenovo ThinkStation PGX as it does on the ASUS box next to it, assuming identical software. What changes box to box is what you're paying for the privilege, how fast the thing boots and loads models off disk, and who answers the phone when it breaks.

LMSYS's GPT-OSS-20B benchmarking on DGX Spark is the clearest proof of that: SGLang serves GPT-OSS-20B at around 70 tokens/sec on the chip, a figure that depends entirely on which serving stack was running, not on which box it ran in. A separate llama.cpp community benchmark thread measured llama.cpp at 79.74 tokens/sec on the same model at zero context depth, on the same GB10 hardware. That's a real gap between two software stacks running on identical silicon. Any OEM box in this lineup will post the same range once you control for the stack running on it.

The Full Lineup: 8 Certified GB10 Boxes

Here's the complete roster as confirmed by NVIDIA's own partner list and independent reviews:

1. NVIDIA DGX Spark (Founders Edition) - the reference design, sold direct and through Amazon, Micro Center, PNY, and TD Synnex.
2. ASUS Ascent GX10 - the OEM box with the most independent review coverage, including a mid-cycle chassis redesign.
3. Dell Pro Max with GB10 - positioned inside Dell's Pro Max workstation line.
4. HP ZGX Nano AI Station (G1n) - runs NVIDIA's own DGX OS, based on Ubuntu 24.04, the same Linux-only base every GB10 box ships with.
5. Lenovo ThinkStation PGX - Lenovo's corporate-channel take on the same chip, reviewed with a focus on business procurement relationships.
6. Gigabyte AI TOP ATOM - ships in 1TB and 4TB configurations, per ServeTheHome's AI TOP ATOM review.
7. MSI's GB10 system - MSI's desktop AI computer built on the same chip, listed as an OEM partner on NVIDIA's own DGX Spark product page.
8. Acer Veriton GN100 AI Mini Workstation - the least independently reviewed of the eight, but listed as a launch partner for the new 64GB tier.

Every one of these is the same chip under a different logo. The question worth asking isn't "which GPU is faster," it's "which box's storage, warranty, and price fit what I'm actually going to do with it."

Spec and Price Table: ASUS Ascent GX10 vs DGX Spark, Dell, HP, Lenovo

The price column is the most useful part of this table and also the most volatile. ServeTheHome's chief analyst, Patrick Kennedy, pinned the cause on memory costs rather than NVIDIA margin: "we are paying $2000 for 64GB ECC RDIMMs, so pricing logically must increase." NVIDIA's own 128GB Founders Edition is up roughly 74% from its original $3,999 to $6,950 inside a year, while the OEM partners are launching a smaller 64GB tier at $4,999, a price that undercuts the current Founders Edition 128GB by nearly $2,000 for half the memory. Check each vendor's own listing before buying; none of these numbers will still be accurate in six months, and several of the table's "not disclosed" cells may fill in once wider retail coverage lands.

NVMe Generation and Cold-Load Speed: Why Storage Is the Real Differentiator

Since memory, compute, and bandwidth are fixed by the chip, storage generation is the one spec in this comparison that actually changes what the box feels like to use day to day. Loading a 70B-parameter checkpoint off disk at the start of a session, or swapping between two large models during an eval run, is bound by drive throughput, not GPU throughput.

StorageReview's teardown of the ASUS Ascent GX10 found the 1TB PCIe Gen4 Phison drive's write throughput peaking at just 0.72 GiB/s at the 16K block size, hit at 32 threads, the slowest peak among the drives it tested in this GB10 class, then declining to 0.62 GiB/s at 128 threads instead of improving further. That's the entry-level ASUS configuration. ASUS's own spec sheet lists a step-up path: the 4TB SKU moves to PCIe Gen5 NVMe, which should clear that bottleneck, though no independent review has benchmarked that specific drive yet.

Lenovo's reviewed configuration ships a 4TB PCIe Gen4 x4 M.2-2242 TLC drive standard. Gigabyte's AI TOP ATOM uses the same drive class: its review unit shipped at 1TB, with a 4TB model also listed for purchase. That puts Gigabyte's base configuration at the same capacity as ASUS's entry SKU, on the same Gen4 x4 drive class rather than the slower Phison drive StorageReview measured in the ASUS unit. HP's ZGX Nano is the one box in the lineup with no published storage spec at all. NVIDIA's own Founders Edition, by contrast, ships self-encrypting NVMe up to 4TB, per its product specifications.

If your workflow involves frequently swapping large checkpoints, the entry-level ASUS configuration is the one box in this lineup with a documented, independently-measured storage bottleneck. Lenovo's standard tier and every step-up option elsewhere in the lineup land on the same PCIe Gen4 x4 class drive or faster.

Support, Warranty, and Regional Availability

This is the thinnest part of the published record, and worth saying plainly rather than papering over. NVIDIA's own warranty terms are public: the standard DGX Spark warranty runs one year from the date of purchase, and it is explicitly voided by what NVIDIA calls "Enterprise Use," defined as large-scale datacenter or GPU cluster commercial deployment in its warranty documentation. That clause matters more than it looks: it means a DGX Spark bought for a developer's desk is not warranted for redeployment as a production cluster node, which is relevant if you're weighing one against renting cloud capacity for the same job.

None of the OEM partner reviews available at publication time, Dell, HP, Lenovo, or Gigabyte, disclosed specific warranty length or coverage terms for their GB10 boxes. Gigabyte's review noted a US partnership with AVADirect for "next-day on-site repair services" without specifying the warranty term those repairs fall under. Dell and Lenovo typically extend their existing business-line warranty and ProSupport-style tiers to new workstation SKUs, but neither company's GB10-specific warranty page was available to confirm that for this comparison, so treat it as an open question to raise with sales rather than an assumed fact. Regional availability is similarly thin in public reporting: NVIDIA's Founders Edition sells through Amazon, Micro Center, PNY, and TD Synnex in the US, while OEM partner availability varies by each company's existing retail and enterprise channels. Confirm both warranty terms and your region's availability directly with whichever OEM you're considering before ordering.

Which One to Buy for Your Use Case

Since the chip, memory, and bandwidth are identical everywhere, the decision collapses to three real variables: how much storage you need out of the box, what support relationship you already have, and what price you can actually get today.

• If you want the most independently verified box and don't mind configuring storage yourself: the ASUS Ascent GX10 has the deepest independent review coverage of any OEM partner, including a documented chassis revision that fixed early thermal headroom issues. Step up past the 1TB SKU to avoid its measured write-speed floor.
• If you need turnkey 4TB storage without a storage decision to make: the Lenovo ThinkStation PGX ships 4TB PCIe Gen4 NVMe as its standard configuration, and the Gigabyte AI TOP ATOM offers the same 4TB drive class as a step-up from its 1TB review unit.
• If your organization already runs on Dell, HP, or Lenovo procurement and support contracts: buying the matching GB10 box keeps the device inside an existing enterprise support relationship, even though none of the three publishes GB10-specific warranty terms independently yet. That existing relationship is worth more in practice than a spec-sheet difference that doesn't exist.
• If budget is the deciding factor and you can wait three weeks: the new 64GB tier launching Oct. 23, 2026 across Acer, ASUS, Dell, Gigabyte, HP, and MSI starts at $4,999, which is cheaper than NVIDIA's current 128GB Founders Edition price for half the memory. That trade only makes sense if your models fit comfortably under 64GB.
• If you specifically want NVIDIA's own build quality and direct support: the Founders Edition is the reference design every OEM box is measured against, at the cost of the steepest price increase in the lineup.

When Any of These Boxes Hits Its Ceiling

Every GB10 box, regardless of vendor, shares the same ceiling: 128GB of unified memory (64GB on the new tier) and 273 GB/s of bandwidth. That's enough to develop, fine-tune, and run single-user inference on models up to roughly 200B parameters with quantization, as our DGX Spark to GPU cloud pipeline guide covers in more detail. It is not enough for 5+ concurrent users hitting the same model, for 24/7 production uptime with an SLA, or for long-context workloads at scale, because KV cache pressure and total throughput both run into the same fixed bandwidth no matter which OEM's case it's in.

That's the point where renting data-center GPU capacity by the hour replaces buying a second desktop box. On Spheron, on-demand rates start at $0.78/hr for an RTX 5090 and scale up through $2.65/hr for H100 and $5.52/hr for H200 as of 05 Oct 2026, billed per minute after a 20-minute minimum with no long-term commitment, which is a materially different cost shape than a second $4,999-$6,950 desktop purchase for a workload that only needs to burst occasionally. Our GPU cloud benchmarks guide has throughput and pricing data across the full GPU catalog for sizing that next step. Spheron doesn't sell desktop hardware and has no stake in which OEM's GB10 box you buy; it's simply where the workload goes once a single box, from any vendor, stops being enough.

Pricing fluctuates based on GPU availability. Spheron rates above are live as of 05 Oct 2026; other providers and OEM prices reflect their most recently reported figures and may have changed. Check current GPU pricing → for live rates.

Picked your GB10 box and outgrown its 128GB ceiling? Rent H100 capacity by the minute instead of buying a second desktop unit.

Rent H100 on Spheron →