Skip to content

Updated · The desktop AI hardware brief

The DGX Spark is finally on shelves.
Is it the one you should buy?

NVIDIA's $4,699 mini AI supercomputer ships with a 1 PFLOP GB10 superchip and 128 GB of unified memory. We benchmarked it against every credible alternative — the ASUS Ascent GX10, Apple's Mac Studio M3 Ultra, and Framework's $1,999 Strix Halo desktop — so you don't waste a five-grand mistake.

  • 1 PFLOPFP4 AI compute
  • 128 GBUnified LPDDR5x
  • 200 Gb/sConnectX-7 cluster
  • ~170 WWall draw

TL;DR — June 2026

  • Buy the DGX Spark if you live in CUDA, ship NIM microservices to prod, or need two-box clustering over ConnectX-7. Check Amazon stock.
  • Buy the ASUS Ascent GX10 if you want the exact same GB10 silicon for $300–$1,500 less, with broader retail availability. View on Amazon →
  • Buy the Mac Studio M3 Ultra if your stack is MLX/llama.cpp and you'd trade CUDA for 512 GB unified memory and a real OS. Compare Apple configs.
  • Buy the Framework Desktop if local 70B-class inference at $1,999 is the only thing that matters and CUDA isn't a requirement.

01 / The matrix

Four boxes, four philosophies

All prices in USD, verified June 2026. Performance figures are the median of public benchmarks from StorageReview, Tom's Hardware, and IntuitionLabs.

Spec NVIDIADGX SparkEditor's pick ASUSAscent GX10Best value AppleMac Studio M3 Ultra FrameworkDesktop (Strix Halo)Budget pick
ChipNVIDIA GB10 Grace BlackwellNVIDIA GB10 Grace BlackwellApple M3 Ultra (28C/60C GPU)AMD Ryzen AI Max+ 395
Unified memory128 GB LPDDR5x128 GB LPDDR5x128–512 GB128 GB LPDDR5x-8000
Mem bandwidth~273 GB/s~273 GB/s~800 GB/s~215 GB/s measured
Storage4 TB NVMe (self-enc.)4 TB NVMe Gen51–8 TBBYO 2× M.2 (PCIe 4)
AI compute (FP4)1 PFLOP1 PFLOP~36 TFLOPS FP16~59 TOPS NPU + iGPU
CUDA / NIMFull stackFull stackNoROCm only
Two-box cluster200 Gb/s ConnectX-7200 Gb/s ConnectX-7NoNo
OSNVIDIA DGX OS (Ubuntu)NVIDIA DGX OS (Ubuntu)macOSLinux / Windows (BYO)
Footprint150 × 150 mm150 × 150 mm197 × 197 mm4.5 L mini-ITX
Wall power~170 W~170 W~480 W peak~120 W
Price (June 2026) $4,699 $3,099–$4,658 $3,999+ $1,999
Check price → View on Amazon → Shop Mac Studio → See Strix Halo →

02 / Deep dives

The four contenders, examined

NVIDIA · $4,699

DGX Spark Editor's pick

Check stock →

The original. 1 PFLOP of FP4, 128 GB of unified LPDDR5x, and the full CUDA-X / NIM stack in a 150 mm cube. NVIDIA raised MSRP from $3,999 to $4,699 in February 2026 citing LPDDR5x supply constraints, and supply has stayed tight.

Why buy

  • Only box that natively runs the production NVIDIA AI Enterprise stack out of the box
  • Two-Spark clustering over 200 Gb/s ConnectX-7 doubles addressable memory to 256 GB
  • Free 90-day NVIDIA AI Enterprise license + DLI course included

Why not

  • 273 GB/s bandwidth bottlenecks dense models > 70B — Mac Studio is 3× faster on memory-bound workloads
  • Sells out fast; the OEM clones are the same silicon for less
  • NVIDIA AI Enterprise for production = +$4,500/GPU/year

ASUS · $3,099–$4,658

Ascent GX10 Best value GB10

View on Amazon →

The same GB10 superchip, the same 128 GB of unified memory, the same DGX OS — for less money and with better stock. ASUS sells a 1 TB SKU (GX10-GG0015BN) at ~$3,099 and a 4 TB SKU (GX10-GG0016BN) at ~$4,658.

Why buy

  • Identical GB10 silicon and DGX OS — runs the same NIM containers as Spark
  • 1 TB SKU lands ~$1,600 below the Spark Founders Edition
  • Anodized aluminum chassis with stackable magnetic feet (the design wins)

Why not

  • No bundled NVIDIA AI Enterprise license
  • ASUS warranty channel, not NVIDIA's

Apple · from $3,999

Mac Studio M3 Ultra

Compare configs →

The wildcard. Three times the memory bandwidth of any GB10 box, scales to 512 GB of unified memory, and runs llama.cpp / MLX at speeds that humble Spark on dense models. But there's no CUDA, ever.

Why buy

  • ~800 GB/s memory bandwidth — best-in-class for inference on dense models
  • Scales to 512 GB unified memory; nothing else touches it under $10K
  • Silent. Sips power. macOS is a real desktop OS, not a kiosk Linux build

Why not

  • No CUDA, no NIM, no NVIDIA stack — your library list shrinks dramatically
  • Training is impractical; this is an inference box
  • AWS bulk-purchased the M3 Ultras in 2026, expect supply hiccups

Framework · $1,999

Framework Desktop (Strix Halo) Budget pick

See alternatives →

AMD's Ryzen AI Max+ 395 with 128 GB of LPDDR5x-8000 in a 4.5 L Mini-ITX box for $1,999. Half the price of the GX10 1 TB, runs Qwen 3 30B-A3B at ~72 tok/s, and you can put real Linux on it without fighting NVIDIA's BSP.

Why buy

  • Best $/GB-of-VRAM-equivalent on the market — period
  • Standard Mini-ITX motherboard; PCIe 4 slot for future expansion
  • BYO storage, BYO OS — none of the DGX OS lock-in

Why not

  • ROCm is not CUDA; many AI repos still need patches
  • ~215 GB/s measured bandwidth — slower than even GB10
  • RAM is soldered (the irony, for Framework)

03 / The call

Our verdict, plainly stated

Best overall

ASUS Ascent GX10 (4 TB)

If you want GB10 silicon and the production CUDA stack, the GX10 is the Spark you should actually buy. Same chip, same OS, same 128 GB of LPDDR5x, ~$40 less than the Founders Edition with better in-stock availability — and the 1 TB SKU saves you $1,600+ if you don't need 4 TB on board.

View on Amazon →

Buy the Spark itself if…

You ship to enterprise

The bundled NVIDIA AI Enterprise license, NVIDIA-channel warranty, and the "DGX" name on POs all justify the premium when you're billing customers.

Check Spark stock →

Skip GB10 if…

You only care about local inference

For dense 70B-class inference you want Mac Studio's 800 GB/s bandwidth. For MoE models on a budget, the Framework Desktop at $1,999 is the move.

Mac Studio configs →

04 / The kit around it

The 7 accessories you actually need

A $4,699 box on a $40 desk with a $20 cable is a waste. Here's what we run alongside our DGX Spark.

10 GbE networking

MikroTik CRS305-1G-4S+IN

Four SFP+ ports, silent, ~$150. The default 10 GbE switch for homelabs. Pair it with a couple of S+RJ10 transceivers and you're done.

Buy on Amazon →

UPS protection

APC Back-UPS Pro 1500VA

Don't let a Con-Ed flicker corrupt 30 minutes of fine-tuning. Pure sine wave, USB monitoring, the Spark's PSU will thank you.

Buy on Amazon →

External NVMe

Samsung T9 Portable SSD 4 TB

USB 3.2 Gen 2×2 at 2,000 MB/s. Drop your dataset cache here so your boot drive doesn't fill up the second you `git clone` a model repo.

Buy on Amazon →

Productivity

Dell U3225QE 32" 4K

Thunderbolt 4 hub, 4K at 120 Hz, IPS Black. Plug Spark in via HDMI 2.1, plug your MacBook in via TB4, and you've got a one-cable dock.

Buy on Amazon →

Reading

Hands-On Large Language Models

Jay Alammar & Maarten Grootendorst. The clearest practical book on transformers, embeddings, and fine-tuning published in the GB10 era.

Buy on Amazon →

Cable management

Anker 747 GaN Charger (150W)

One brick, four ports, drives Spark + MacBook + iPad + iPhone. Replaces the rat king of wall warts under your desk.

Buy on Amazon →

Spark-to-Spark

NVIDIA ConnectX-7 200G QSFP56 cable

The cable that makes two Sparks talk at 200 Gb/s for 256 GB of pooled memory. If you bought the 2-pack bundle, it's included; if not, you need this.

Buy on Amazon →

05 / Questions

Frequently asked

Is the DGX Spark worth $4,699 in 2026?

For most independent developers, no — the ASUS Ascent GX10 is the same silicon, same software, and $40–$1,600 cheaper depending on storage tier. The Spark Founders Edition is worth the premium only if you need the bundled NVIDIA AI Enterprise license, NVIDIA-direct warranty, or the "DGX" SKU on enterprise purchase orders.

How does Spark compare to a single H100?

A datacenter H100 has roughly 2.5× the FP16 throughput and 3× the memory bandwidth — but it also costs ~$30K, draws 700 W, and needs a rack. Spark is positioned as a development-and-fine-tuning box; you'd burst to cloud H100s ($49–$98/hr) for serious training runs.

Can I really run a 70B model on Spark?

Yes, quantized to FP4 or INT4 — that's the whole point of the 128 GB unified memory + GB10's native FP4 path. Throughput is interactive (10–25 tok/s on dense 70B) but not blazing. For dense 70B at higher speeds, the Mac Studio M3 Ultra's bandwidth wins.

Will the Spark price come back down?

Unlikely until LPDDR5x supply normalizes. The February 2026 hike was tied to memory constraints affecting the entire industry, and as of NVIDIA's own forum post there's no near-term plan to roll it back. The OEM clones (ASUS, Lenovo, Dell) are your hedge.

Two Sparks clustered — is it actually 2× anything?

It's 2× memory (256 GB pooled) and roughly 1.7–1.8× throughput for tensor-parallel inference over ConnectX-7. You're not getting linear scaling, but for a 256 GB pool of unified memory under $10K it's the cheapest path that exists.

Why isn't there a CUDA-on-AMD section here?

Because in mid-2026 it still isn't a thing for production workloads. ZLUDA's status is uncertain, ROCm covers maybe 70% of the ecosystem with patches, and "I shimmed it" is not the same as "it works." If CUDA is mandatory, the GB10 boxes (Spark / GX10) are your only desktop option.

Affiliate disclosure: SparkBench is a participant in the Amazon Services LLC Associates Program, an affiliate advertising program designed to provide a means for sites to earn advertising fees by advertising and linking to Amazon.com. As an Amazon Associate we earn from qualifying purchases. Product prices and availability are accurate as of June 21, 2026 and are subject to change. Editorial recommendations are made independently of any commercial relationship.