How much does the DGX Spark cost in 2026?

$4,699 as of August 2026. NVIDIA raised the price from $3,999 by an official $700 in February 2026, citing memory supply constraints. This was an MSRP change, not street drift. The ASUS Ascent GX10, built on the same GB10 chip, lists at $3,999 for the 1TB configuration and is now the cheapest GB10 box.

How fast is the DGX Spark at running local LLMs?

It depends heavily on the software stack, which is the most misunderstood fact about this box. On gpt-oss 120B, community benchmarks recorded roughly 11.7 tokens per second decode under Ollama, 38.5 under a tuned llama.cpp build (numbers posted by the llama.cpp author), and around 50 under SGLang. Same hardware, roughly a 4x spread. The 273 GB/s memory bandwidth is the hard ceiling underneath all of it.

Is the DGX Spark good value compared to a Ryzen AI Max 395 box?

On tokens per second per dollar, no. A 128GB Ryzen AI Max+ 395 box costs roughly $2,000-2,400 and posts broadly similar generation speeds on MoE models, because both machines are bandwidth-limited to a similar order. What the Spark uniquely offers is the NVIDIA stack: CUDA, TensorRT, NIM containers, and behavior that matches the datacenter hardware you deploy to. If that sentence means nothing to your workflow, the AMD box keeps $2,300-2,700 in your pocket.

Who should actually buy a DGX Spark?

Three buyers: developers who prototype locally and deploy to NVIDIA datacenter GPUs and need CUDA-identical behavior; teams using the DGX/NIM software ecosystem who want a desk-side node; and buyers who need two Sparks linked for the 200B-class and long-context configurations NVIDIA markets. For everyone else — hobbyists, agent-runners, model tinkerers — a cheaper unified-memory box or a used-GPU build does the same work for less.

Can the DGX Spark run a dense 70B model well?

It fits easily in 128GB, but a dense 70B must read about 40GB of weights per token, and at 273 GB/s that caps generation in the single digits of tokens per second. The Spark is at its best on MoE models with small active parameter counts, and on prefill-heavy and batch workloads, where its compute and FP4 support matter more than bandwidth.

← All guides

Hardware August 11, 2026

Is the NVIDIA DGX Spark Worth $4,699 for Local AI? (August 2026)

The DGX Spark is the most argued-about box in local AI. It packs a GB10 Grace Blackwell chip and 128GB of unified memory into something that sits on a desk — and its 273 GB/s of memory bandwidth is a sixth of an RTX 5090's. Whether it is worth $4,699 depends entirely on which of three buyers you are, and most people asking are the wrong one.

Bottom Line

Price: $4,699 as of August 2026 — NVIDIA added an official $700 in February, blaming memory supply. The ASUS Ascent GX10 (same GB10 chip) now undercuts it at $3,999/1TB.
Speed: stack-dependent to a degree nobody expects: gpt-oss 120B decode ran ~11.7 tok/s on Ollama, ~38.5 on tuned llama.cpp, ~50 on SGLang — same box, ~4x spread.
The real product: CUDA parity with your deployment target, on a desk. Not tokens per second.
Verdict: worth it for NVIDIA-stack developers. Not worth it as a pure inference box — a ~$2,000-2,400 Ryzen AI Max+ 395 machine does that job at half the price.

What $4,699 actually buys

GB10 Grace Blackwell, 128GB unified memory, ConnectX networking for pairing two units, and the full NVIDIA software stack in a quiet desktop box. The price history matters: it launched at $3,999, and in February 2026 NVIDIA raised it to $4,699 in an official notice citing memory supply constraints — the same shortage that reset RAM and GPU prices everywhere.

The spec that defines it is 273 GB/s of memory bandwidth. An RTX 5090 moves ~1,792 GB/s — about 6.5x more. Token generation is bandwidth-bound, so no amount of Blackwell compute makes single-stream chat fast on this box. That is not a flaw discovered by reviewers; it is the design tradeoff: capacity and ecosystem over bandwidth.

The benchmark chaos, explained

The Spark’s launch reviews disagreed with each other by 4x, and both sides were measuring honestly:

Stack	gpt-oss 120B decode	Source
Ollama	~11.7 tok/s	community review benchmarks
llama.cpp (tuned)	~38.5 tok/s	posted by the llama.cpp author
SGLang	~50 tok/s	NVIDIA developer forum testing

Two lessons. First, on this machine the software stack is worth more than 3x the hardware delta between it and its rivals — if you buy one, do not run Ollama defaults and conclude you were robbed. Second, all three numbers obey the same ceiling: gpt-oss 120B is MoE with a small active set, which is why it moves at all. A dense 70B reading ~40GB per token caps in single digits — if dense 70B is your goal, that money buys a much better rig.

The three buyers it is right for

The deploy-to-NVIDIA developer. You prototype locally, ship to A100/H100/Blackwell in production, and need identical CUDA/TensorRT behavior. The Spark is the only 128GB desk box that gives you that. This is its honest, load-bearing use case.
The DGX/NIM ecosystem team. Desk-side node, same containers, same tooling as the fleet.
The two-Spark buyer. NVIDIA’s 200B-class and 1M-context configurations pair two units over ConnectX. Niche, real, and nothing else at this price does it.

If you are none of these, you are buying a badge.

What to buy instead

Same chip, less money: the ASUS Ascent GX10 — the same GB10 with 128GB — lists at $3,999 (1TB) as of August 2026, making it the cheapest way into this exact platform. The Spark’s extras (2TB, branding, ConnectX topology details) are worth checking against a $700 gap.

Same job, half the money: a 128GB Ryzen AI Max+ 395 box (~$2,000-2,400 street, volatile) is bandwidth-limited to the same order and runs the same MoE models at broadly comparable speeds — our model picks for it show gpt-oss 120B at 31-55 tok/s. No CUDA. If your stack is llama.cpp/Ollama anyway, that is $2,300-2,700 saved.

More speed, less capacity: if your models fit in 24-48GB, a used-GPU build demolishes every unified-memory box on generation speed.

The three boxes from this page:

NVIDIA DGX Spark 128GB — the CUDA-parity dev box.
ASUS Ascent GX10 — same GB10 chip, now the cheaper way in.
AMD Ryzen AI Max+ 395 128GB box — the half-price alternative if you don't need CUDA.

Amazon affiliate links — we earn a small commission at no cost to you.

Is the NVIDIA DGX Spark Worth $4,699 for Local AI? (August 2026)

Bottom Line

What $4,699 actually buys

The benchmark chaos, explained

The three buyers it is right for

What to buy instead

Sources

See Also

Need OpenClaw fixed live?

Read next