At ¥35-45k with 128GB unified memory and the ability to run 100B-scale LLMs locally, the NVIDIA DGX Spark (GB10, ~¥35k) and Apple Mac Studio M5 Max (~¥42-45k) are natural rivals. Yet they have completely different personalities: one 'thinks fast but talks with a stutter', the other 'talks smoothest but is lopsided at work'. This post uses third-party benchmarks to show exactly where each wins.

Hardware specs explain a lot, but the runtime decides the rest.

The core difference: prefill (understand) vs decode (output)

An LLM call has two phases: prefill reads the whole prompt at once (compute-bound, throughput), and decode emits tokens one by one (bandwidth-bound, latency). DGX Spark packs 1 PFLOP FP4 but just 273 GB/s; the M5 Max's 614 GB/s is 2.25× higher. The result splits naturally:

On prefill, DGX is ~3× the Mac (~1200 vs 400 tps), great for long-doc analysis, RAG and tool-calling where you feed big prompts in; on decode the Mac nearly doubles it, great for reading long outputs. Repeatedly ask the AI to work → DGX; read long AI output → Mac.

Finetuning & batch: DGX wins big (but most users won't need it)

Reason: DGX Spark runs the full CUDA stack (vLLM, complete finetuning tooling) while the Mac only has MLX/Metal with no CUDA-optimized batch pipeline. But if you're building a personal assistant or doing day-to-day inference, the finetuning/batch edge rarely matters — it serves professional research.

Power & electricity: the Mac is far cheaper

Run inference 24/7 and the Mac's power bill is under half of DGX's; it also wins on noise and heat. DGX draws ~240W at full load despite the tiny box. For marathon runners, the Mac — the electricity gap adds up over a year.

One M5 Max vs two DGX Sparks (the $10K showdown)

ZDNet staged an interesting matchup: one M5 Max 128GB ($7,999) vs two DGX Sparks ($9,398), running DeepSeek-V4-Flash (284B MoE) across Terminal-Bench 2.1's 89 tasks.

Task completion was nearly tied, but the dual Sparks win on decode speed, 8-way concurrency and context length. One person → M5 Max; many agents or a shared team → dual DGX Spark.

A 'spicy' hack: PD separation

36Kr measured DGX Spark at just 20 tok/s decode at batch=1 (370 at batch=32). Someone pipelined it with a Mac Studio M3 Ultra (819 GB/s): DGX does prefill, the Mac does decode, speeding up total inference 2.8× — at triple the cost (~¥100k). A tinkerer's flex, most users can just watch.

So which one?

• AI research/finetuning/batch, need CUDA, want prefill speed → DGX Spark (expandable, but Linux-only, power hungry)
• Personal assistant/creative work, want smooth output, live in macOS → Mac Studio M5 Max (top bandwidth, quiet and efficient, but pricey with weak tuning)
• Not sure → both are pro-oriented; for occasional tinkering, a discrete-GPU build or waiting for DRAM prices to fall is more sensible.

Buy the Mac for a personal assistant; buy the DGX for developing, batching and finetuning. Both are great machines — just built for different people.

Data notes: speed/finetune/power figures from third-party reviews (AI Bytes, Presenc AI, ZDNet, 36Kr, Zhihu, 2026), vary by environment, for reference only.