DGX Spark

128 GB · 128GB unified memory · $5,100 Full rig

Quick decision verdict

DGX Spark outputs 48.7 tok/s at $0.097/M tokens (rank #15 of 28 measured rigs on unit cost), breaking even in 1,062 days at full load.

  • Generation speed rank #19 of 29 · System power rank #6 (170 W).
  • For a lower upfront build cost, compare against RTX 4090 ($4,199)
  • If you need faster interactive token speed, compare against RTX PRO 6000 Blackwell (171.1 tok/s)

Can I Run It? → · Best → · Method →

How you'll use it

Fits?Source
DGX Spark
128GB unified memory
128 GB2,228.548.7Roomy170 W$5,10035.2$0.097$0.61$5.411,062 daysMeasured

Selected hardware

DGX Spark

128 GB · 128GB unified memory

Read tok/s
2,228.5
Write tok/s
48.7
Power
170 W
Output/day
35.2
Cost/M tok
$0.097
Power/day
$0.61
Saved/day
$5.41
Payback
1,062 days

Details · See source

FAQ

How fast does DGX Spark run local LLMs?
DGX Spark (128 GB) measures 48.7 tokens/s generation and 2,229 tokens/s prompt reading on Qwen3.8-Flash-Next (snapshot 2026-10-07).
Is DGX Spark worth buying over cloud LLM APIs?
At $0.097/M tokens local cost, DGX Spark breaks even in 1,062 days at 24h/day utilization, or 6,371 days at 4h/day — making daily utilization the main deciding factor.