Breaking

Desktop-Sized NVIDIA DGX Spark Serves 64 Concurrent Users at 700+ Tokens/Second on 38 Watts in New Benchmark

July 12, 2026 at 13:15 EDT

Infra & Chips
Foundation Models
Sales & Business

A single NVIDIA DGX Spark, the compact desktop machine built on the GB10 Grace Blackwell Superchip, has been shown serving 64 simultaneous users at more than 700 tokens per second while drawing just 38 watts—a demonstration positioning the device as capable of running a small company's entire AI workload from a desk.

Mid-July 2026 · DGX Spark Benchmark

64 Users, One Desktop, 38 Watts

A single NVIDIA DGX Spark served 64 simultaneous users at more than 700 tokens per second while drawing just 38 watts — running the Qwen3.6-35B model and positioning the compact machine to handle a small company's entire AI workload from a desk.

simultaneous users, each with an independent prompt & KV cache

700+

tokens/sec aggregate output across all sessions

38 W

measured power draw during the test

Why concurrency lifts throughput

With continuous batching, GPU utilization rises sharply as more users share the machine — aggregate token output scales up dramatically.

Low

Single user
memory-bound

700+ t/s

64 users
continuous batching

GB10 Grace Blackwell Superchip

20 cores

Arm CPU (10× X925 + 10× A725) + Blackwell GPU

128 GB

unified LPDDR5x · 4TB NVMe · ConnectX-7 NIC

1 PFLOPS

FP4 AI perf · up to ~200B params quantized (NVFP4)

150×150×50.5 mm chassis · ~$4,000 (Founders Edition) · DGX OS + 90-day AI Enterprise license

The upside

High aggregate throughput and low power in a desk-sized box — practical for local LLM inference, running multiple agents at once, and fine-tuning. Units can be clustered for larger models.

The caveat

It remains fundamentally a development kit, not a hardened production rig. ARM64 architecture and a still-maturing ecosystem can require extra effort for compatibility and optimization.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.