Sunk Cost sunkcost.ai Data checked 2026-09-03

DGX Spark GB10 Grace Blackwell, 128GB vs Strix Halo GMKtec EVO-X2, 128GB for local AI

Both hold 39 of the 48 current models here. The GMKtec EVO-X2, 128GB costs $1,199 less. On Qwen3.8 27B, the strongest model both hold, the DGX Spark, 128GB is about 1.1× faster: 11 tok/s against 10, both estimated from memory bandwidth. The GMKtec EVO-X2, 128GB pays for itself sooner, in 267 years against 346 years at 500k tokens a day.

DGX Spark GB10 Grace Blackwell, 128GBStrix Halo GMKtec EVO-X2, 128GB
Price$4,699$3,500
Memory128 GB128 GB
Usable by the GPU119.5 GB96 GB
Memory bandwidth273 GB/s256 GB/s
Power under load150 W150 W
Models that fit3939
Best model it runsQwen3.8 27BQwen3.8 27B
Speed on that model11 tok/s estimated10 tok/s estimated
Pay-back on that modelPays back in 346 yearsPays back in 267 years

Run the numbers on the DGX Spark GB10 Grace Blackwell, 128GB · or the Strix Halo GMKtec EVO-X2, 128GB

How much use it takes to pay back

Everything above is at 500k tokens a day. Pay-back moves with how much you actually run, so here are both machines on Qwen3.8 27B, the strongest model both hold, at the five levels of use the calculator names. The Strix Halo GMKtec EVO-X2, 128GB pays back sooner at every level of use, so this is not a choice that turns on how hard you work it.

A day's useDGX Spark GB10 Grace Blackwell, 128GBStrix Halo GMKtec EVO-X2, 128GB
50ka few chats a day3,456 years2,670 years
200klight assistant use864 years667 years
1Ma moderate coding-assistant day173 years133 years
4Mheavy coding with an agent43 years33 years
20Magents running most of the day11 yearsits ceiling9.4 yearsits ceiling

On Qwen3.8 27B neither machine can generate 20M tokens a day: the DGX Spark GB10 Grace Blackwell, 128GB manages at most 15.2M and the Strix Halo GMKtec EVO-X2, 128GB at most 14.3M. Both figures on that row are for the most each can do.

Run the DGX Spark GB10 Grace Blackwell, 128GB at 20M tokens a day · or the Strix Halo GMKtec EVO-X2, 128GB

The same models, not to the same length

Every model on this list that fits one machine fits the other at 32k of context, so memory does not change what they run. What it changes is how far you can take the context on one of them. The DGX Spark GB10 Grace Blackwell, 128GB has 119.5 GB usable against 96 GB, and spare memory is what the KV cache grows into as you keep more tokens.

ModelDGX Spark GB10 Grace Blackwell, 128GBStrix Halo GMKtec EVO-X2, 128GB
Devstral 2 123B75 GB of weights64k32k

Each figure is the longest context the calculator offers that the machine still holds that model at, and no model is taken past its own context limit. The other 47 models the calculator counts reach the same length on both machines, at every setting from 4k to 256k.

Run Devstral 2 123B on the DGX Spark GB10 Grace Blackwell, 128GB at 64k · or on the Strix Halo GMKtec EVO-X2, 128GB at 32k

The DGX Spark, 128GB is also head to head with another computer: Mac mini M5 Pro, 24GB · Mac Studio M5 Max, 128GB · MacBook Air M5 (15-inch), 16GB · MacBook Pro M5 Pro (16-inch), 64GB · Corsair AI Workstation 300, 128GB. With a graphics card: RTX PRO 6000 Blackwell, 96GB · Radeon AI PRO R9700, 32GB.

The GMKtec EVO-X2, 128GB is also head to head with another computer: Mac mini M5 Pro, 24GB · Mac Studio M5 Max, 128GB · MacBook Air M5 (15-inch), 16GB · MacBook Pro M5 Pro (16-inch), 64GB · Framework Desktop, 128GB · Mac Studio M5 Max, 64GB. With a graphics card: RTX PRO 6000 Blackwell, 96GB · Radeon AI PRO R9700, 32GB. With the same machine at another memory size: 64GB.

The assumptions behind both columns

Both columns use the same usage: 500k tokens a day at 15:1 input to output, 32k of context, $0.17 per kWh, and today's API prices held flat. Speeds marked estimated are worked out from memory bandwidth rather than measured, and pay-back scales with them. Change any of it in the calculator.

More head to head: everything the DGX Spark GB10 Grace Blackwell, 128GB runs · everything the Strix Halo GMKtec EVO-X2, 128GB runs · every other match-up · the quickest pay-back at each level of use · every model against the frontier