THE CRUNCH
Apple's Mac Studio with the M5 Ultra chip has taken a decisive win over Nvidia's DGX Spark platform in detailed local AI testing by Tom's Hardware. Using Artificial Analysis' AgentPerf-Local benchmark, which replays recorded inference traces across agentic workloads, the Mac Studio finished the test in just under half the time of a Dell Pro Precision GB10 box, a system nearly identical to the DGX Spark. With both machines running the dense Qwen 3.8 27B model through llama.cpp, the Mac also consumed about 16% less task energy, giving it an energy-delay product under half that of the Spark-like system.
The comparison is less lopsided than it sounds on paper. The tested 4TB Dell GB10 now costs around $10,000, so pitting it against the roughly $12,000 Mac Studio is a fair fight on price, even if the Mac's 80-core GPU and 1.2 TB/s of memory bandwidth dwarf the GB10's 48 Blackwell SMs and 273 GB/s on spec sheets.
The picture shifts with larger models. Quantised versions of GLM 5.3 Flash and DeepSeek V4.1 Flash 1017 fit natively in the Mac Studio's 256GB of unified memory, but cannot fit on a single 128GB DGX Spark, forcing buyers into two-machine clusters. A dual-Dell cluster with 4TB SSDs would run to $20,000 or more, and even a more typical $15,000 setup is about 25% dearer than one Mac Studio. Rising GB10 prices, with Lenovo listing its ThinkStation PGX as coming soon during testing, have muddied the value picture.
Context from the machine's September launch: Apple discontinued the Mac Pro earlier in 2026, leaving the Mac Studio as its top-end desktop, and the tested configuration cost $12,299 with 256GB of unified memory and 4TB of storage. The publication cautions that a different inference engine could flatter the Spark, and that Apple Silicon's limited engine choice forces llama.cpp for DeepSeek V4.1 Flash.