The Desktop AI Hardware Flood Is Real, and Most of It Ships in Weeks
DGX Spark, DGX Station, RTX Spark, Agentic PCs -- local inference is becoming a consumer category.
October was the month local AI hardware stopped being a conference demo. NVIDIA's DGX Station for Windows, announced October 7, carries 748GB of coherent memory and up to 20 petaflops and is slated for a fourth-quarter release. The 64GB DGX Spark ships October 23; RTX Spark laptops target personal agents with 128GB of unified memory and one petaflop of AI compute.
On the CPU side, AMD said Agentic PC shipments passed 500,000 units, and Microsoft reported Copilot+ PCs crossing two trillion local AI inferences a month. Samsung's first AI glasses are dated for November on Gemini, and Apple's homeOS details point to a Siri-centered home hub.
The memory math deserves scrutiny. Reports indicate the 128GB DGX Spark SKU exposes only about 110GB to the GPU once Windows claims its share -- a gap that decides which model actually fits. Buy the spec you'd really run, not the trillion-parameter headline.
For builders, the message is that private, offline inference is now a procurement decision, not a research project. The question shifts from 'can it run locally' to 'which local model do I trust with my data.'
Updated October 28, 2026: the 64GB DGX Spark began shipping October 23 as scheduled.
Frequently asked questions:
Q: Which local AI hardware shipped in October 2026?
A: NVIDIA's DGX Station for Windows (748GB coherent memory, Q4), the 64GB DGX Spark (shipped October 23), and RTX Spark laptops targeting personal agents with 128GB unified memory.
Q: How big is the local AI market now?
A: AMD said Agentic PC shipments passed 500,000 units, and Microsoft reported Copilot+ PCs crossing two trillion local AI inferences a month.
Q: What spec should I actually check?
A: Usable memory, not headline memory - the 128GB DGX Spark SKU reportedly exposes only about 110GB to the GPU once Windows claims its share. Buy the spec you'd really run.