Nvidia is introducing a 64 GB version of its DGX Spark / GB10 local-AI desktop, Tom’s Hardware reported, with systems from Acer, Asus, Dell, Gigabyte, HP, and MSI slated to start at $4,999 when they launch Oct. 23. The cut comes as LPDDR5X prices soar and dense open models such as Qwen 3.8 27B can fit in far less than the original Spark’s 128 GB for many inference jobs.

The 64 GB Spark keeps the same ConnectX-7 RDMA networking, 20-core Arm CPU complex, and about 273 GB/s shared memory bandwidth as the 128 GB models, so users can still cluster units later with Nvidia’s Sync Cluster Assistant. In-stock 128 GB systems have been selling roughly $7,000–$9,000, well above adjusted MSRP, turning excess soldered memory into a barrier. Forum reaction noted some retailers still list older 128 GB packs near $4,900 — a reminder memory markets are chaotic.

Nvidia’s broader blog push on GPUs powering OpenAI GPT-6 / Astra work underscores that cloud demand remains the company’s core business; Spark is a developer lifeline, not a replacement for datacenter GB clusters.

Backstory: local AI boxes rose as cloud GPU rents and RAM shortages blocked hobbyists. Engage: partner store pages on Oct. 23 and Nvidia Sync docs.

What we know:
64 GB Spark SKU from ~$4,999; Oct. 23 partner launch; same NIC/CPU/bandwidth as 128 GB; clustering still supported.

What we don’t:
How long the $4,999 street price holds as RAM moves; real-world LLM context limits at 64 GB; whether 128 GB MSRPs stabilize.

Agree (full): Tom’s Hardware and Nvidia materials agree a lower-memory Spark SKU is arriving to address RAM cost and local inference needs.
Differ (full): Nvidia pitches a lifeline; community voices call $5,000 still steep versus Mac mini-class alternatives and question value versus leftover 128 GB stock.