NVIDIA’s DGX Spark 64GB goes on sale on 23 October 2026, starting at US$4,999 through Acer, ASUS, Dell, Gigabyte, HP and MSI (NVIDIA). It is the cheapest way yet into NVIDIA’s deskside line, and it lands in the same month as RTX Spark laptops and a preview of DGX Station for Windows.
Put together, October 2026 is the month serious AI compute became a desk item.
What changed in October 2026?
Three launches moved large-model inference from the server room to the desk. As of 23 October 2026:
| System | Memory | What NVIDIA says it runs | Status |
|---|---|---|---|
| RTX Spark laptops | Up to 128GB unified | A 125B model, unmetered and offline | On sale 16 Oct 2026 |
| DGX Spark 64GB | 64GB unified (128GB with two units) | Up to 100B parameters (200B clustered) | On sale 23 Oct 2026, from US$4,999 |
| DGX Station for Windows | 748GB coherent | Up to trillion-parameter scale | Previewed 7 Oct 2026 |
Sources: NVIDIA on DGX Spark 64GB and NVIDIA on RTX Spark and DGX Station for Windows.
Two DGX Spark 64GB units link over a single QSFP cable at 200 gigabits per second and pool their memory to 128GB. NVIDIA says that pair delivered up to 1.7 times the performance of one unit in its own Qwen 3.8 27B test. Futurum notes the new unit starts above the original 128GB model’s US$3,999 launch price, but below today’s US$7,000 to US$9,000 range for 128GB systems (Futurum).
Why does deskside AI matter?
Deskside AI is data centre class inference that fits under a desk, runs on office power and needs no cloud account. Two years ago, running a 100 billion parameter model meant renting a GPU cluster. Now it means ordering a box.
That matters most where the cloud is a poor fit: a ground station between satellite passes, a substation control room, a vessel, a forward site. Those places never lacked the will to run AI. They lacked compute they could put on site. That excuse is gone.
Memory is still the constraint to plan around. Total parameters, precision and context length decide what fits, as our guide to GPU memory planning shows. A 64GB machine that is “up to 100B” at 4-bit is a very different machine at FP16.
What is still hard about local AI?
The hard part has moved from hardware to operations. A box on a desk answers “can we run it here?” It does not answer the questions that decide whether anyone should trust it:
- Which model is approved to run? Someone has to choose the model and version, and stop anything else loading.
- Does it fit and stay fitted? Context grows, models update, and memory runs out at the worst moment.
- How do you update fifty of them? One deskside box is a project. Fifty across remote sites is a fleet, and every update needs a rollback.
- What happens when the link drops or the room gets hot? Sites lose connectivity and cooling. The system has to keep serving and recover on its own.
- Can you prove what ran? After an incident, an operator, regulator or insurer will ask which model gave which answer, and when.
None of these is a hardware problem. All of them are why most AI at remote sites is still a pilot. We made the same point about outages in the cloud has a blast radius: the failure that hurts is rarely the chip.
What it means for operators
- Energy and utilities: local inference in the control room is now affordable. Approval, rollback and a record of what ran are what make it acceptable.
- Space and ground segments: a deskside system can sit in a ground station and work through pass windows without a cloud round trip.
- Defence: compute that fits in a vehicle or a tent is real. Running it when the network is gone or untrusted is the requirement.
- Everyone: budget for operations, not just the box. The purchase order is the easy part.
Where Sector88 fits
Sector88 is one system to deploy, run and govern AI at every site, connected or completely cut off. That covers choosing which model runs, keeping it serving when the network is gone, and keeping a durable record of what ran, so control stays with the operator. More on that in keeping operator control of AI at remote sites.
Compute just became a desk item. Control did not.