From CES to your desk: the DGX Spark rollout and NVIDIA's ecosystem push

It is easy to forget how quickly the GB10 went from a slide behind a keynote to a box humming on someone's desk. The chip that powers GB10 Studio did not arrive fully formed; it arrived as a preview, then a product, then an ecosystem. This is the short story of how that happened — and why it is the reason a marketplace for GB10 compute can exist at all.

The CES reveal

NVIDIA first showed the world a GB10-based personal AI supercomputer at CES, under the name Project DIGITS. The pitch was almost absurd on paper: take the Grace Blackwell architecture that powers data-center racks, shrink the relevant slice of it down to something that fits on a desk, and let a developer run serious local inference without a server room or a cloud bill.

The reaction was the kind you only get when a demo crosses a threshold people did not expect to be crossed yet. A petaflop-class machine you could put next to your monitor, with enough unified memory to hold large models in one place, reframed what "local" could mean. For the first time, the answer to "where does the model run" could plausibly be "right here," and it would not be a toy.

Project DIGITS becomes DGX Spark

The preview matured into a product. Project DIGITS was rebranded the DGX Spark and moved toward general availability — the same GB10 Grace Blackwell silicon, now with a name that placed it squarely in NVIDIA's DGX family of AI systems rather than off to the side as a curiosity.

Crucially, NVIDIA did not keep the form factor to itself. Its hardware partners began building their own GB10-based machines, so the same core platform started shipping in more than one chassis from more than one vendor. That matters more than it sounds: a single reference box is a demo, but a class of machines from multiple system builders is a category. The GB10 stopped being a thing one company made and became a thing you could go buy from several.

The software catches up

Hardware is only half of a platform. What made the DGX Spark land was that the software people already knew ran on it on day one. Because GB10 is Grace Blackwell, it speaks the standard NVIDIA AI software stack — the same CUDA, the same libraries, the same NIM microservices — rather than some bespoke embedded fork you would have to learn from scratch.

On top of that, the local runtimes developers actually reach for came along too. Ollama, LM Studio, and vLLM all run on the platform, which means the tools you used on a laptop or a workstation behave the way you expect on a GB10 — just with far more headroom. NVIDIA's developer blog followed with software optimizations and walkthroughs specifically for the DGX Spark, turning early-adopter folklore into documented, repeatable practice.

The arc, compressed, looks like this:

1. CES — a GB10 personal AI supercomputer previewed as Project DIGITS.
2. GA — rebranded DGX Spark, generally available, with partners shipping their own GB10 systems.
3. Ecosystem — CUDA, NIM, Ollama, LM Studio, and vLLM all running on the box, with optimizations published on NVIDIA's developer blog.

What it means now

Put it together and the GB10 has quietly completed the journey from keynote demo to a real thing on real desks. It is no longer hypothetical compute you have to imagine; it is hardware you can buy from NVIDIA or its partners, running software you already use, exposing an API any client can talk to.

That last part is the whole reason GB10 Studio works. When the platform standardized on OpenAI-compatible inference and common runtimes, every GB10 became interchangeable from a buyer's point of view — a session is a session, no matter whose box answers it. A marketplace needs supply that is real, uniform, and idle some of the time. The DGX Spark rollout delivered exactly that: a population of genuine Grace Blackwell machines, on desks all over, waiting to be put to work.