Enterprise AI Infrastructure

Every AI pilot that matters gets stuck at the same gate: legal will not let the data leave the building. A desktop box that costs about $4,000 walks straight past that gate, and a $189 management license is what makes the walk defensible.

By Shashi Bellamkonda · July 6, 2026

Legal says no. That is where most enterprise AI pilots die in healthcare, financial services, defense work, and anywhere the General Data Protection Regulation applies. The moment inference runs in someone else's cloud, every prompt becomes a data handling event that needs contractual assurance, a security review, and often a signature from someone who does not want to give one. The NVIDIA DGX Spark removes the reason for the no. It is a desktop AI supercomputer that runs models up to 200 billion parameters entirely on the machine, with the option to operate air-gapped, on an isolated network with no path to the outside internet (NVIDIA; 2026).

That is the sandbox a chief information officer can approve this quarter. Not a data-center project, not a procurement cycle for graphics processing unit racks, not a nine-month sovereignty architecture. A box that sits next to a monitor, draws about the power of a bright desk lamp, and keeps every byte of the experiment inside the building.

The catch shows up when the sandbox works.

The pilot that succeeds is the one that gets you in trouble

One DGX Spark handed to a data science team is a controlled experiment. Give the finance group its own once they see the results, then the clinical informatics team, then two research labs, and the controlled experiment becomes a scatter of unmanaged machines that each hold sensitive data and each run whatever model someone downloaded last week. That is the exact profile of shadow information technology, except the shadow now has a petaflop of compute and a copy of your regulated data on local storage.

I have written this sovereignty argument from almost every layer of the stack. Open-weight models that a firm can self-host so matter data never leaves its walls. Inference silicon that runs a capable model at the power of a light bulb. On-premises servers that resolve the governance constraint that blocked deployment, where control is the product and the cost saving is secondary. The DGX Spark is the smallest and most approvable version of that same argument. It is also the one most likely to multiply past the point where anyone is tracking it.

A sandbox nobody governs is not a sandbox. It is an exposure with good intentions.

Fleet management is the difference between a sandbox and an incident

On June 30, Progress Software put a price on closing that gap: $189 per system per year for Progress Chef Enterprise Management for DGX Spark (Progress; 2026). Progress Chef, the configuration management platform Progress acquired in 2020, treats a fleet of these boxes the way it has long treated servers. It sets an approved system state and holds every unit to it, monitors what software and models each machine is running, stages updates through controlled waves rather than all at once, detects when a box drifts out of policy, and enforces who is allowed to touch what, with an audit trail behind every change (Progress; 2026).

For the air-gapped sandbox, that audit trail is the whole point. A regulator does not accept "the data stayed on the box" as a claim. It wants evidence: which machine, running which model, patched to which level, accessed by whom. Fleet management is what converts a physical fact, the data never left, into a documented control a compliance team can sign. Without it, the CIO who stood up the sandbox to satisfy legal has handed audit a harder problem than the cloud pilot they blocked.

Fleet management is what converts a physical fact, the data never left, into a documented control a compliance team can sign.

Progress is one of three vendors NVIDIA named for this job, alongside Perforce Puppet and Canonical Landscape (NVIDIA; 2026). NVIDIA built its own manageability framework for DGX Spark and then pointed enterprises toward outside tools to run governance through consoles they may already operate. For the buyer, that openness is the reassuring part. You are not forced onto a proprietary control plane invented for one hardware vendor. You govern the sandbox with tooling your platform team already knows, which is what lets a security review approve it in the first place.

Price the fleet, not the box

The hardware math is easy and it is not the decision. Four thousand dollars for a box that ends a sovereignty stalemate is trivial against the legal hours one blocked pilot burns. The number that matters is the recurring one. $189 per system is an introductory figure attached to a fleet most organizations expect to grow, and the renewal terms for a thousand governed units are not something anyone has seen yet.

A sandbox is supposed to be where you take a cheap risk to learn something. The risk here is not the experiment. It is signing up for a per-seat governance subscription whose price you cannot see at the scale you are hoping to reach.