Trend
Local AI agents on Windows: what changes if your data cannot leave the building
NVIDIA and Microsoft have opened a vendor-supported path for AI agents to run on the Windows machine instead of a cloud API, but the privacy and speed claims attach to new RTX Spark systems, not to your existing office PCs. The useful move this month is an audit, not a purchase order.
What the two companies actually announced
NVIDIA says Jensen Huang and Satya Nadella used a Microsoft event in San Francisco to lay out hardware and software co-engineered so AI agents run on the Windows PC itself (NVIDIA).
Three pieces matter to an operator.
Microsoft made Microsoft Execution Containers generally available, described as OS-level infrastructure that lets agents run safely and persistently in the background under operating system control (NVIDIA).
NVIDIA's RTX Spark puts its AI stack into Windows laptops and compact desktops: a Blackwell RTX GPU with up to 6,144 cores and an up to 20-core Grace CPU connected at 600 GB/s, with availability stated for October 16.
Microsoft's Windows blog puts that at 1 petaflop of AI performance and up to 128GB of unified memory, and says it is optimising Windows for building and running agents with OS-enforced identity, containment and manageability (Microsoft).
Local does not mean offline
Most coverage reads this as AI without the internet. It is not that. It is a managed agent stack with operating system controls over identity and containment, which is a different and more useful thing: the OS decides what a long-running agent is allowed to touch.
The second thing coverage skips is hardware specificity. The privacy story and the performance story both sit on the new systems. Nothing here improves the eight-year-old PC in your estimating room on its own.
What this means for a 10 to 60 person business
If a workflow touches drawings, cost structure or customer records you have promised to keep in-house, there is now a vendor-supported Windows path to run an agent on that data without it crossing a cloud API, on new hardware, for the workflows you can actually name.
That last clause is the work. Most shops I talk to cannot name them.
The audit table, finish it this week
| Workload | What it touches | Cloud API acceptable | Where it should run |
|---|---|---|---|
| Quote follow-up email drafts | Names, quote totals | Usually yes | Cloud |
| RFQ drawing read and extract | Customer drawings under NDA | No | Local |
| Price lookup for a proposal | Your cost structure | No | Local |
| Website chat triage | Public product info | Yes | Cloud |
| Supplier invoice matching | Supplier terms | Depends on the contract | Read the contract |
Budget about 60 minutes with whoever owns the files, and about 30 minutes more to count how many of your current automations sit in each row.
Arithmetic on your own numbers
Call A the number of automations you run and L the number that land in a Local row. If A is 12 and L is 2, you are considering a hardware purchase for two workflows, and the other ten keep running on a cloud API at no capital cost. If L is 9, the hardware case is real. Run that division before you read another spec sheet.
For a sense of what sits in the Local column: I built a datasheet pipeline for a lighting manufacturer covering 6 product families and 32 dimension sheets with a daily catalogue refresh. Product drawings and dimension data are exactly the kind of input a company tells you may not leave its network.
Limitations
- I have not tested RTX Spark or Microsoft Execution Containers. I own neither.
- I could not verify that these local-agent features reach ordinary Windows 11 business PCs beyond the announced RTX Spark and partner systems.
- I could not verify that a manufacturer can keep all drawings, pricing and customer data off-cloud using only the announced stack, with no external model or telemetry dependency.
- Neither company framed this as a shop-floor data story. That is my reading of who it helps, not their claim.
My take
This is the first on-device agent story I would repeat to a manufacturer, because the containment is at the OS layer rather than in a vendor's promise. It is also not urgent. My guess is that the interesting hardware for a 20-person shop is the second generation, not the October systems, so the thing worth doing now is the audit. Other notes on running this kind of pipeline are on the blog.
Questions people ask
Does this mean I can run AI with no internet at all?
Not necessarily. The announcements describe a managed Windows agent stack with OS-enforced identity and containment, not a guarantee of a fully offline system. I could not verify that every dependency, model and telemetry path can be kept inside your network using only the announced stack.
Do I need to buy new hardware to get this?
The strong claims are tied to the new RTX Spark systems, stated at 1 petaflop of AI performance and up to 128GB of unified memory, with availability stated for October 16. Whether these features reach ordinary Windows 11 business PCs is something I could not verify.
What should I do this month if I am not buying hardware?
Write the audit table. List every automation, mark the ones touching drawings under NDA, your cost structure or records you promised to keep in-house, and divide that count by your total. Budget about 90 minutes. If only two of twelve workflows are local-only, the purchase can wait.
Have a process like this leaking time or leads?Thirty minutes, no deck, and you leave with a plan.
Book a call