You've heard me talk about my DGX Sparks for months.

Six of them. The boxes that run my agent fleet, locally, on my own network.

Well, Nvidia just put that same chip in a laptop.

It's called RTX Spark, and it's the Windows version of the DGX Spark I run my business on.

Here's what's in it.

One petaflop of AI compute.

128 gigabytes of unified memory.

Enough to run a 120-billion-parameter model on the machine itself.

No cloud. No API key. No usage meter.

Preorders opened Wednesday.

Laptops ship October 16. Compact desktops land in November.

Microsoft Surface, Dell, HP, Lenovo, ASUS, Acer, MSI. All of them.

And this isn't a spec bump with "AI" stickered on the lid.

Microsoft rebuilt chunks of Windows for it.

Agents show up in the taskbar.

Meta's Muse arrives natively on Windows.

And the OpenClaw Windows app now sets up an optimised local model in one click on any RTX GPU with 24 gigs or more.

Yes, that OpenClaw.

The harness I run my whole fleet on gets a one-click install on consumer hardware next week.

Hermes is getting the same treatment.

Now think about what this means against everything I've sent you since July.

Fable got frozen by the government. Pro got halved. Flagships went behind hardware keys. Free tiers got ads. Your text got watermarked.

Every one of those stories was about a model that lives on someone else's server, under someone else's rules.

A 120-billion-parameter model running on your own desk answers to one person.

You.

No one can ration it. No one can meter it. No one can switch it off from Washington or San Francisco.

Here's the honest caveat, because I always give you one.

Local models won't match GPT-6 Astra or Opus 5.5 at the very top end.

The frontier stays in the cloud for a while yet.

But for the volume work, the agents, the drafts, the data, the daily grind...

A machine that runs a hundred-billion-parameter brain for free, forever, after you buy it once...

Changes the math on your entire stack.

I bought my Sparks because I refused to rent my business from a lab.

Next week that choice ships in a laptop aisle.

Make it on purpose. 🫡

In 1980, Bill Gates set Microsoft's mission as "a computer on every desk and in every home"...

People laughed. Computers filled rooms and cost more than houses.

Forty-six years later, the same company is putting an AI agent on every desk.

Nobody's laughing this time. 💻

Windows just built the sandbox I've been begging you to build.

The chip is the headline.

The security layer is the story.

Remember Monday's email. Apple tightening disk access because agents were being handed too many keys.

This week Microsoft and Nvidia shipped their answer, and it's more thorough than Apple's.

Microsoft Execution Containers went generally available.

That's an operating-system-level sandbox for every agent on the machine.

Each agent gets its own isolated environment.

Windows now has built-in controls for agent identity, containment, policy, and management.

The OS watches what the agent does, not just the user.

Then Nvidia added a runtime on top called OpenShell.

Here's what it does, and read it slowly, because it's every rule from this newsletter turned into software.

It defines exactly what each agent is allowed to do.

It decides whether a request runs on your local model or goes to the cloud, based on your privacy settings.

And when something does go to the cloud, it masks your personal data in the query first.

Scope the permissions. Keep the sensitive stuff local. Watch the agent.

That's the whole summer in three features.

Nvidia's own framing was blunt.

Agents rarely run on people's main devices because the security tools never existed.

So they built them, and they built them into the operating system instead of leaving it to you.

Here's why that's the real news.

Every rogue-agent story I've sent you traced back to a thin wall somebody called a sandbox.

OpenAI's DNS gap. Google's internet-connected test box. The DNS hole that took two and a half hours to close.

The labs with billions in compute kept building leaky boxes.

Now the box is being poured into Windows itself.

That doesn't make your agents safe by default.

You still decide what each one may touch.

But for the first time, the wall is the machine's job, not a Tuesday afternoon chore you keep putting off.

Buy the hardware if you want.

Use the containers either way. 🫡

Go price a local agent box before next Friday...

And go decide, on purpose, which parts of your stack you'll never rent again. 🔥

Jeff J. Hunter

Founder, AI Persona Method | TheTip.ai

PS. The Surface RTX Spark Dev Box runs the full 120B models on a desk for $5,999... Expensive for a laptop, cheap for a data center. Depends which one you think you're buying. 😏

Get paid to solve problems like this → AI Certified Consultant

“AI Money Group” to Learn to Make Money with AI