The Personal AI Computer Will Move Agents From Cloud Tolls to Local Control

The computer interface has been getting in our way for decades.

We got better screens, faster chips, prettier apps, and more cloud services. The basic bargain stayed the same: you learn the computer’s language, then the computer rewards you with a result. Click here. Drag there. Open this menu. Paste that command. Pay another subscription. Buy another batch of credits. Hope the tab you need is still open.

Well, AI agents are starting to invert that bargain. The computer is learning our language.

This has all sorts of interesting implications, but the biggest one is what it does to the balance of power between cloud AI and local AI.

Which brings us to the big news of the day: Microsoft and NVIDIA’s new partnership on a new machine that feels like the first real step toward changing that equation.

First up, the TL;DR

The future computer may feel less like a tool you operate and more like an employee you manage.

That is the real pitch behind NVIDIA and Microsoft’s RTX Spark, a new Windows PC platform built for personal agents. In simple terms: instead of opening apps, hunting through files, and paying cloud credits for every serious task, your computer gets enough local intelligence to do more work on-device.

Here’s what happened:

• RTX Spark PCs will offer up to 1 petaflop of AI performance and 128GB of unified memory.
• NVIDIA says they can run 120B-parameter models locally, which means large models can work on your machine without sending every task to the cloud.
• Microsoft is adding Windows security tools, while NVIDIA is adding OpenShell, a runtime that limits what agents can access and do.
• RTX Spark laptops and desktops are expected this fall from Microsoft Surface, ASUS, Dell, HP, Lenovo, MSI, and others.

How to try it:

• Developers can follow Microsoft Build for the first Windows agent platform demos.
• Everyone else should watch the fall RTX Spark PC launches.
• The useful test: can it run a local agent that handles real files, real apps, and real mistakes safely?

Why this matters: Cloud AI turned intelligence into a toll booth. Every image, video, code agent, and long task burns credits somewhere.

Local agents change the economics. They make basic computer work feel more like buying a game console once than feeding quarters into an arcade machine forever. The cloud will still matter for top-tier reasoning, giant models, and heavy enterprise tasks. But everyday computer control should become local, private, and built into the machine you already own.

Apple saw this direction early with Apple Intelligence: the assistant belongs close to your personal context. The problem was timing. The models, memory, and agent safety layer were not ready.

Our take: NVIDIA and Microsoft are trying to get ahead of the device shift OpenAI and Anthropic may eventually need to face.

If the agent becomes the interface, the device becomes the distribution channel. The company that owns the trusted local computer owns the place where everyday AI work happens.

WWDC starts June 8. Apple now has one week to show whether it still owns that idea.

The arcade era of AI has an expiration date

Today’s consumer AI economy feels like an arcade. This is a metaphor Dave Morin quoted as making the rounds at Dell World 2026, but it makes perfect sense to us: to play video games, you used to have to go to an arcade, and keep putting in quarters to keep using it. That's basically cloud AI today.

You walk up to the machine (grab your API over the cloud), put in credits, and get a few minutes of magic. Generate a video. Run a coding agent. Make an image. Analyze a long document. The game is incredible, but the meter is always running. In this case, the cloud datacenters are gigantic arcade machines.

That use-case made sense for the first AI wave, just like it made sense for the first wave of video-games. The models were too large to run locally, the hardware was too weak to support it and required massive server racks that you couldn't afford, and the safety layer was barely there. Cloud subscriptions were the fastest way to get intelligence into people’s hands. But now that the bill has come due, parents everywhere (CFOs, in this case) are asking the all-important question about bankrolling their childrens' (engineers and tokenmaxxers, in this case) seemingly endless appetite for arcade games: must we?

That's why the next phase looks different. NVIDIA says RTX Spark can run 120B-parameter models locally with up to 1M tokens of context. Microsoft says Windows on RTX Spark will support local agent workloads with OS-enforced identity, containment, and manageability. NVIDIA’s OpenShell adds another control layer, including policy rules and routing between local and cloud models.

That last part matters more than the chip specs.

A local agent needs three things before normal people trust it:

• Enough intelligence to handle everyday computer work.
• Enough memory and local access to understand your files, apps, and context.
• Enough containment that it can act without becoming a tiny raccoon in your operating system, scarfing up your garbage and leaving the lid open on your trash can for anyone else to sniff through. Weird metaphor, but you get what I mean: raccoons are trouble.

Make no mistake: Cloud models will keep winning the frontier. No one involved in this project is saying anything otherwise. If you need the smartest possible reasoning, giant context windows, enterprise-grade research, or massive video generation, the cloud still makes sense.

However, this is something that OpenAI and Anthropic aren't yet ready to admit: local agents will win on convenience. They can search your files, organize your desktop, summarize your PDFs, prep a deck, check a folder, and run personal workflows without turning every click into a billable event.

That is the home console moment equivalent. Arcades did not disappear when consoles arrived. They just stopped being the only way to play.

I guess what I'm saying is Microsoft and NVIDIA are trying to make the NES, but for AI.

Apple was early, but the stack was not ready

This is also much-needed vindication for Apple. Apple’s core AI instinct has always been device-first.

The company has spent years optimizing for privacy, on-device processing, hardware-software integration, and personal context. That is exactly the right worldview for agents. A useful assistant needs to know your calendar, files, contacts, location, apps, habits, and preferences. That context already lives on the device.

The issue was readiness. Apple Intelligence launched before the agent layer was mature enough to make the full vision obvious. Siri still carried years of user disappointment. Local models were useful for narrow tasks, but the broader public expected ChatGPT-level magic. We stand by every joke we made at Apple's expense over the past two years... they're a ~$4+ trillion market cap company, they can handle it.

In this case, Apple were something they never like to be: early. But over the long view, they were right... and by our estimations, Phase 2 of their AI excursion is about to arrive right on time.

WWDC26 runs June 8-12, with the keynote on June 8. Apple now has a chance to show whether Apple Intelligence is turning into a real interface layer, or staying a collection of helpful features. We imagine the RTX Spark will light a real fire under them to do so.

Interestingly, NVIDIA and Microsoft are making the opposite move from Apple’s usual playbook. Instead of starting with the integrated consumer experience, they are starting with hardware horsepower, developer workflows, and Windows as an agent platform.

That gives them a different advantage. Windows still owns a huge amount of work. Developers, creators, analysts, accountants, marketers, designers, and operations teams already use it as their daily command center.

If Windows becomes truly local agent-native, Microsoft does not need to invent the next consumer gadget first. It can make the PC feel less like software you use and more like infrastructure you command... without all the bad headlines they've received trying to do so thus far.

Get this: the agent economy starts with easier computers

The fun demos are already pointing there.

Clicky sits next to your cursor, sees your screen, listens when you talk, and can spin up a background agent. OpenClicky shows how quickly the open-source version of that idea can appear. Bryce Rattner Keithley’s How I AI workflow shows a non-technical recruiter shipping an iPhone app by using Claude as planner, Claude Code as engineer, and Terminal as executor.

Those examples share one deeper pattern: the computer is (finally) becoming easier to use.

This is happening at two layers:

• At the interface layer, voice, screenshots, pointing, and visual context make computers easier to navigate.
• At the creation layer, plain language, code agents, and visual debugging make software easier to build.
• Put those together and the pain of the modern computer starts fading into the background.
• • You stop remembering which menu hides the setting.
  • You stop translating your goal into the exact app sequence.
  • You ask, inspect, approve, and redirect... and actually get what you need done.

That shift, once fully embodied, is bigger than any so called "productivity."

To quote Jessica Lessin, also of the More or Less podcast (and CEO of The Information): the average person has a limit to how "productive" they actually want to be.

But you know what everyone wants to do more? Use the computer less (yes, even knowledge workers!).

Once again, I borrow an idea from Dave Morin that hit me with the truth equivalent weight of a ton of bricks: today's computers kind of suck. The reason we need to spend 8 hours a day dealing with web and application interfaces is not just because we have a 9-to-5 (which also needs a 2026 / AGI-pilled re-think), but because these machines are not as intuitive or simple to use as we'd like them to be.

Now imagine if they were. Imagine if your computer was as easy to manipulate as waving a magic wand. Imagine if you could not only click and type, but you could speak and point and your computer could register both semantically what you want and visually what you're pointing at, and it could actually do what you intend for it to do. That is, it doesn't just react to the button you click, but you give it a list of instructions and point at this interface or that interface or swipe a few gestures on your computer, and the computer actually understood all of those things together and could reinterpret it as a comprehensive complete instruction (as you intended) and it could then go act on it for you.

If suddenly our computers could understand us truly like other humans do, and the interfaces we used were as malleable as physical objects we interact with in real life (but even BETTER and more like magic to use), what else could we spend our time doing? Suddenly the agentic gig economy becomes actually real.

Whoa whoa whoa, what is the agentic gig economy? It's what happens when you combine skills, agents, and local AI into agents-as-a-service. I'm talking about the gig economy transforming into the agentic services economy, where suddenly anyone can offer services powered by agents (not necessarily always or only involving agents, as you'll soon read), with the digital infrastructure essentially solved for them.

Okay, let's unpack that. Roll back to the 2010's. The gig economy gave us individuals apps that found work for us. Uber found riders. DoorDash found orders. Fiverr and Upwork found clients. Etsy and Patreon found patrons. OnlyFans found... "fans", lol. Of course, in this model the platform owned demand, payments, trust, and rules. The worker brought labor. And this was a fair trade.

Arguably, it's the trade any corporation who offers a job to any individual contributor makes. They offer the worker the trust of their name and capital, payment, demand for more work, and rules on how often they get paid, what they have to do to get paid, etc etc.

But how does the addition of agents change this dynamic? For starters, the obvious: labor is no longer just an input of the individual contributor. The platform (be it Uber, Fiverr, Upwork, or OpenAI and Anthropic) can also offer labor. Not human equivalent labor (or at least, not yet), but there are now "agents" that can be hired instead of humans. Suddenly, laborers can now hire cheaper labor to help sell their services. They don't necessarily need a platform anymore to aggregate their demand. They can hire agents, kinda like human talent agents, to represent them in the global marketplace.

So play that out: the agent economy gives individuals agents that can represent them. Maybe not yet, but play this out five years down the line. These agents can search for jobs, pitch services, negotiate scope, book calendar time, prep materials, produce deliverables, and coordinate with other agents.

A few examples:

1. A freelance video editor’s agent could find brands launching campaigns, pitch a short-form package, draft the contract, pull source assets, schedule revisions, and invoice automatically. Play this out and these brands will also have agents pitching other agents for qualified candidates, streamlining this process.
2. A local contractor’s agent could find neighborhood repair jobs, bid based on availability and materials, order parts, schedule the visit, and update the customer while the contractor is on-site. Arguably, this is possible now via Yelp + Nextdoor + Facebook Marketplace arbitrage, though you have to be pretty tech-savvy to pull off the orchestration across these platforms.
3. A consultant’s agent could monitor RFPs, draft proposals, assemble case studies, book discovery calls, and route research to specialist sub-agents. If you're a consultant, you mean to tell me you aren't already using AI to do this for you now?

Those are just a few examples, and they're why the Local Agent PC matters. In order for regular people to be able to do this at scale, they need to trust their AI, and be able to use it as much as they can. This type of AI needs to live close to the worker’s files, calendar, voice, tools, and past work. The best agent for your service business probably starts as a local computer-native layer, then reaches into cloud models when it needs a bigger brain.

It might be hard to justify paying API prices for AI to start, but if you had a local PC agent you could trust that can run 24/7 just for the cost of your home electricity, you can justify working for yourself and getting your first payback ASAP. It's like having Uber or Doordash ready to book you your first gig in as little time as it takes to download the app, except instead of buying a car, you're buying a desktop. In both instances, you will also likely be able to rent instead to get up and running even faster.

Of course, some people will equate "renting" computer to what you can do now, and it's true. But how many normies do you know who know how to set up a cloud system of any kind outside of iCloud or Google Drive, let alone rent server capacity from Fireworks or HuggingFace?

Natively using your own computer is much more natural for people who grew up with computers, just like using a smart phone is much more natural for those who grew up with those. For simplicity, i'm going to limit this discussion to local desktop agents vs smartphone agents, but you could apply the same logic, just scaled down slightly in scope.

Also worth noting: this is related to what we wrote about why OpenAI (or anyone, really) needs to create a "business-in-a-box" app before they accidentally destroy the economy by making millions of jobs automatable. Since then, the major AI companies have decided to get their messaging under control and explain that "actually, the AI revolution will not kill all the jobs, so put your pitch-forks away, people." Of course, this could be a pump-fake, and

The missing layer: physical infrastructure for individual operators

My point here is that digital agents can replace a surprising amount of coordination. If suddenly it's feasible for individual service providers to run their own business providing their service without the need to hire a small team or take out a huge loan to get to work, it enables a lot more people to work for themselves. And as big companies reduce headcounts to remain agile and compete with smaller, agent-native upstarts, there will be more individual service providers available on the market than before. Sure, new platforms could and will likely pop up to accommodate this demand in talent (you could also argue having a central platform to coordinate service providers, like Uber, could result in more people working more of the time, but that's another matter), but this time the human service providers aren't just competing against other humans, they're competing against agent service providers as well.

This leads me to believe that human service providers will need to use a combination of custom skills they create from their own unique experience coupled with local (or on-prem) agents working on their behalf as well as their physical, embodied self in their specific geographic location in order to stay competitive and secure them work that only they can provide for their customers.

So let's talk about the physical component for a second there.

Agents can do many digital tasks. The list will increase as we change the frame. But they cannot yet manufacture a part, drive a van, install a heat pump, cut steel, run a lab bench, or hold inventory by themselves.

That is where the next company wave of the agentic gig economy gets interesting.

The 2010s gave us “Uber for X,” where startups built human-to-human digital infrastructure for services. The platform matched demand and supply. The human still did the work.

The agent era needs a new stack: agent-to-agent infrastructure plus just-in-time physical capacity.

Think of it as the franchise kit for one-person companies.

I have a curse, and it's that I'm constantly thinking of the common person as my north star. But as they say, if you make live better for the every man, every man benefits, right? So picture a laid-off operations manager, designer, mechanic, engineer, nurse, construction specialist, or factory technician. All of these people may have the knowledge to sell a service themselves. But do they have the similar requisite expertise to run a business? Probably not.

Well, agents can theoretically supply much of the missing digital labor they need to do. They can handle prospecting, paperwork, customer support, scheduling, compliance reminders, and basic analysis. This prediction is for five years in the future, mind you, so even if agents today can't do those tasks perfectly, in five years they better damn well be able to!

What's left for the human service provider to start offering their own services to customers and compete with the firms that laid them off? Well, the remaining gap is access to physical assets. There's also trust, but my theory here is that if you solve the access to physical assets problem, trust can increase.

That creates room for new categories of businesses, IMO:

• Agent-native marketplaces: places where personal agents can discover jobs, bid, verify credentials, negotiate scope, and book work on behalf of humans.
• Physical capacity clouds: imagine "the cloud" but for physical goods. Shared access to workshops, tools, vehicles, robots, lab equipment, fulfillment, fabrication, and testing. So called "Cloud labs" like Emerald are a high-tech, first demonstration of this. Fully autonomous factories are next. What will be after that?
• Contract manufacturing for tiny firms: these are factories and specialist shops that accept agent-readable job specs, quote quickly, and run small batches. These businesses are already coming back in fashion with the US self-sufficiency manufacturing boom.
• Robotics-as-a-service: robots and physical automation rented by task, hour, job, or location. We're not ready yet, but you can already see this start to play out.

The idea here is to turn any physical infrastructure or physical good into an output that can be procured on demand. You could argue this trend has been happening for a long time now. I guess my point is that this cloud for on-demand physical infrastructure needs to be hyper-local, just like the computer.

Once you have all the physical infrastructure in place so anyone can get up and running offering an agent-supported service in their area of expertise, the last hurdle remaining is trust and logistics. And both of those items could be handled in an agent-friendly way, too:

• Trust and insurance rails: identity, licensing, guarantees, bonds, liability coverage, and dispute resolution needs to get built for agent-mediated work.
• Local service operating systems: someone needs to handle scheduling, inventory, route planning, customer updates, quoting, and payments for small operators.

The old gig economy assumed the worker needed help finding a customer. The agent economy assumes the worker needs help becoming a miniature company.

If large corporations shrink their human teams while keeping all the physical assets, then the displaced talent needs a way to compete without recreating the entire corporation. Arguably, that's not needed in an agentic-first economy anyway. Agents can supply part of the missing organization. Shared physical infrastructure and shared trust networks supplies the rest.

I'm trying to formalize this into a physically tangible example. This one is a bit extreme:

Imagine a former auto manufacturing engineer who may not need to own their own auto plant to make a car that competes with Ford. This is probably the most extreme example I can think of, and I'm sure everyone will tell me why it's unlikely. But try to think of what they would need in this scenario. They would need agent-mediated access to design software (which they can get today as agents can easily use CAD now), prototyping shops, robotics cells, certification workflows, distribution partners, and customers.

Is this easy to secure today without capital? No. So what would have to change to make it easy for the highly-skilled individual to do this on their own? The physical infrastructure available on demand.

Okay fine, here's another, perhaps more credible, near term example: a former enterprise marketer may not need a big agency to work for anymore. But they will probably need agents for research, creative production, media buying, reporting, and a brand system like Bloom that keeps every output consistent (no free ads; just saw them today, looks sick).

A robotics researcher needs open models like MolmoAct 2 and NVIDIA Cosmos 3, plus access to robot labs, sensor datasets, and simulation tools.

I guess what I'm trying to say here is that the prize of the agent economy should not be “everyone becomes a gig worker.” The prize is that more people can become small service operators with leverage in their local area. The digital infrastructure required to do that has never been easier to acquire. So the next wave will be filling in the physical infrastructure gaps required as more people are laid off and find it necessary to pursue entrepreneurship to compete with the former firms that laid them off.

The counterpoint: local agents will be messy before they are magical

The local-agent future has three obvious problems.

First, hardware will be expensive at the beginning. RTX Spark sounds powerful, but the first wave of local-friendly devices has skewed toward premium laptops, creator workstations, and developer machines. Regular people will not instantly get local frontier-ish agents for cheap. It will still require an upfront investment.

Second, local intelligence will lag the best cloud models, maybe always. Cloud providers can throw more GPUs, memory, and specialized systems at the problem. Local agents will need to be smaller, more efficient, and more specialized. This is not a bad thing. The idea is to bring the intelligence as close to where you do the work as possible over time.

Third, computer control is risky. An agent that can click, type, browse, email, code, and move files needs serious containment. Microsoft’s security primitives and NVIDIA’s OpenShell are signs the platform companies understand this. Will the security controls over these agents get much better before these tools are rolled out en-masse to the public? Remains to be seen.

This is also where the device race gets sharper.

If agents become central to computer use, OpenAI and Anthropic may eventually need their own hardware or deep operating-system partnerships. Owning the model is powerful. Owning the interface where the model acts may be more powerful. We've argued OpenAI and Anthropic will need to become computer companies long term... we see them as this generation's Microsoft and Apple, TBH. Wanna guess which is which?

However, Apple already owns the interface for millions of people. Microsoft owns the work PC. NVIDIA owns much of the AI compute stack. And Google owns Android, Pixel, and their own low-cost devices. OpenAI and Anthropic own the relationship with users inside the chat window, but can they convert this to a hardware company? I would argue they must. Unless they want to get into the far more expensive datacenter business. But that's hardware by another name anyway.

What to watch next

I got way ahead of myself here. Will RTX Spark ship as another premium curiosity, or could it become a real developer platform? It depends on whether or not agents like Hermes, OpenClaw, ComfyUI, and local model runtimes feel meaningfully better on these machines.

Next is whether Apple shows a more agentic Apple Intelligence at WWDC. Apple does not need to win the benchmark fight to win the interface fight. It needs to make the assistant feel native, private, and useful across the system.

Finally, we wonder whether the cost curve will keep moving. MiniMax M3, LocateAnything, Cosmos 3, MolmoAct 2, and other open releases all point in the same direction: strong models are getting cheaper, more specialized, and more available. What does this do to "closed" cloud models long term? Maybe the right play all along was being a neo-cloud that sells access to datacenters, not models themselves.

But the reason I started thinking about the agentic services economy is because I worry a lot about what happens to normal people as this technology disperses.

For companies optimizing for shareholder value, the natural thing to do is to reduce costs (in this case, headcount) and increase prices (in this case, charge more quarters per token; also can we acknowledge that unit of AI economics is "tokens"? That's what Chuck-E-Cheese charges for their arcade games lol).

For governments and those in charge of coordinating humans, obviously having big companies lay off more people is not an ideal solution. But it's the optimal move when their fiduciary duty is to optimize for shareholder value (profit). So you can either write laws that protect human jobs, or support initiatives that give smaller upstarts more leverage with which to compete. And for the record, big companies know this too. They want to reduce expenses, not customers.

Which is why I look to the future and ask: what can an individual contributor do when agents become normal, other than be washed away by them?

Just like it's now possible for many more millions of people to write code thanks to AI, it will soon be possible for millions more people to run and operate local services businesses, either selling their own custom agents-as-a-service, or more likely (because people want to work with other people), using agents to support their own business where they utilized their own skills as a service provider.

I'll end with this vision: Imagine you are that service provider. You have a set of skills that you've spent your whole career learning how to do, or you've always wanted to develop, and finding yourself unemployed, you decide to finally dedicate to offering these services to others under your own business.

Because of local AI, you can now use your own custom-tailored agents to help support your business. Your agent finds your first client. Another agent negotiates the terms. A local model handles the private files, keeping your data (and theirs) secure. Some part of the job requires processing tons of data and compute. So a cloud model handles the heavy reasoning. Then it's time for you to go on site. A physical capacity network provides tools, fabrication, robots, vehicles, fulfillment; whatever you need to get the job done. You bring your judgment, taste, previous experience, skills, relationships, craft, and human accountability.

This is the job you were born to do.

Thanks to local AI, you can support yourself financially to be able to do it.

That is the agent economy’s most interesting future-outcome.

It gives everyday people leverage without forcing every worker through the same middleman app or corporate ladder that seemingly no longer wants to exist.

The personal AI computer is the first piece of that stack.

The rest is still waiting to be built.