The Intelligence Estate Gets a Street Address

A computer the size of a hardcover book can now run a 200 billion parameter AI model on your credenza, for about the cost of a trade show booth. For a company serious about owning its intelligence, the estate just became real property.

For roughly $4,000, your company can now own a computer that runs artificial intelligence at a scale that required a data center two years ago, on your premises, on your data, under your rules. That is the plain fact behind three hardware announcements most executives scrolled past. It changes the opening move for any company between $100 million and $3 billion in revenue that has been waiting for a practical way to own AI capability instead of renting it.

In April I wrote Stop Renting Intelligence, arguing that most companies are treating AI like a hotel room, and that the alternative is to build an Intelligence Estate, the way George Vanderbilt built Biltmore. It was a fair argument with a fair objection: Vanderbilt had to clear the land and quarry the stone. Where does a midsize company even break ground?

This summer, the answer went on sale.

What just happened in local AI hardware?

In the span of sixteen months, three of the largest computer makers in the world shipped desktop machines built specifically to run large AI models locally, and priced them like office equipment rather than infrastructure. Apple got there quietly first. The Mac Studio that shipped in March 2025 can be configured with up to 512GB of unified memory, enough to hold AI models of over 600 billion parameters entirely in memory.¹ NVIDIA named the category. Its DGX Spark, unveiled as Project DIGITS and shipping since last October, is a "personal AI supercomputer" the size of a hardcover book: 128GB of unified memory, models up to 200 billion parameters, $3,999.² Then in June, AMD confirmed the market. Its Ryzen AI Halo developer platform matches the Spark's price and memory, opened for pre-orders immediately, and will refresh this fall to 192GB of memory addressing models of roughly 300 billion parameters.³

One product is a stunt. Two is a race. Three is a market.

Notice the language AMD chose for its announcement. Not workstations. Not mini PCs. "Agent computers." The vendors themselves are telling you what these machines are for, and it is not faster spreadsheets.

One more detail worth an executive's attention. This spring, Apple quietly withdrew the 512GB memory option from the Mac Studio because AI data center demand is soaking up the world's memory supply.⁴ Read that as a signal. The window where this class of hardware is cheap and available is open right now. Windows also move.

The Intelligence Estate was always a metaphor. This summer it acquired a price tag, a spec sheet, and a delivery date.

Why does economical hardware matter for a midsize company?

Economical hardware completes a convergence: for the first time, the hardware, the models, and the platform layer needed to own real AI capability are all affordable at the same moment, and the entry decision fits inside a quarter's discretionary budget rather than a capital program.

The hardware curve is the story above. A serious machine now costs less than the audiovisual budget for your annual sales kickoff.

The model curve crossed at the same time. In my last piece, Liberty Ships Built in 4 Days, I pointed to the analysis converging on open weight models like GLM-5.2: the models you can download and run on your own hardware now excel at the center of the distribution, the everyday, high-volume work of content generation, document analysis, coding, and data synthesis. The premium frontier models keep their edge at the exotic margins. The center is where most of your company's daily work actually lives.

The third curve is the platform layer, and it is the one executives hear about least. Open standards now let AI systems connect to the software you already run. Durable memory frameworks give AI structured recall instead of goldfish sessions. Local runtimes serve models reliably on your own machines. The pieces to encode your processes exist. Two years ago, assembling all of this took a research team. Today it is an engineering exercise, and the platform that assembles it into one owned, governed system now exists and has a name. I will come to it.

Here is why this convergence lands hardest on a company in the $100 million to $3 billion range. You have proprietary data worth protecting, processes distinctive enough to be worth encoding, and compliance exposure that should make you think twice about shipping every contract, formula, and customer record to a third party's servers. What you do not have is the enterprise excuse. This does not require a program office, a data center, or a nine-figure commitment. It requires one executive decision and a place to plug in the box.

Isn't a desktop AI box too slow to compete with the cloud?

The tokens-per-second critique measures the wrong thing, because these machines are built to produce finished outcomes, and an outcome is worth more than a fast chat reply. The tech press has spent months benchmarking these boxes on single-conversation speed and finding them slower than the cloud.⁵ True, and beside the point. The same DGX Spark that looks modest in a one-on-one chat delivers roughly 695 tokens per second when running 256 jobs in parallel.⁶ These are plow horses. They are built to work all day and all night on many things at once, and that is exactly the work you need done.

Measure the machine by what it produces instead of how fast it types:

Memory is stored intentionally. What your company learns gets kept, organized, and compounded, instead of scattered across a vendor's session logs.

Policies, rules, and preferences are understood and applied. Your pricing floors, your approval thresholds, your tone with customers, enforced consistently.

Access to information is respected by role. The system knows what the controller may see and the sales rep may not.

Skills are built for your desired process. The work happens your way, to your standard, every time.

Inconsistencies and issues are surfaced. The contract term that contradicts the master agreement, the invoice that does not match the PO, found before they become problems.

Overnight processing plans the next day. Your team arrives to prioritized exceptions and prepared drafts instead of a cold queue.

Not one of those outcomes is a chat reply. Every one of them compounds.

The spec sheets leave one thing out. These machines are sold as developer platforms, and the label is accurate. Making one productive is a project. That is precisely the argument for moving now. Every prior technology cycle rewarded the companies that built while a capability was still a project and punished the ones that waited for the appliance. By the time Sovereign AI in a Box arrives, professionally packaged and safely late, the builders will have years of compounding advantage.

What turns a box into an Intelligence Estate?

A box is not an estate. A model running alone on a desktop is impressive and idle. Turning it into a working capability takes a platform: software that connects the intelligence to your systems, gives it durable memory, holds it to your rules, and puts it to work on your schedule.

That platform exists now, and it has a name. It is called SOVR, from a company called Autarko. SOVR keeps ownership and control of a company's AI where it belongs, with the company. The intelligence your people actually talk to, the one that remembers your business and works through your night, is named Hewston. SOVR provides the control. Hewston does the work. Everything either of them learns belongs to you.

Six elements combine to provide sovereign control.

Integration connects the intelligence to the systems you already run, your ERP, your CRM, your document stores, so nobody has to learn a new interface. The work meets your business where it already lives.

Context gives it a memory that is yours. The intelligence accumulates durable, structured knowledge of your business the way a twenty-year employee does, and that memory is held on your hardware and owned by you.

You create the Trust layer. Every action runs under rules you wrote, with information visible by role and a record you control, so you can always answer what was done, on whose authority, and why. The controller sees what the controller may see. The sales rep does not.

Governance is part of the core architecture, not bolted on at the end. When actions are taken on your behalf, someone owns the consequence, and SOVR is built so that someone is you, on your terms and inside your walls, rather than a third party whose interests are not yours.

Skills encode your way of working. The month-end close checklist, the quote approval flow, the way your best estimator scopes a job, captured as repeatable capability rather than tribal knowledge.

Tools let it act. They serve the models, run the overnight work, and tell you in the morning what got done while you slept.

I have written about this architecture before as The Sovereign AI Operating System, and the elements will be familiar to readers of It Has to Sound Amazing and Earn a Profit as the sections of the Intelligence Orchestra. The idea has been on the record for months. What changed is that it now fits on a bookshelf, and the platform to run it is real and named. Proxigee Services, based in Pittsburgh, helps midsize companies stand up exactly this, the hardware, the SOVR platform, and the internal capability to grow it.

What should you move onto your own hardware first?

Own the center, rent the tails. Bring home the high-volume, proven, data-sensitive work first: contract and document review against your own playbook, order and invoice exception handling, correspondence drafted to your standards, the weekly reporting nobody wants to build by hand, institutional knowledge made queryable. This is the work that touches your most sensitive data, which is exactly why it belongs on hardware you own.

The tails stay rented. Novel analysis, the hardest reasoning problems, the once-a-quarter question that genuinely needs the most capable model on earth. Your cloud AI bill becomes small, targeted, and defensible, and your negotiating position at every renewal improves.

What does independence actually look like?

Independence looks like this: your box, your data, your rules, your standards, your skills, your tools, your results. That is what SOVR was built to deliver, and Hewston is the intelligence that delivers it.

This country just marked 250 years since a group of businessmen, farmers, and lawyers decided dependence had become intolerable and signed their names to say so. A quieter declaration is available to any company willing to make it. This one gets signed with a purchase order.