Hi,

This is Plain Strata, the Tuesday Pulse.

Software has always needed a landlord. Any company that wants a frontier AI model has had to rent one, because building your own means months of computation inside a datacenter only a handful of labs can afford.

This week the terms of that rental got a lot more visible, and so did the alternative. The same day the US government finally let a gated frontier model go on public sale, a training company banked 130 million dollars on the opposite bet: stop renting, train your own.

Here is what one day of news tells you about who actually controls the smartest software in the world.

Listen:

Spotify: https://open.spotify.com/episode/07kgqVQqG6OmTluniqZhl7?si=RsQ0PpOkReCBjpY3hcVAPw

Apple Podcasts: https://podcasts.apple.com/kg/podcast/plain-strata/id6783455764?i=1000776787005

YouTube: https://youtu.be/ThF1YDlXjLY

The full piece, no need to click through:

One Wednesday this July, two announcements landed a few hours apart, and together they told one story.

In Washington, on July 8, the Commerce Department signed off on something no American AI lab had ever waited for before: government permission to sell its own product. GPT-5.6, OpenAI's most capable model family, had spent two weeks available only to a short list of roughly twenty companies, each name individually approved by the US government. That gate is now open. The broad rollout began the next day.

In San Francisco, on the same July 8, a company called Prime Intellect announced it had raised 130 million dollars, led by Radical Ventures, with NVIDIA, Intel, and Dell all putting money in, at a valuation reported around one billion dollars. Prime Intellect made its name trying to train large AI models across the open internet, on hardware scattered around the world, instead of inside one datacenter. The new money is not for that. It is for helping any company train its own AI, so that it never has to stand at a gate like the one that just opened.

That is the spine of this week: the most powerful AI in the world just became something you wait for permission to rent, and on the very same day, 130 million dollars bet that you will train your own instead.

Two terms, defined in one breath each.

A frontier model is the most capable AI system available at a given moment. The frontier moves: GPT-5.6 is on it today, something else will be next year. Training one takes months of computation in a single building full of specialized chips, which is why only a handful of labs can do it.

And there are two very different ways to train an AI. The first, called pretraining, is how a frontier model is born: the model reads essentially everything, trillions of words, and learns the shape of language and knowledge. Think of it as putting someone through every library on earth. The second, called reinforcement learning, comes after: the model practices a specific job, over and over, and gets scored on each attempt, so the behaviors that score well get stronger. The word is borrowed from psychology, where to reinforce a behavior, from the Latin fortis, meaning strong, is to strengthen it with reward. Practice with a scorekeeper, not reading with a library card.

Hold that distinction. The whole week turns on it.

First, Washington. GPT-5.6 launched on June 26 under an arrangement with no precedent in American AI: the government reviewed and approved, case by case, which companies could use it. We covered that arrangement two weeks ago as the permission slip, and the question we left open was whether it would formalize into a standing process.

This week answered part of that. The Commerce Department's Center for AI Standards and Innovation ran additional technical testing, OpenAI sent its engineers to Washington to answer questions, and on July 8 the government cleared a broad release. All three tiers of the family, Sol, the flagship, Terra, the mid-priced workhorse, and Luna, the fast cheap one, are now rolling out to everyone.

So what actually happened here? On the surface, a gate opened, and that is good news if you were waiting behind it. But look at what the episode establishes. For the first time, a leading American lab released a frontier model on the government's schedule rather than its own. OpenAI said openly that it does not want this process to become the default, and it is not hard to see why: a government that can schedule a launch can also stop one, a power it has already used elsewhere this year when it ordered another lab's models switched off for users abroad. The review framework this all sits inside was only established on June 2. Five weeks later it has already produced a state-approved customer list and a state-timed release.

The gate opened this time. The gate also now exists, and every company building on a frontier model just watched it operate.

Now San Francisco. To understand why Prime Intellect's announcement matters, you need to know who they are, because the company itself is the story.

Prime Intellect spent the last two years proving, step by step, how far you can push training AI across the open internet. In 2024 they trained a 10 billion parameter model on hardware spread across three continents, a first. In 2025 they pushed to 32 billion. But late last year, when they built their largest model yet, 106 billion parameters, they trained it the old way, in one centralized cluster. The reason is physics: at frontier scale, the machines have to exchange so much data so often that an ordinary internet connection becomes the bottleneck, and the whole thing only works if every machine is in the same room, wired together. The frontier stays in one room, and the company that had the strongest incentive in the world to prove otherwise proved it instead.

This week's 130 million dollars is what they decided to do about it. Instead of fighting that wall, they are going around it. The pitch, in their own words: pretraining concentrated frontier AI in a handful of labs, but reinforcement learning breaks that open. A company does not need to pretrain its own frontier model, that part stays behind the wall. What it can do is take a good open model of moderate size and use reinforcement learning to make it excellent at the one job that company actually needs done, trained on its own data, its own workflows, its own product.

And here the abstract pitch has one concrete case worth slowing down for. Ramp, a finance company, used Prime Intellect's platform to train a 35 billion parameter model for a single task: finding answers inside spreadsheets. By Ramp's own account, that small specialized model beat Claude Opus, a frontier model many times its size, on accuracy at that task, while running faster and far cheaper than even the frontier lab's lightweight model. One honest caution before anyone gets carried away: that is a company reporting its own benchmark, and this show's standing rule is that a claim becomes evidence when someone outside the company reproduces it. Nobody has yet. But the shape of the claim is familiar. Regular listeners will remember a 19 megabyte model on a decentralized network beating the giants at one narrow vision task back in June. The pattern is the same: at a fixed, well-defined job, a small model trained hard on that job can beat a giant model trained on everything. What is new is that this pattern now has a price tag and a sales team. Prime Intellect says over six thousand customers use its stack, and that demand has passed 100 million dollars in annualized revenue in under a year. Whatever you think of the branding, that is not vapor.

Why does this belong on a show about decentralized AI, if the training happens on rented clusters rather than across the open internet? Because decentralization was never only about where the computers sit. It is about where the control sits. A world where one lab trains the one model everyone rents, behind a gate a government can operate, is centralized in the way that matters. A world where thousands of companies own their models, trained on their own workflows, is not, even if every one of those models was trained in a perfectly ordinary datacenter.

Name the shape, because it is much older than AI. It is the rent-or-own decision, and you have made it yourself: a home you rent is cheaper to start and someone else fixes the roof, but the landlord sets the terms and can change them. A home you own costs more upfront and the roof is your problem, but nobody can raise your rent. Businesses made the same choice about computing over the past two decades, and mostly chose to rent: the cloud is other people's computers, cheaper and better run than your own server room. AI has so far followed the cloud's path. Everyone rents the frontier.

What this week changed is the visibility of the landlord's terms. When GPT-5.6 spent two weeks behind a government gate, every company that depends on a frontier model got a live demonstration of what renting means: your access is a policy decision made by other people, revisable at any time. And on the same day, the ownership option got 130 million dollars of funding and a working example. That simultaneity is the story. Not a coincidence to marvel at, but the same underlying fact showing both faces at once: control over frontier AI has concentrated so far that both the government gate and the escape from it became inevitable products.

The honest implication: the competitive edge in AI is quietly moving from who has the biggest model to who has the best data about their own work. Reinforcement learning turns a company's daily workflow into training fuel, and that fuel is the one thing a frontier lab cannot buy. If the pivot works, "which model do you use" becomes a boring question, like asking which brand of server a company owns.

The honest counterweight: do not mistake this for the frontier labs losing. Prime Intellect's own funding round is led by the most centralized names in hardware, NVIDIA, Intel, Dell, who win no matter which way the training happens. Most companies training their own specialist model will still run frontier models beside it for everything general. And the phrase the company hangs over all of this, an open superintelligence stack, should be treated as branding until proven otherwise. The open question worth carrying forward: when a company built on open, distributed ideals starts earning 100 million dollars a year from enterprises, does the open part survive? The next open model they release, or do not release, will be the tell.

Two things worth watching next. First, the next frontier release: Google's Gemini 3.5 Pro is reported for mid-July. If it also passes through the government's review process, the permission slip stops being an OpenAI episode and becomes the industry's operating procedure. Second, an independent reproduction of the Ramp benchmark, or any customer-trained specialist model evaluated by someone other than its maker. That single result, verified, would move this week's bet from funded thesis to demonstrated fact.

The two voices are AI. The research and writing are mine.

Decentralized AI, layer by layer.

Dastan,