Sovereignty is the idea I've built my career around, but it's a word that's become so widely used, especially in AI, that it's lost some of its gravity. Take Palantir's Alex Karp, who went viral last week for a combative CNBC segment on the danger of building your company on top of centralized frontier labs. He's right, but the irony is he’s there on a roadshow for an AI-sovereignty partnership between two of the world’s biggest frontier tech firms Palantir and Nvidia.
Strip the buzzword away, and it’s the difference between technology that works for you and technology that works on you. Sovereignty is a spectrum, and working towards it is a worthy mission, especially for those building in frontier tech. It comes down to who holds the keys, who can cut off your access, and who gets to use your data as it moves through the system.
When Anthropic's Fable arrived as the most capable model anyone could use, it was pulled by the government on national-security grounds in a matter of days. Palantir spent the same stretch making its case in a sovereignty manifesto, anchored on "Controlling your weights is controlling your fate." Chamath and the All-In besties spent the top of last week's episode on it, largely making the enterprise case for open weights and working through the economics against the frontier.
I wrote during the Fable blackout that frontier AI is rented, not owned, but I didn't expect the point to be proven many times over inside a single month. For most of AI’s short history, the open-source-versus-frontier tradeoff was an argument among researchers and policymakers. Now it’s more relevant for boardrooms and executives.
What you hand over with every prompt
A company pays a couple hundred dollars a month for a Claude seat. When an employee inevitably maxes out utilization, the compute behind that seat costs Anthropic something closer to ten thousand, and it’s likely that the token subsidies will fizzle before their commoditization drives prices back down. Through it all, Anthropic's run rate went from $19 billion in March to $30 billion in April to $44 billion by early May. The clock on all of this started ticking in April, when a few hundred CFOs looked at that number and realized they had torched their entire 2026 AI budget by the end of Q1. The models kept improving and the budgets broke anyway, because the only unbounded variable in the contract is consumption, and consumption scales like a parabola.
Cost is the part everyone can see. The part that should worry them more is that renting intelligence never stops at intelligence. Every prompt carries your context out with it, and your context is your edge: the workflows, the proprietary data, the accumulated judgment that make a model worth anything inside your business. Send it through a third party's API a few million times a day and you are handing that edge to a frontier lab to be distilled, and no one has to train on it against the terms for that to hurt you.
Aggregate usage already shows them where the value is. This is the platform-risk playbook and it isn’t new. Zynga was a $10 billion company until Facebook decided the traffic was worth keeping, and Amazon watched what sold on its marketplace and shipped private-label versions of the winners. In AI it just happens faster. It’s how a lab ships a feature that lands squarely on a public company's core product and takes half its market cap in a session, and then does it again in the next category, and the one after that. Watch it happen five or six times and the question stops being whether they can read the signal and becomes how fast they can build on it. In their seat, with their cost of capital, I would build on it too.
Markets eventually price those two things separately, and when they do they fall back on an old rule: you rent commodities, and you own your edge.
Rent or own, one call at a time
That is the argument the industry is now widely having. Build or buy, own the models or rent them. The case for owning got much stronger this year, because the supply side finally moved.
Open weights got good enough, and they will keep getting better. GLM 5.2 shipped at a level that stands up against Opus 4.8, and the part that mattered was the capital, because the hundred-billion-dollar capex wall that was supposed to protect the labs got cleared by someone willing to give the result away. People fixate on the fact that the lab was Chinese, and for a regulated US enterprise that is a real problem today. It is also the reason the gap closed at all. China is running a scorched-earth AI strategy, pouring capital into open models specifically to knock the American labs off their pricing, and once the frontier has been matched and handed out for free, open models from friendlier jurisdictions follow. I would expect Meta back in serious open-weight work inside a year, because they fell off the frontier pace and going open is the only way back into the fight.
An open weight out of Shenzhen is not something a bank runs as-is, and it shouldn't. But the industry has solved this exact problem before. IBM paid $34 billion for Red Hat in 2019, and because Linux itself is free, what IBM actually bought was the adult in the room: the entity that takes free software, strips out what enterprises can't accept, indemnifies it, and commits to a decade of support with no surprise end-of-life. Someone becomes the Red Hat of open weights, taking a capable model, cleaning the provenance and the censorship out of it, and republishing it under a name auditors trust with the enterprise wrapper attached. At that point you are paying for a support agreement and a services model, not for per-token intelligence, and that is a radically lower place to sit on the curve.
The other half is the orchestration layer, the software that feeds a model your context, runs the fine-tuning and the reinforcement work, and manages it inside a real enterprise environment. That is what Palantir's ontology architecture is, and it is what Accenture and every systems integrator with a pulse is now building a practice around. It isn't there yet, and neither is the workforce, because very few people can run a post-training pipeline properly right now. That is a human capital problem, and human capital problems get solved with time and money.
The strongest case against all of this comes from the people who spent their careers getting regulated institutions through audits, and it deserves acknowledgement. Tell an examiner you use a Thales HSM and the conversation moves on. Tell them you built your own HSM and you've signed up for three weeks of them going through your life choices. A model you trained yourself gets treated like an HSM you built yourself, which is a large part of why the most regulated shops move last, and why they are right to.
The build-versus-buy history points the same way. Plenty of organizations saved money self-hosting open source and then met the payroll of the team it took to keep it running, though the thing being kept running here is a compute stack these companies already operate at scale, not a bespoke application. And the labs may simply cut prices. Anyone who worked at a bank in the mid-nineties (I'm looking at you, David Schwed) remembers paying for internet access metered per employee, because a T1 line ran $3,000 a month, and that meter died the same way AOL's per-minute pricing did. Flat-rate AI may well arrive before self-hosting does. But a flat price solves the CFO's problem and leaves the harder one alone. A subscription does not change whose infrastructure your context runs through, or who is allowed to switch it off.
What resolves the argument is routing, the architecture that makes the rent-or-own choice one call at a time instead of one company at a time. Coinbase's CEO posted the numbers in late June: their third-party model spend fell by nearly half while token usage kept climbing, once they routed lower-value work to open-weight models they run themselves behind a gateway. Coinbase is a sophisticated shop, granted, but the same test is running well beyond crypto. On All-In this week Chamath described wrapping an open model in 8090's own harness and landing near a sixteenth of frontier cost for his use case, and then he said 8090 will run that optimization across the major open-weight models, Chinese and American, starting with Nvidia's Nemotron, and publish the results. The publishing is the part that should worry a frontier lab, because once those optimizations are open, every enterprise inherits them at once, and the open models climb toward the frontier in public instead of behind a subscription. Armstrong figures 80% of workloads land on radically cheaper models inside twelve to eighteen months.
The realistic setup is a secure, compliant shared services environment you can trust, running a model you post-trained on your own context, at a dollar per million tokens instead of fifty, and it is almost as good for almost everything you ask of it. For the hard five percent, the critical code review, the work where being wrong is expensive, you route out to a frontier model and pay the fifty. Standing that up the right way with the current tooling is a several month project for three to five people, not a moonshot. That is what we are doing at SVRN. I would put the endpoint more aggressively than Armstrong does. Most of the Fortune 500 adopts some version of this within two years, and the sharp end reaches 95% of workloads on owned, post-trained open weights, with the frontier holding the last 5% where $50 a million tokens is worth it the way outside counsel is worth it. The only open questions left are the ratio and the timing.
There are degrees of this
What the manifestos skip is that ownership is not binary, and neither is sovereignty. There is a spectrum between them. At one end you rent everything and hope the vendor's interests stay aligned with yours, and at the other you own your weights, your keys, and the environment your agents run in. Almost every real company lives somewhere in the middle, and moving a single step toward the owned end carries real cost. I am not going to pretend a mid-market business should go build its own confidential-compute cluster next quarter.
What matters more than where you sit today is the direction you move, because value rarely stays with the scarce input. It left bandwidth for the networks that carried it and compute for the cloud that made it usable. Cloud revenue roughly tripled between 2019 and 2025, long after anyone had stopped debating whether servers were a business. Intelligence is on the same path. As it becomes a commodity, the value moves to the layers that give it context and settle what it does.
The same split runs one level down, from the enterprise to the individual. The argument a company is having about its weights is the one each of us is about to have about our own agents. Once software is transacting and negotiating on your behalf, whoever controls its keys and its execution controls the economics underneath it. NEAR made that bet in 2018, building chain abstraction and verifiable execution for a world where software would act for you, and you can see it now in NEAR AI running models inside confidential compute and IronClaw refusing to hand an agent your credentials. It is the enterprise fight rewritten for a personal account.
I spent years as an investor before building SVRN, and the pattern holds across every cycle I've watched: scarce inputs grab attention, and the infrastructure beneath it takes the returns. Frontier intelligence turns into an input you route to, owning your context turns into the baseline for anything that matters, and the infrastructure that lets a company or a person hold their own weights, keys, and agents turns into where the value settles. That is the layer NEAR was built for, and getting it adopted at institutional scale is the problem I’ve chosen to work on as an operator. Arguments like this get settled in the market over the next several years, which is the only scoreboard that counts, and it is why the positioning has to happen now.
overview
