Data Sovereignty & The End of Vendor Lock-In
On Friday, two of the most capable models companies had built on went dark for everyone, all at once.
The US government issued an export-control directive, citing national security, ordering Anthropic to suspend access to its two newest models — Fable 5 and Mythos 5. To comply, Anthropic disabled both models for every customer, effective the same day the order arrived. Not a tier. Not a region. Everyone. A team that had wired Fable 5 into a product on Thursday had no Fable 5 on Friday. They didn’t get a vote, a deprecation window, or a migration path. The decision was made several layers above their contract, and their access was revoked..
Set aside, for a moment, the merits of the order. Anthropic is contesting it, and reasonable people will argue about national security and export control for months. The part that matters for anyone building a business on top of a model is simpler and harder to argue with: the most capable models on the market can become unavailable to you overnight, for reasons that have nothing to do with you, decided by people you will never meet. You can do everything right and still lose access by Friday afternoon.
That is the cleanest version of a lesson the rest of the month had already been spelling out. Earlier, Anthropic updated its enterprise terms to claim rights over customer data for model training. What happens to your data became a clause in someone else’s document, revised on someone else’s schedule. And Microsoft, a company with an effectively unlimited budget and an army of lawyers, suspended its own employees’ access to new Claude models, deciding the prudent posture toward a frontier vendor was to cut its own people off. Tracy Alloway named the underlying anxiety in one sentence: “the smart money is worried about all the AI companies that are being thinly built on top of foundational models that they don’t actually control.”
That worry was already correct. Friday made it concrete. It is worth being precise about why.
Every frontier API is a dependency, and dependencies have owners
When you build on a closed frontier API, you are not buying a capability. You are renting one, on terms a landlord can change, and on access a third party can pull.
This is not a hypothetical risk anymore. Over the past eighteen months the market has watched closed vendors reprice tiers, deprecate models that products were built on, rewrite usage policies, restrict access by customer category, and reach toward training rights over enterprise data. Then, on June 12, it watched two flagship models disappear for the entire customer base in a single day under a government order. None of those moves required your consent. None of the next ones will either.
The people making these decisions are not villains. Anthropic did not want to switch Fable off, and it is fighting to turn it back on. Frontier labs are full of careful engineers who care about a lot of the same things we do, facing a brutal equation: training runs that cost billions have to be recovered somewhere, and the levers available are price, terms, and data. And sitting above all of them is a layer of risk that no lab controls: regulators, courts, export authorities, and national-security desks that can reach a vendor’s API and, through it, reach you. When a vendor needs margin, your invoice is where margin lives. When a vendor needs training data, your terms of service are where training data lives. When a government acts on a vendor, your access is what disappears. Ask anyone who was running on Fable 5 on Thursday.
So name the dependency honestly. If a model you don’t control sits in your critical path, then your roadmap has a silent co-author. Your unit economics have a floor you don’t set. Your data governance story carries an asterisk you can’t remove. And your continuity plan rests on the assumption that a company under enormous capital and regulatory pressure will keep behaving next quarter the way it behaved last quarter — an assumption that just failed, in public, for everyone at once.
“Thinly built on top of foundational models that they don’t actually control” is not an insult. It is an accurate balance-sheet description of a large share of the AI industry right now.
What owning your stack actually means
Open weights change the shape of the relationship, not just the price of it.
A closed model is a service. An open-weight model is an artifact. Once the weights are released, the model is infrastructure — it can be served by anyone, anywhere, indefinitely, and the best open models now go toe-to-toe with closed frontier on the work most businesses actually run. The capability gap that justified the closed premium has mostly closed. What remains is a difference in who holds the keys.
Three things follow, and each one answers a risk named above.
Your data stays home. When inference runs on open weights served on infrastructure you, or a serving partner whose incentives run with yours, control, prompts and outputs never transit a lab’s systems. No terms-of-service update can reach data it never touches. The enterprise-terms story is only alarming if your data sits in someone else’s pipeline; the structural fix is for it never to be there.
Your access cannot be revoked. Nobody can deprecate a model you hold the weights to. Nobody can suspend your tier, sunset your endpoint, or decide your use case is no longer served, and no directive aimed at a vendor’s API can reach a model you already serve. When Fable 5 and Mythos 5 went dark on June 12, every open-weight model running on private infrastructure kept running, exactly as it had the day before. That is not luck. It is the difference between a service that can be switched off and an artifact you already hold.
Your roadmap is yours again. Model choice becomes a routing decision you make per task, not a marriage you renegotiate per contract cycle. When a better open model ships (and one ships every few months now) you adopt it the way you adopt any infrastructure upgrade: on your schedule, with your data exactly where it was.
To be clear about what we are not saying: owning your stack does not mean every company racking GPUs in a closet. Self-hosting is its own unglamorous discipline, and most teams should not take it on. The point is that open weights make ownership possible, served via API by a layer built for it, on infrastructure where your data stays yours, with no training-rights clause waiting in the next terms update and no third party able to switch the model off. Ownership is about where control sits, not where the servers sit.
The cost gap is structural, and it runs the same direction
Sovereignty would be worth paying a premium for. The strange fact of this market is that it comes at a discount instead.
A closed frontier lab has to price every token to recover the cost of training the current flagship, fund the next one, and return a margin on tens of billions of invested capital. Serving an open-weight model that has already been trained only has to cover inference plus a margin. Those are different equations, and the open one has a lower floor by construction. In practice the gap runs from roughly 5× on reasoning-heavy work to 20× on routine volume; when we priced three real jobs end to end this month, routing each to the right open model came in about 15× cheaper than running everything through a single frontier model, with output a reviewer would struggle to tell apart. Against frontier rate cards, that is on the order of 80–95% less for the same finished work.
This gap does not close from the expensive side. The closed vendor’s floor is set by an R&D burden the open-weight server simply does not carry. Which means the pressure you are watching, terms reaching for data, repricing, access games, and now access pulled by forces outside the vendor entirely, is not a phase the industry grows out of. It is what the closed cost structure, and the closed point of control, require. The longer you rent, the more of that pressure routes to you.
Where this lands
The question facing every company building with AI has quietly changed. It used to be “which model is best?” It is becoming “who controls the layer my business now depends on?” After June 12, that question has a sharper edge: it is no longer only the vendor who can answer it. The companies that can say “we control it” will compound. The rest will keep paying rent that only moves in one direction, on access that someone else can end without warning.
This is the thesis BasedAI exists to serve. BasedAI is the acceleration and commercialization layer for open source AI. We build products that turn the best of open source into reliable, useful work for real people and real businesses. Hirebase, our AI coworkers product, already runs this way: every job routes to the right open-weight model, served on infrastructure we control, and customer data never leaves to train anyone’s model. For teams that want this layer as raw infrastructure rather than finished work, that’s BasedAPIs — open weight models served reliably for production workloads, coming soon.
The smart money is right to worry about companies thinly built on models they don’t control. Friday showed what “don’t control” can cost on an ordinary afternoon. The answer is not to build less on AI. It is to build on AI you own.
Own your stack. Don’t rent it.