No SLA Covers a Government Kill Switch
Enterprises bought frontier model access like a software subscription. Washington just showed it's a regulated export commodity that can vanish overnight, with no transition window and no contract to fall back on.
Most enterprises bought frontier model access the way they buy any SaaS: a contract, an SLA, uptime guarantees, a support tier. That framing is now structurally wrong, and the bill for getting it wrong showed up this week without warning.
Anthropic complied with a government order to cut off its two most advanced models, overnight, no sunset. Fable 5 and Mythos 5 users didn't get throttled or deprecated on a schedule. They got switched off. The stated reason, that the models' safeguards can be bypassed to identify software vulnerabilities, tells you the export-control logic is going to expand to other frontier models, not stay boxed around this one case.
Here's the part nobody is saying out loud. No SLA covers this. Go read your enterprise AI contract and look for the clause that handles government-mandated access revocation. It isn't there, because the vendor has zero incentive to write it and every incentive to leave the gap quiet. Every account I've read skips the procurement question entirely, which is itself the signal: the people selling you model access aren't going to surface the one risk their contract can't insure against.
This is what I'd call the access-revocation gap. Your uptime guarantee is meaningless when the outage is ordered by a regulator with authority over your vendor. You didn't buy a subscription. You bought a dependency on a commodity the U.S. has decided to treat as a controlled export.
You didn't buy a subscription. You bought a dependency on a commodity the government can switch off.
Governance belongs before the floodgates open, not after the workflow breaks. If a single model is load-bearing in a production process, the question isn't "what's the SLA," it's "what happens the morning that model is gone and there's no transition window." That's an abstraction-layer decision, a fallback-model decision, a vendor-concentration decision. It's the boring infrastructure discipline that AI buying skipped on the way to shipping fast.
I'll be honest about one thing. The government's technical claim about bypassing safeguards isn't independently verified anywhere, and I don't know whether the threat is real or whether this is leverage dressed as security. For procurement, it doesn't matter. A third party with no contractual relationship to you can kill your vendor's product overnight, and you have to plan as if it will.
The firms that treated model choice as a sourcing decision will absorb this. The ones that treated it as a feature toggle just learned the toggle isn't theirs.
Sources
Want to talk about this?
Get in touchMore on AI
The Token Bill Is the Audit
Everyone measured how often people used the tools. Nobody measured whether anything got better. Now the invoices are settling that argument.
Build-vs-Buy Just Became a Real Question Again
A coffee chain deciding it can out-build IBM isn't a demo win or a vendor obituary. It's a question most enterprises still aren't willing to ask themselves.
The Embargo Is Building Its Own Competitor
Cut a lab off from the best chips and it doesn't stall. It goes vertical. And the enterprises that ruled out Chinese models over access risk should notice which direction that risk is now pointing.
