The portability premium — why vendor-independence is the new enterprise moat

THE AI NEWSLETTER THAT TURNS NEWS INTO ACTION

DEEP DIVE · COMPANION TO ISSUE #140

The portability premium — why vendor-independence is the new enterprise moat

On a Friday afternoon, one government letter made the most capable AI model on earth vanish from millions of accounts — no warning, no fallback. By Monday a Tokyo lab had sold the cure. The real lesson isn’t about export law or geopolitics. It’s that the thing worth paying for is shifting from the model itself to the freedom to lose it and keep working — and a market is already forming around that freedom.

BY R. LAURITSEN · 24 JUNE 2026 · 8 MIN READ

On 12 June, at 5:21pm Eastern, Anthropic received a letter. Within hours, the two most capable models it had ever shipped — Fable 5 and the Mythos system beneath it — went dark for every customer on the planet. Not throttled. Not deprecated with a migration window. Gone, mid-session, because a US export-control directive ordered the company to bar access by foreign nationals, and there was no reliable way to verify nationality at the API layer in real time. The only compliant move was to switch the models off for everyone. Eleven days later, they’re still off.

Plenty of ink has gone on the politics of that — who flagged the jailbreak, what the NSA did or didn’t tell a Senate committee, whether the President softened after a G7 handshake. That’s a story, but it isn’t the useful one. The useful one is colder and applies to you whether or not you ever touched Fable: a capability you do not control can be taken away on a schedule you don’t set, by people who don’t answer to you. This piece is about why that flips the entire value equation, and what to do before the next letter arrives.

What actually broke that Friday

Strip away the geopolitics and look at the mechanics of the failure, because the mechanics are what repeat. A workflow somewhere — a coding agent, a research pipeline, a customer-support bot, a contract-review tool — was pointed at a single model endpoint. When that endpoint stopped answering, the workflow didn’t degrade gracefully. It threw an error and stopped. The people who felt it worst weren’t the casual chat users; they were the ones who had quietly made one model load-bearing in something that runs without a human watching it.

This is the part that should unsettle a CTO more than the export law. The directive itself was extraordinary — and worth labelling honestly: the suspension is officially confirmed by Anthropic, the jailbreak rationale is the company’s reported understanding, and the more lurid claim doing the rounds (that the underlying model cracked classified systems “in hours”) traces to a single relayed briefing the original reporter has since cautioned against reading literally. Treat that one as contested. But the failure mode underneath was utterly ordinary. Models get deprecated on published schedules. Prices move. Rate limits tighten. A safety classifier shifts and a prompt that worked last week starts getting refused. A vendor has an outage; a region gets geofenced. The blackout was a sudden, dramatic instance of a risk that was present all along — ignored only because the earlier failures were gradual enough to absorb.

Capability got cheap. Continuity didn’t.

Here’s the shift hiding under the noise. For most of this technology’s short life, capability was the scarce thing — one clearly-best model, and access to it was the prize. That era is closing. Three or four labs now ship models close enough to the frontier that, for the overwhelming majority of real work, the difference is a rounding error: on most tasks you’d struggle to tell whether Opus, GPT-5.5 or Gemini 3.1 Pro wrote the answer. When several suppliers can all do the job, being able to do the job stops being a moat. It becomes table stakes.

So what’s scarce now? Not raw capability — resilience. The ability to lose your primary model on a Friday and not lose your Monday. Call it the portability premium: the value that accrues to whoever can treat any single model as swappable rather than essential. The week’s most-discussed launch is the clearest evidence the market already believes this. Sakana AI’s Fugu doesn’t try to out-build the frontier labs; it sits above them, routing each request across a swappable pool of models, and it was pitched — explicitly, in the launch copy — at teams and nations wanting resilience against exactly the kind of disruption that had just happened. It sold a beta to roughly 500 users before launch day on a promise that has nothing to do with being the smartest model and everything to do with never being hostage to one.

The counter-argument, taken seriously

Three honest objections deserve more than a wave-away. First: routing is a softer hedge than it sounds, because an orchestration layer still rents its intelligence from the same handful of vendors — if the frontier labs all face the same regulatory pressure, routing between them protects you from a one-vendor outage but not from an industry-wide squeeze. Second: for most people, none of this matters, because they use one chat app, never touch an API, and a blackout costs them an inconvenient afternoon, not a business. Third: the very best model still wins — Fugu’s own benchmarks show its top tier trailing the one model it couldn’t include, Fable 5, so portability buys resilience at the cost of a capability ceiling.

Each lands a real hit, and none is fatal. Routing-is-soft is true and worth saying aloud: resilience comes from the diversity of the pool, not the routing layer, so a serious strategy keeps at least one genuinely independent option — a different provider, or an open-weight model you can self-host — outside the main pool. Most-people-don’t-care is correct, and it’s exactly why this stays a professional discipline rather than a consumer panic: the audience that needs it is whoever runs AI inside something with a deadline or a customer attached. And best-model-wins quietly concedes the argument — yes, the frontier model is better, and yes, it was the one switched off. A ceiling you can reach beats a peak you can lose without warning. For the work that pays, reliable-and-available outscores brilliant-but-revocable nearly every time.

What to do before the next letter

If you run a company: treat single-vendor AI dependency as a continuity risk on the same register as a single payment processor or a single cloud region — because that’s precisely what it is. You already demand a fallback for those; AI has quietly become load-bearing without earning the same scrutiny. Put it through a continuity review. The question to force is not “which model is best?” but “what is our time-to-recover if our primary model becomes unavailable, for any reason?” If nobody can answer, that’s your first finding.

If you buy AI tools: read the continuity terms before the pricing page. The lens that matters: can the vendor explain, in plain terms, what model sits under each feature, what happens when that model is withdrawn or upgraded, whether they can fail over to an alternative without you rebuilding, how they pin versions so behaviour doesn’t drift, how they handle an outage, when they’d notify you, and what their recovery time looks like. A vendor who can answer those crisply is selling you resilience; one who can’t is selling you their supplier’s risk along with the features — and as of this month, that risk has a face.

If you build on AI: design for substitution from day one. Put a routing layer between your application and the providers, even a thin one, so swapping models is a config change rather than a rebuild. Pin exact model versions in anything automated — never “latest” — so a silent upgrade can’t change behaviour underneath you, and name a fallback model that takes over automatically when the primary errors. Keep one alternative genuinely warm: run a real task through it monthly, because an untested fallback isn’t a fallback, it’s a hope. And look hard at your own vertical — every industry that comes to depend on frontier AI will need a resilience layer between its workflows and the models. Someone gets to build that layer. It’s a business.

The bet, stated plainly

A thesis with no falsifiable claim attached is just commentary, so here are the terms. By the end of Q4 2026, at least one major enterprise software platform will ship multi-model routing as its default configuration — not a power-user toggle buried in settings, but the out-of-the-box behaviour, with single-model setups reframed as the exception you opt into and have to justify. The forcing function isn’t fashion; it’s procurement. Once a CISO or a CTO has watched a vendor go dark overnight, “single model, no fallback” becomes a line item that fails a security review, the way “single data centre, no backup” already does. Vendor lock-in stops being a default you quietly accept and becomes a risk you’re required to disclose.

What would prove me wrong: the labs settle into boring stability, blackouts stop happening, and the perceived risk fades faster than the tooling to manage it — leaving routing a niche for the paranoid. Or the capability gap between the best model and the rest widens again so sharply that buyers decide the frontier model is worth the dependency after all. What I am not predicting is the death of the frontier labs or the end of wanting the best model. The argument is narrower and, I think, harder to dodge: when the best model can be switched off by someone you’ve never met, the freedom to switch away from it quietly becomes the thing worth paying for.

Everyone spent the week asking which model is best. The more useful question — the one this blackout forced into the open — is what it costs you to lose it. Until you can put a number on that, you don’t fully know what your AI stack is worth. That’s the number to go find. If you think I’ve called this wrong, reply to this week’s issue and tell me why — I’ll print the sharpest counter-argument.

iPrompt

PUBLISHED BY FRONTWAVE MEDIA LTD · LIMASSOL, CYPRUS · IPROMPT.COM