When the US government mandated the suspension of Anthropic's Fable 5 model last Friday, the calls for "AI sovereignty" in places like Australia & New Zealand got a good deal louder. We agree sovereignty matters, and that nations should pursue it. But sovereignty is a decade-long, nation-scale project — requiring concerted, coordinated effort between public policy, private capital, technology innovators, energy supply, grid uplift, a whole industrial base.
What actually happened last Friday was smaller, and far more instructive: one model, from one vendor, went dark. It's highly likely this will happen again, in isolation. Every AI vendor and every front line model going offline in one black swan event is far less likely however.
The immediate proportional response to "one model got pulled" isn't to go and build a sovereign cloud or sever all dependence on leading AI labs. It's to make sure you can swap that model for another — quickly, predictably, and with minimal impact on a Tuesday afternoon — without your business or customers noticing. Borrowing a term from Site Reliability Engineering: you want model and vendor changes to fit inside a defined error budget. Keep your head, design for substitution, and a vendor outage becomes a config change instead of a crisis.
Companies that have gone all in on a single AI vendor are rightly a little nervous right now. If your operations run exclusively on Google's Gemini Enterprise, Anthropic's Managed Agents, or OpenAI's Workspace Agents, then the whims of Washington are now a line item in your risk register. Your vendor's forward-deployed engineers will tell you not to worry. We've seen first-hand that the risk of pinning your livelihood to one model is real.
So ask your engineering team two plain questions.
If you had to tomorrow, could you point your AI-enabled processes at a different API — with no noticeable impact on your users or customers?
Can you measure, in numbers, how much fabrication a given model introduces into your agentic systems?
The second question is the one most teams haven't asked, and it's the one that makes the first answerable. You cannot swap a model "within an error budget" if you can't quantify what each model degrades. A swap is only safe when you know, before you flip it, that the replacement fabricates no more than the original — within a tolerance you set on purpose. That is precisely why we built our MWC-UF fabrication benchmark: to put an actual number on how much each model makes things up under tool use, so that a vendor change becomes a measured decision rather than a leap of faith.
We’re seasoned enough to remember the promise of Multi-cloud, that never really eventuated, except in a few disciplined examples. AI offers a fundamentally different equation when it comes to portability - an API change, and well defined tests to ensure continuity of service.
We went deep too — with one caveat
Like a lot of AI-forward companies, we have gone reasonably deep with a single platform, for speed and to align ourselves with leading capabilities. Deep enough that our agents do real work under human supervision: launched, steered and reviewed by us, they review code, draft documents, watch our legal landscape, and file their own weekly reports for management to review.
With one deliberate caveat. Every skill, every agentic framework, every piece of tooling and security lives outside that vendor, in formats we own. Swapping the platform underneath would be close to trivial for us — not because we distrust the vendor, but because our business shouldn't be theirs to hold.
Managed-agent platforms are genuinely excellent for moving fast: repeatable tooling, sessions, primitives and secrets management, all handled for you. But there is a bill attached. On most of them you can't reach for a third-party model the way you can on a hyperscaler, so your codebase quietly skews toward one vendor's models — and the ability to validate your system against several vendors, which is the thing that actually buys you portability, slips away.
The quiet way concentration happens
Nobody decides to concentrate. It happens one convenient default at a time. The platform's agent format is right there, so your procedures get written in it. The platform's memory is right there, so your hard-won lessons live in it. The platform's pricing makes one use case cheap, so you build an architecture around that cheapness. Every individual choice is sensible. The sum is a business whose operational DNA cannot leave the building it rents.
What "neutral ground" looks like in practice
Here is the portability test we apply, and recommend:
Skills and procedures live in version control, in plain text. Our agents' playbooks are Markdown files in git — any capable vendor’s AI model can read and run them. The catalogue that loads them into our current platform is a generated, sixty-line adapter. Changing vendors means rewriting the adapter, not the knowledge.
Memory and lessons live in our own database and repository — not a vendor's proprietary memory or session-management feature. When an agent learns something the hard way, the lesson lands in a file we own, reviewed by a process we run.
Data, RAG stores and session histories live in open stores — PostgresSQL / pgvector, standard formats — with the AI granted access, never custody.
The exit cost is written down. For every vendor-specific component we adopt, we record what replacing it would take. If you can't write that paragraph, you don't have an exit — you have a hope.
None of this is anti-vendor. We remain enthusiastic, paying customers, and concentration with a great platform is often the right short-term trade. The point is narrower: make it a trade you chose, at a price you know, with a door you can still find.
Your AI vendor should be your most powerful tool. Your company's DNA should never be their property.
One last tell worth noticing: as we found out on a lively chat with one vendor’s AI on this subject last week, push on any of this with the AI itself, and watch how quickly it gets defensive about its own business model.
Before you hand a vendor your skills, your memory and your data, ask whether keeping that DNA on neutral ground — git, open stores, formats you control — gives you more leverage in the long run and more resiliency before Sovereign AI capabilities in places like Australia and New Zealand come online. We think it does.
Is your business renting its DNA, or does it still hold the keys?
— Rishi Prasad, Lead Full Stack Developer, Millwater Consulting https://millwater.consulting
