Anthropic's recent US restriction on Fable 5 teaches us something about the changing landscape in AI models, infra, restrictions and regulations. Are you the tenant who builds on someone else's AI Infra?
By Jedidiah
I've been in a lot of enterprise AI conversations over the past year. And lately, something has quietly changed.
For most of 2024, the question was capability. Which model is best? How fast can we get to a demo? The implicit assumption was that speed to deployment was the goal, and infrastructure was a detail to figure out later. That made sense at the time. We were all still learning what was possible.
Then Anthropic updated Fable 5. Enterprises that were building real workflows around that model's behavior found themselves scrambling. Not because of a catastrophic failure. Because something they depended on changed in ways they didn't see coming and had no power to prevent.
I keep coming back to that. Because the scramble itself is the message: what exactly is the ground we're building on?
"Speed to deployment is not the same thing as a strategy. One is a sprint. The other is a foundation."
There's a real tension here, and I want to name it
The API-first approach works. It gets you moving fast, gives you access to frontier capabilities without heavy upfront investment, and lets you iterate while the technology matures. For early-stage use cases, it's often the right call.
But there's a structural catch that only becomes visible once AI is deeper in your operations: when the intelligence powering your workflows lives outside your walls, so do the decisions that affect it. Model behavior evolves. Pricing changes. APIs get deprecated. None of this is malicious. Vendors are running businesses, and businesses change.
The question is whether your architecture absorbs that reality gracefully, or whether it inherits the vendor's decision as an emergency.
| Renting your intelligence | Owning your intelligence |
|---|---|
| × Model behavior changes on the vendor's schedule, not yours | ✓ You validate model versions before production and upgrade on your own schedule |
| × Proprietary data flows through infrastructure you didn't build and can't inspect | ✓ Data stays inside your perimeter and never needs to move to be useful |
| × Compliance depends on third-party promises, not your own controls | ✓ Compliance is demonstrable end-to-end, with audit trails you actually control |
| × Pricing and availability sit outside your planning horizon | ✓ Infrastructure costs are predictable, not subject to external repricing |
| × Vendor sunset decisions land in your lap as emergency migrations | ✓ Your AI roadmap follows your business strategy, not a vendor's product calendar |
The regulatory ground is shifting faster than most realize
In Singapore especially, the question of where AI runs and who controls it is moving from boardroom conversation to regulatory requirement. NAIS 2.0 and MAS guidelines are asking harder questions: where does AI processing happen, who owns the audit trail, and how can you demonstrate that your AI-driven decisions are traceable?
For regulated industries — financial services, healthcare, government — "we use a reputable frontier model" isn't a governance posture. It's an incomplete sentence. The rest of it needs answers about data residency and processing sovereignty that public API deployments often can't provide.
Singapore's positioning
NAIS 2.0 frames sovereign AI infrastructure as a strategic asset, not a compliance checkbox. The organizations that build for this now will have a real advantage as regulatory expectations tighten across APAC. The interesting question isn't whether to build for sovereignty — it's how to do it without slowing down.
What sovereign AI actually means
A misconception I run into often: people think sovereign AI means building your own foundation model. It doesn't. That's neither realistic nor necessary for most organizations.
What it actually means is control. Where your AI runs. What data it can see. How outputs are governed. What the audit trail looks like. You can use excellent models without giving up control over those things.
Three principles tend to matter most in practice:
Bring the intelligence to the data, not the data to the intelligence. Sending sensitive enterprise data out to an external model to get it processed is a risk that accumulates quietly over time. Zero-copy architectures flip that: the model comes to the data, inside your perimeter. The answer to "where did this data go?" becomes "nowhere." That's a much cleaner compliance story.
Traceability is not optional in regulated environments. When an AI system informs a decision, you need to know which documents it drew from, what access policies governed the retrieval, and which model version was running. Good RAG does this. It's not just about returning relevant results — it's about maintaining a chain of custody for every answer.
Your AI roadmap should answer to your strategy. The practical benefit of owning your stack is that you adopt new capabilities when you're ready — tested, validated, deployed on your own terms — not because a vendor deprecated something you were relying on.
What Fable 5 was actually telling us
I don't think the Fable 5 situation was a failure. I think it was clarifying.
The enterprises that scrambled learned something valuable: the speed of their deployment had outpaced the resilience of their architecture. That's a common pattern in any technology wave. The ones who come out ahead are rarely the first movers. They're the ones who built thoughtfully enough to absorb the surprises.
Owning your foundation isn't a reason to slow down. It's actually what makes sustained speed possible — because you're not managing fragility underneath every decision.
That's the thinking behind Cohesity Gaia. Secure, cited RAG that keeps your data where it belongs, gives your teams answers they can stand behind, and puts your organization in charge of its own AI trajectory. Not renting intelligence from someone else's infrastructure. Building on ground you own.
The conversation is changing. I think that's worth leaning into.
The future of enterprise AI belongs to those who build on a foundation they own.
SOVEREIGN AI ENTERPRISE AI DATA GOVERNANCE RAG COHESITY GAIA AI STRATEGY NAIS 2.0