Anthropic Fable Shutdown Forces Businesses to Control Their AI

Political risk is now an AI risk. The Anthropic Fable shutdown made it obvious.

TL;DR

• If your business runs on Claude or ChatGPT, you rent AI.

• The vendor controls the access, usage, costs, and limitations.

• They quietly learn your expertise as you use it.

• Anthropic turned off the most powerful model available, signalling to the AI industry that cloud-provided AI can be pulled at any time.

• Z.ai released a very powerful model that's downloadable, with open code to configure as you like.

• CTOs and business leaders are now prioritising ownership over their AI stack so that it can't be switched off, watered down, or trained on your work and sold back to your competitors.

A Fable is short, fictional story intended to teach a moral lesson, the lesson here is not to build your business on AI you don't control

What happened

On Tuesday 9 June 2026, Anthropic launched Claude Fable 5 and Claude Mythos 5. They were, by most public benchmarks, the most capable commercial AI models in the world.

Three days later, on Friday 12 June at 5:21pm Eastern time, the US government issued an export control directive citing national security. Because Anthropic could not filter foreign nationals out of its product in real time, the only compliant move was to turn both models off for everyone. Millions of users, including every paying business customer inside the US, lost access to the best model available at once. It was the first time a frontier model deployed at consumer scale had been pulled by government order.

The next day, Saturday 13 June, Z.ai (the consumer brand of Tsinghua spinout Zhipu) released GLM-5.2. A 744B parameter mixture-of-experts model with a 1M token context window, scoring 62.1 on SWE-bench Pro, sitting just behind Claude Opus 4.8 on coding tasks. MIT licensed. Open weights.

In plain terms: an incredibly powerful AI model, ranked nearly as good as Claude Opus 4.8 in globally recognised AI benchmarks, is available to download. You can edit the code and tailor it for your company. The license says you can use it for whatever you want, commercial or personal, no problem. It can read and remember massive amounts of information at once (about 750,000 words, or an entire software project), and if you have the hardware, you can run it for free. You can also use Z.ai's servers and pay a much cheaper usage fee than US rivals.

The contrast was almost cinematic. The most capable model in the West, gone in an afternoon by executive directive. A close-second model from China, sitting on any disk that wanted it.

Alex Stamos, former chief security officer at Meta and Yahoo, organised an open letter at freefable.org. It now carries 150+ signatures from executives at Adobe, Zoom, Sophos, Nvidia and Stanford HAI. The technical case the letter makes is narrow: the "jailbreak" cited as justification was a handful of previously known minor findings, not a unique capability that warranted recalling a model at scale.

The bigger point Stamos made was about second-order effects. By yanking the model, the administration injected a type of political instability and political risk into the US AI industry that did not exist a week earlier.

That sentence is the whole story.

What rented intelligence is doing to your business

A week after the Fable takedown, Microsoft CEO Satya Nadella published an essay on X titled "A frontier without an ecosystem is not stable." It drew more than 28 million views in a few days. The core warning was directed at every CEO using AI:

"The last thing any of us want is a world where every company across every sector is ceding value to a few models that eat everything they see."

Nadella's argument is that if you just rent intelligence from a handful of external APIs, those models quietly absorb your expertise on the way through and sell the lessons back to your whole industry. Your competitive edge erodes, query by query, into someone else's model.

Translation: rented AI is moat dissolution.

What is a moat

A moat, in business, is the thing that makes it hard for competitors to copy what you do. It's become more popular through AI advancement tearing down barriers to entry. The medieval version was literal: a trench of water around your castle that meant the next king over had to plan a year ahead and bring siege engines before they could take your harvest. The business version is figurative but works the same way. A brand, a network of customers, a patent, a proprietary process, a dataset no one else has, a team that has done a job ten thousand times. Anything that lets you charge more, or hold a market share, that an outsider with money cannot just replicate next week.

For some Australian CRE businesses, the moat has historically been people and relationships. Your team knows your clients. Your operators know your processes. Your leaders have judgment built over decades. AI arrived and started building bridges over those moats. New, AI-native market entrants are crossing them faster than longstanding businesses can dig new ones.

Nadella's reframe of the moat is what he calls "token capital." It is the AI capability a company actually owns rather than rents: proprietary learning loops, private evaluations, internal feedback that improves a model on your real work, and the operational traces of how your business actually executes. Keep those trajectories inside your perimeter and your business compounds an advantage no rival can copy. This increases the quality and capacity of your services while keeping your IP yours. Push them out through a rented API and you are training someone else's product on the way to losing your edge to whoever buys it next.

This is the same conclusion the Fable week forced on enterprise buyers, arrived at from a different direction. The first direction is continuity: your vendor can be turned off. The second direction is competitive: your vendor can absorb you. Either way, the conclusion is the same. You want more of your AI inside your own boundary.

What AI sovereignty actually means

"AI sovereignty" is one of those phrases that has been stretched to cover almost anything. Let me be specific about what I think it means.

There are three things you want to control.

The model. What it knows, what it refuses, what biases sit in its responses. With a closed model accessed via cloud or API, you have effectively zero control. The vendor changes the model weekly. Yesterday's safety filter is today's refusal. Yesterday's tone is today's confused customer. The vendor can also disappear at 5:21pm on a Friday, which is the lesson of the Fable week. With an open weight model, you choose the version. You decide when to upgrade. You decide when to stay. You decide what fine-tuning to apply. The model behaves the way you want, because you are running it, not asking permission to use it.

The infrastructure. Whose hardware the model runs on. For Claude and ChatGPT today, that is the US. For most "sovereign" Australian AI offerings, the data centre is here but the model still phones home or sits on a US-controlled stack. With open weights you have real Australian options: a managed AU-hosted private AI service that keeps your data onshore and under Australian contract, a dedicated deployment on Australian cloud, or your own hardware in the office. The choice of country becomes a procurement decision, not a vendor decision, and the practical setup is fast if you bring in someone who has done it before.

The IP layer that sits on top of the model. This is the moat in Nadella's sense. The prompt engineering, agentic workflow design, evaluation procedures on your own real work, the operational data that demonstrates which prompts and which workflows perform on the tasks your business actually does. If those things live inside your perimeter, on a model you control, they compound into capability no competitor can rent. Push them out through a hosted API and they compound into someone else's product roadmap.

The interesting thing is that all three layers are now achievable for a small business. Two years ago the best argument against sovereign AI was that the open models were too far behind to take seriously for production work. That gap has closed faster than almost anyone predicted. GLM-5.2 sits behind only the very best closed frontier model on coding benchmarks, and the gap on most business tasks is invisible. You do not need to be Microsoft to own your AI any more.

Why this matters more in Australia

Australian businesses are downstream of every move the US makes on AI. Almost every frontier model used in this country is a US model on US infrastructure under US jurisdiction. We have built our 2026 AI strategies on top of an assumption that the supply of intelligence from American vendors is stable, infinite, and apolitical. The Fable week was the proof that none of those three things is true.

Anthropic was instructed to restrict foreign nationals from accessing this technology, and we sit in that bucket.

AI is being embedded across all industries. Whether you personally use it a little or a lot, there are competitors in your industry becoming AI-native. Deciding to use vendor AI is technical debt that grows the longer you use it, and can be taken away without notice. Taking control of your AI strategy is how you build a moat that compounds, the same way good positions compound in finance and real estate.

For the compliance lens on this (privacy, expectations, cross-border data flow), the full picture is in our pillar on AI client documents in Australian CRE finance. The rest of this piece is about something more fundamental: the capability itself, and whether you own it.

What to do about it

You do not need to host a 700B model in your office to take this seriously. Three things, in this order.

• Identify the sovereign workloads. Not everything has to be. The ones that do are mission-critical and repeatable: how you solve client problems, how you assess deals, how you handle client information, how you brief your team. Those are the workloads where the operational traces feeding a rented model are training your competition, and where a vendor outage costs you real revenue. A general-purpose research assistant can stay on a hosted vendor.

• Take control of the model for those workloads. Pick an open-weights model on a deployment you control: a managed Australian-hosted private AI service, a dedicated tenancy on Australian cloud, or hardware in your office. Nobody outside your business can switch it off, change its behaviour, or read its inputs and outputs.

• Build your moat on top of that model, not someone else's. Your prompts, agentic workflows, evaluations and task-specific configurations are IP. Develop them against a model you own. The same work built on a hosted API is visible to that vendor and unportable when terms change.

The real point

The Fable takedown was not a one-off. It was a demonstration of a capability the US government always had and had never used. That capability now sits permanently in the toolkit of every administration to come, and of every other government that decides it would like the same lever.

Add to that the slower-burning problem Nadella named. Even when your rented AI vendor is online and friendly, every query you send through them is value leaving your business. Some of that value is a usage fee, which is fine. Some of it is operational data that improves their product and erodes your moat, which is not.

Owning your AI is a set of architectural choices made before the day you need them. Pick the workloads that are core. Run them on a model you control. Build your IP.

If you would like a practical view of what owning your AI looks like for a small Australian business, including model selection, hardware, and the parts of the stack that are worth owning versus renting, get in touch.