The models are back. Anthropic's Fable 5 and Mythos 5 are live again after roughly two weeks offline, and the company will understandably treat the episode as resolved. It isn't. The fact that the US government could walk into Anthropic's infrastructure, flip a switch, and shut off the best models in the world for every customer on Earth, and then, just as easily, flip it back, tells you something important about how Anthropic's business actually works. It runs on permission. Permission from the cloud providers who own the hardware. Permission from the government that regulates what runs on it.

There is a version of Claude that wouldn't need that permission. A smaller model, running directly on a phone or laptop, that does useful work without calling home to a data center. Anthropic never built it. Not because they couldn't. Because they had every financial reason not to.

Follow the Money

Anthropic charges by the token. Every question you ask Claude, every document it summarizes, every line of code it writes. Each one registers as usage on a meter, and the bill goes up. That model has produced extraordinary growth: a $47 billion annualized run rate, the kind of number that makes an IPO filing look very attractive. The whole financial story depends on more people sending more queries through Anthropic's servers.

Now imagine an alternative. A compact version of Claude that sits on your device and handles the routine stuff: drafting emails, summarizing documents, basic analysis, all without ever touching the cloud. That model doesn't generate per-token revenue. It generates a one-time licensing fee, maybe an annual renewal. The economics are steadier but slower. You can build a good business on it. You cannot build the kind of revenue curve that justifies a near-trillion-dollar valuation.

So Anthropic made the rational choice. It signed roughly $80 billion in compute commitments through 2029, including a $1.25 billion monthly contract with SpaceX's Colossus 1 facility and a multi-year chip expansion deal with Google and Broadcom. Every one of those contracts exists to serve the cloud model. None of them would be necessary if a meaningful share of Claude's workload ran on the devices themselves.

What June 12 Actually Proved

When Commerce Secretary Howard Lutnick's directive landed on June 12, Anthropic had seventy-two hours to comply. The order required disabling Fable 5 and Mythos 5 globally, not just in specific countries, but for every user, including Anthropic's own foreign-national employees. The company pushed back, pointing out that competing models with the same jailbreak vulnerability faced no restrictions. It didn't matter. Anthropic complied because it had no choice. When your product runs entirely on servers you control, the government only needs to tell you to turn it off.

A model already running on someone's device is a different proposition. You can't remotely disable software that doesn't phone home. That is, of course, precisely why no frontier lab has shipped one. It would mean giving up the kill switch that lets you comply with orders like this. But here's the thing about the kill switch: it protects the company's relationship with the regulator. It does not protect the company's relationship with its customers. Those are different things, and on June 12 they collided.

The models came back after about two weeks. But every enterprise customer who watched it happen now knows the switch exists. Restoring access doesn't erase that knowledge. It confirms it.

Europe Got the Message

Within two weeks of the shutdown, Siemens, Renault, and Orange each publicly confirmed they were adopting multi-model strategies that include DeepSeek and Alibaba's Qwen. Orange's chief executive called the suspension "a clear demonstration of how crucial it is for Europe to have an AI service it can control." That isn't ideology. It's a procurement officer doing the math on vendor risk.

DeepSeek V4-Pro and Qwen's open-weight models can be downloaded and run on a company's own hardware. No API key from a US company. No server that a US agency can reach. That is the competitive property Anthropic cannot offer as long as Claude exists only in the cloud. And the restoration of Fable 5 and Mythos 5 doesn't change this calculus. If anything, it underlines it. The models were turned off, and then they were turned back on. Both actions required the same infrastructure. The Europeans noticed.

The Safety Argument Anthropic Could Have Made

The standard objection to on-device AI from frontier labs is safety: if you put a powerful model on a device, you lose control over how it's used. That objection has real weight. But Anthropic, more than any other lab, has the toolkit to answer it.

Constitutional AI, the approach Anthropic pioneered, trains the model to behave safely on its own, not because a server is monitoring it. The whole point is that the guardrails live inside the model's weights, not in a filter sitting between the model and the user. If that's true, and Anthropic has staked its reputation on the claim that it is, then a local Claude should be safe by design. The safety doesn't depend on the cloud. The revenue does.

Meta already ships on-device models across Instagram, WhatsApp, and Facebook, serving billions of users without a round trip to the data center. The technology works. Anthropic has deeper safety research than Meta. The missing ingredient was never capability. It was willingness to cannibalize the business model.

The Window Is Closing

There is still a path. A hybrid architecture: smaller Claude on the device for everyday tasks, full Claude in the cloud for the hard stuff. That would give enterprise customers something no other frontier lab offers: continuity when the cloud goes dark. A hospital running on-device Claude across ten thousand tablets for clinical documentation keeps working even if the next export control order lasts a month instead of two weeks. That customer is locked in by integration, not by dependency.

But Anthropic has announced nothing like this. Its roadmap runs toward longer conversations, more autonomous agents, and deeper cloud integration. Every product decision moves in the direction of more compute, not less. And the IPO, expected by October, will lock in the public market's expectation that per-token revenue keeps compounding. After that, pivoting toward a licensing model won't look like strategy. It'll look like the growth story is breaking down.

Anthropic didn't fail to build an on-device Claude. It chose not to, for reasons that made perfect financial sense at every step. The compute contracts, the revenue model, the IPO timeline: each decision followed logically from the one before it. That's what makes the June 12 episode so clarifying. The vulnerability isn't a bug. It's the architecture working exactly as designed.

About Shashi Bellamkonda

Loading bio…

LinkedIn Profile

Get new posts by email:

Subscribe

Powered by