What happened on June 12 was unprecedented in the history of commercial software: The U.S. Department of Commerce imposed export controls on Anthropic’s flagship models, Fable 5 and Mythos 5, effectively forcing them offline—out of concern over their cyber capabilities. Companies worldwide that had built their workflows around the model were left without it overnight. On June 30, the agency lifted the controls; since July 1, Fable 5 has been up and running again globally, including on AWS, Google Cloud, and Microsoft Foundry.

Forced Hiatus Sets a Precedent

More interesting than the return itself is what Anthropic is making of it. The company is publicly calling for government intervention in AI releases to follow a transparent, legally enshrined process in the future—and for all Frontier Labs equally. At the same time, it is working with Amazon, Microsoft, and Google to develop an assessment framework that rates the risk posed by security vulnerabilities, so that a borderline case does not trigger another total shutdown. And according to the Financial Times, the White House is already negotiating voluntary standards for the release of state-of-the-art models. The forced hiatus is thus setting a precedent: it is defining the rules of the game between the government and the AI industry.

The Shift to Hardware

While the battle for software sovereignty raged in Washington, the search for answers to the question of power also began one level deeper—at the silicon level. OpenAI kicked things off: On June 24, the company unveiled “Jalapeño,” its first in-house AI chip—an inference processor developed in collaboration with Broadcom that, according to the company, delivers more computing power per watt than the competition. The objective is clear: to reduce dependence on NVIDIA GPUs, whose prices and delivery times represent a significant cost factor for the industry. It’s no coincidence that an inference chip, of all things, is leading the way—as the business shifts to mass production, the cost focus is moving from training models to their continuous operation, and that’s exactly where every watt counts. The labs are thus following the lead that hyperscalers have long set: Amazon has been relying on its own silicon for years with Trainium, and Google with its TPUs.

Capital and Manufacturing Converge

Anthropic’s response was prompt. On July 2, The Information reported that the company was in talks with Samsung about manufacturing its own AI accelerator—with an eye toward the 2-nanometer process and Samsung’s packaging capabilities. Nothing has been decided yet; neither the intended use nor the performance specifications have been finalized. But the personnel move behind it speaks volumes: Anthropic recently hired Clive Chan—one of the first engineers on the very OpenAI team that built Jalapeño. The backdrop to both projects: According to estimates by The Information, Nvidia continues to hold 74 percent of the AI chip market—and despite all attempts to diversify, dependence on the company has grown rather than shrunk. Particularly noteworthy in this context: Samsung invested in Anthropic’s $65 billion funding round as recently as May, alongside memory manufacturers SK Hynix and Micron.

Export Controls Trigger a Backlash

The episode has had ripple effects all the way to China: The 19-day ban visibly accelerated work there on models trained entirely on domestic chips—Meituan’s LongCat 2.0, released as open source in early July, is already being explicitly marketed as an alternative free from U.S. dependence.

Two Sides of the Same Coin

Both reports tell the same story from opposite perspectives. At the top of the stack, Anthropic experienced just how quickly a government can pull the plug. Further down the stack, the company is working to ensure that, at least when it comes to silicon, no one else has their hand on the switch. Control over one’s own value creation—that is the currency in which the AI industry will be calculating in 2026. The fact that OpenAI and Anthropic—the two leading labs—took the same path within a matter of days means this is not an isolated case, but rather an industry consensus.

A Blueprint for the Industry to Follow

For industrial companies, this case serves as a blueprint for their own operations. Anyone who uses AI productively in design, production planning, or quality assurance should view these 19 days as a stress test: Which processes would grind to a halt if the model from the in-house supplier were unavailable tomorrow—for regulatory, geopolitical, or commercial reasons? Multi-model strategies, defined fallback plans, and contractual exit clauses are no longer optional extras but essential for operational security. Anthropic is diversifying its chips. The industry should diversify its models.