When the frontiers keep shifting
THE COMPUTE / BY TOKENANDO Issue VII / June 18, 2026
On the evening of June 12, 2026, at 5:21pm in Washington, a letter reached Anthropic from the United States Commerce Department. By then, the most powerful model the company had ever released to the public was a few days old. By the end of the night it was gone, switched off for everyone, all over the world.
More than 100 security researchers have since signed an open letter asking Washington to reverse course. This does not change the fact that for organisations all over the world (including the US in this case) this sets a significant precedent, where a frontier capability used as the basis for meaningful product and development work (one must assume that was the case, and not counting R’s in strawberry) can be recalled within a few hours by a government decision, over a concern that is not made clear.
As I write this issue, I am on my way to Paris Expo, where the second day of VivaTech 2026 is starting in an hour or so. Day one was full of interesting and eye-opening sessions. I will hold off until the event concludes this Saturday to collect all my thoughts and insights, which will make their way into next Thursday’s issue. If you cannot wait to hear more, contact me and I can give you the debrief starting Monday next week!
And now without further ado, we dig deeper into the Fable situation and more… read on!
Behind the scenes of the Fable lockout
Researchers at Amazon, Anthropic’s largest investor and the hyperscaler hosting much of its AI cloud, found that Fable 5 could be coaxed into surfacing software flaws, leading chief executive Andy Jassy to highlight the finding to Washington.
US commerce secretary Howard Lutnick signed a directive addressed to chief executive Dario Amodei, ordering Anthropic to suspend all access to Fable 5, and to its more capable sibling Mythos 5, by any foreign national, whether inside or outside the United States, including the company’s own foreign-born staff.
Since Anthropic could not check nationality of Claude users in real time, it was forced to pull both Fable 5 and Mythos 5 for all users around the world, irrespective of nationality. Fable 5 had been public for three days only, the first model of its class the company had ever opened (beyond a vetted few who had access to Project Galsswing). Every other Claude model stayed online.
What was unique about last Friday’s order is that the Commerce Department reached for the Export Control Reform Act of 2018, a law written for physical goods, and applied it for the first time to software delivered over the internet, demanding an individual licence before any foreign national could use either model. Lawyers do not agree the law even applies. A model reached through an interface is not obviously “exported” in the sense the statute imagines, and until a court says otherwise, Washington can pull the same lever again, on any lab.
Human capital and token capital
Two days after the letter, Microsoft’s chief executive Satya Nadella published an essay on what a company becomes in an AI economy. He split a firm’s worth in two: human capital, the judgement and relationships of its people, and token capital, which he defined as “the firm’s AI capability it builds and owns.” The point, he argued, is the learning loop a company builds on top of whatever model it rents, so its institutional expertise stays in-house even when the model underneath is swapped out.
Mr Nadella called that the real test of control: a company should be able to change its general-purpose model without losing the company-veteran knowledge it has accumulated. Fable 5’s customers had just faced that test the hard way. Anyone who built a product straight onto a single closed model, with no fall back, lost both the capability of said model and all the work they had done, in a matter of hours.
It is important to note that Mr Nadella runs the largest backer of OpenAI and one of the heaviest spenders on the very models he now wants firms to depend on less; the same week, Microsoft told Axios it was weighing an Azure-hosted version of China’s DeepSeek for its Copilot Cowork agent.
Too big to switch off
Canadian prime minister Mark Carney, who has run two central banks, put it in the language of the 2008 global economic crisis. On his way to the Group of Seven (G7) summit in Évian, he likened a market that leans on a few models to the bank linkages that turned a handful of failures into a system-wide crisis, and called for more redundancy and diversification.
The world leaders sitting around the table had a fresh example in front of them: a model their own governments, banks and national companies had been using was switched off by the US government.
French president Emmanuel Macron spent the summit pressing Mr Lutnick on a trusted-partners scheme that would carve allied countries and companies out of the ban. The episode, Mr Macron said, had “clarified the stakes”; a Washington able to revoke access overnight would, he argued, shrink the market for American AI as much as anyone else’s.
Mr Amodei, whose own models were at stake, told the same table to “resist the temptation to splinter,” lobbying for allied access, a few seats from the president who had revoked it. Anthropic and the Commerce Department broke off talks on June 16 without a deal. As this issue goes out, both models remain dark.
Cleared for launch
Rewind to Friday, just before that letter was sent: SpaceX had the opposite kind of afternoon. With its IPO price set at $135 the night before, the Musk behemoth opened on the Nasdaq as SPCX at $150 and closed its first day at $160.95, up about 19 per cent, near $2.1tn (the stock closed at $191.82 yesterday, after crossing $220 earlier on Tuesday). The $75bn raise was the largest initial public offering (IPO) on record, and roughly 30 per cent of the shares went to retail buyers.
The closing bell rang a little over an hour before the Commerce Department’s letter reached Anthropic: one frontier company cleared for launch, another grounded, on the same afternoon.
A note to readers
We held a closed session of Follow The Token at Station F, the Paris innovation and startup hub, for a small group of startup founders and investors. There was good, specific feedback which we are using to fine-tune the 60-minute session, and more importantly to nourish the half-day executive briefing.
We are planning the next in-person session in Barcelona next week, but in the meantime mark your calendars for the first online session that will take place on the afternoon of July 16.
Watch this space.
Wirelessly yours,
Ziad Matar
Co-founder, Tokenando
Editor-in-chief, The Compute
After Hours
This is my first work trip to the French capital, so I decided to do my own version of Emily in Paris (which, in all honesty, I think is a stroke of commercial genius, despite the fact that I couldn’t endure more than two minutes of it). What I did enjoy, however, was a hilarious play called Madame Joconde at Theatre du Marais, which delivered non-stop laughter throughout its 90-minute run.
Sticking to the script though, this issue of The Compute was written while basking in the aroma of freshly baked French bread and enjoying buttery croissants served by a jovial French waiter at Le Nemrod.
I am now hopping on a dott electric bike and zipping across the city to catch the first sessions of VivaTech Day 2… à très bientôt!
References
Anthropic, statement on the directive to suspend Fable 5 and Mythos 5 / https://www.anthropic.com/news/fable-mythos-access
Fortune, “Fix this code”: the three words behind the shutdown / https://fortune.com/2026/06/15/fix-this-code-three-words-behind-us-government-shut-down-anthropic-fable-mythos-ai-models-katie-moussouris-open-letter/
The Decoder, on Satya Nadella’s warning about value concentrating in a few models / https://the-decoder.com/microsoft-ceo-satya-nadella-warns-of-a-small-number-of-ai-systems-capturing-all-the-economic-returns/
Financial Times, Anthropic’s Amodei at the G7 / https://www.ft.com/content/573925dd-6d41-4185-810d-2b848195903d
CNN, SpaceX’s market debut / https://www.cnn.com/2026/06/12/business/live-news/spacex-goes-public-ipo