Permission to compute, Mr President?

Hello friends, and welcome to The Unsophisticated Investor! Brought to you by Scott & Rob, the founders of Shuttle!

If you want to invest alongside the VC funds who've backed breakout companies like Revolut, Asana, JustEat, Bolt, Lets Get Checked, Loom, Runna, Charlotte Tilbury, Deel, Aircall, AngelList, Carta, TransferWise and many more, regardless of your knowledge, network or net worth, then sign up using the link below.

Now, let’s get into it 👇

In 1991, Phil Zimmerman, a world-leading software engineer and cryptographer created PGP, an encryption program for email. Software counted as a munition under American export law, and putting PGP on a server a foreigner could download had left Zimmerman under federal investigation for three years. But Zimmerman’s supporters printed the code of his program into a hardback book and posted it to Europe. A book was speech, could be exported freely, and scanned back into working code at the other end. The law it mocked was already useless: strong encryption was just mathematics, it sat in textbooks and in other countries, and no export rule could put it back in the box.

Four years later Apple built an advertisement mocking the rule. The Power Mac G4, released in 1999, was fast enough to cross the one-gigaflop line that classified a computer as high-performance and subject to export licensing, so it could not be sold freely into around fifty countries. Apple surrounded a desktop with tanks and called it the first personal computer classified as a weapon. The restriction lasted months before the threshold was raised, and a gigaflop is now a trivial amount of computing power that ships in every laptop sold. The fact is that in a digitally global economy, governments can only restrict technological developments for so long.

Three days of Fable  

Last week I updated my desktop installation of Claude to see how Anthropic’s latest and most capable model compared against the Opus series. Sadly for myself, and for every other non-US citizen around the world, this was short lived. At 5:21pm on 12 June the US Commerce Department, in a directive signed by Secretary Howard Lutnick, ordered the company to cut off Fable 5 and its more powerful sibling Mythos 5 to any foreign national anywhere, including Anthropic’s own non-citizen employees. The company could not screen users by passport across its cloud partners in real time, so it shut both models down for everyone. Fable had been public for just three days.

Anthropic’s competitor, Amazon, found that Fable would decline to review code for security flaws but would happily repair the same code when asked to fix it, and repairing it meant locating the flaws first. Katie Moussouris, the security researcher Anthropic brought in to assess the finding, called the trick trivial and said removing Fable did more harm to the defenders who relied on it than to any attacker. Amazon’s chief executive, Andy Jassy, reportedly raised the matter with the White House by phone. It is the first time Washington has switched off a commercial AI system the public was already using. But this isn’t the first clash between the US Government and Anthropic.  In February the Pentagon designated Anthropic a supply-chain risk after the company refused to let its models be used in autonomous weapons and surveillance, a dispute still unresolved in the courts.

Software has been treated as a weapon before

American export controls on strong cryptography in the early 1990s failed because the thing they restricted was already everywhere. PGP spread worldwide within hours, foreign firms wrote their own, and by the time electronic commerce needed encryption the rules were effectively useless. Clinton moved commercial cryptography off the munitions list in 1996, and the controls were ultimately gone by 2000. 

Semiconductors are a more challenging containment case, because there the controls both worked and rebounded. Since 2022 the United States has denied China the most advanced chipmaking equipment, the extreme-ultraviolet machines made only by the Dutch firm ASML, and it has genuinely held back Chinese, so far, in competing with the US incumbents. But in 2023 Huawei shipped a 7-nanometre phone chip built by SMIC, which has no such machine, on older equipment pushed past its limits, having rebuilt its supply chain around domestic firms after losing access to TSMC. The Chinese lab DeepSeek now runs inference on Huawei's Ascend chips, even as it has leaned back on Nvidia for frontier training. The controls slowed China down but they also pushed it to build at home the capacity it had previously been content to import. 

Why Europe spent the week panicking

For a decade the obvious choice for a European bank or hospital buying AI was an American model on an American cloud. It performed best, it cost least, and the risk that Washington might one day interfere felt abstract. Last week’s actions by the White House will only encourage European leaders, both at political and commercial levels, to make the case for building and using European models.

France’s former prime minister, Édouard Philippe, called AI infrastructure as basic as electricity and pointed out that infrastructure someone else owns can be switched off by someone else. Politicians throughout Europe have been, and are, making the case that Europe can’t just regulate AI, but must compete by building its own. One Finnish politician said Europe could not keep building on foundations a foreign government could remove overnight. 

The dependence is real and measurable. The European Commission reckons the bloc relies on non-EU providers for more than 80% of its digital products, services and infrastructure, and over 70% of its cloud market sits with three US hyperscalers. Other estimates suggest that the United States and China between them hold about 90% of the world’s AI computing power. Days before the shutdown, on 3 June, the Commission had published a Technological Sovereignty Package, with a Chips Act sequel and a plan to triple Europe’s data-centre capacity, putting the investment need at roughly €422 billion across the package (about €120bn chips, €200bn data centres, €100bn cloud and AI, €2bn open source), with no clear account of who pays. 

Washington’s counter-offer

Within days the administration began trying to turn the ban into something more workable. American and European officials have started discussing a trusted-partner scheme that would grant close allies privileged access to the most advanced models, an idea Lutnick floated with European diplomats at the G7 and that leaders were to take up with the heads of Anthropic and OpenAI on the sidelines of the meetings. To its supporters it solves the problem, keeping allied hospitals and banks supplied while shutting out genuine adversaries. To Europe it looks like the same dependence in better clothes, because access that Washington grants is access Washington can take back. The Commission’s technology chief, Henna Virkkunen, warned against treating partners like the EU as security risks and said Europe should not be so reliant on a single company or country, while asking the Americans to specify the danger they were worried about.

The objection was not only European. The SIIA, the American trade body that represents Apple, Amazon and Google, broke with its usual caution to call the move unprecedented and warn that arbitrary power over frontier models would damage the global uptake of American AI. 

How does Europe’s scene compare?

Mistral, the only European company in the frontier-model conversation, is in talks to raise about €3 billion at a roughly €20 billion valuation, nearly double its €11.7bn Series C last September, selling open-weight models that customers can run on their own hardware. In March the London infrastructure company Nscale raised $2 billion at a $14.6 billion valuation, the largest funding round of its type in European history, to build data centres. The same month the Stockholm legal-AI firm Legora, three years old, raised $550 million at $5.55 billion, and the Cambridge self-driving company Wayve had raised $1.2 billion in February at $8.6 billion.

Set side by side, the list flatters Europe less than it first looks. The money is flowing generously into the parts of the stack that sit on top of a model, the data centres, the defence systems, the applications, and into precisely one company building a frontier model of its own. The layer the Fable episode exposed, the frontier itself, is the thinnest part of Europe’s holdings. An inherent issue in Europe is the ability to compete at the funding level against the mega-VC funds of the US. For example, Accel led Legora, Dragoneer and Lightspeed are leading Helsing, and Nvidia is routinely on the cap table of many of Europe’s largest rounds.

What the unsophisticated investor could take from this

Mistral at 23 billion is, to a large degree, a bet that the jurisdiction risk on display this week is permanent, that European buyers will keep paying to own their AI outright rather than rent a better and cheaper American one. That bet looks shrewd in a week when an American model is switched off by its own government, and a good deal less shrewd in the week Fable returns and outperforms the home-grown alternative. The distance between wanting sovereign AI and having it is still significant. Europe lacks the compute, the cheap energy and the depth of frontier labs, and the deep liquidity of US public and private markets. The likeliest immediate beneficiary of pulling an American model is not a European champion but whatever cannot be pulled, which increasingly means open-weight models, a fair share of them Chinese.

For years American AI was the default because American dominance was complete enough that the government let its companies sell to everyone. The Fable order says that is no longer reliably the case, and the trusted-partner club is Washington’s attempt to keep allies inside the tent on its terms before they drift off and build their own. For Anthropic the timing is cruel: a company about to ask the public markets to value its lead has just shown that the lead can be revoked by its own government in an afternoon.

Zimmermann’s supporters understood the mechanics in 1995. You can take down a server, prosecute an engineer and force a company to disable its best product, and none of it changes what people now know can be built. It changes only where they decide to build it, on whose machines and under whose flag. That decision is being made right now, and not in Washington’s favour.

What we’ve been working on at Shuttle

• Lining up an end-of-month Drop and scanning our co-investor network for what's in market 🚀

• Rolling out new investor dashboards with a weekly AI briefing that watches your portfolio and surfaces what matters 🤖

• Locking in our new payment partner: dedicated IBANs and instant KYC, then the full upgrade goes live ⚡

The Unsophisticated Investor is brought to you by Scott & Rob, the founders of Shuttle. We’re both sick of private markets being a playground exclusive to the ultra-wealthy so we started a company to challenge the status-quo. Shuttle’s singular focus is to unlock private markets for Millennial and Gen Z tech professionals and help them build wealth through the highest performing private market opportunities.

Scott & Rob
Shuttle Co-Founders