The cost of control

THE COMPUTE / BY TOKENANDO Issue VIII / June 25, 2026

Jun 25, 2026

I am back in Barcelona this week, where the heat is at least less biblical than in Paris, which warmed up over last week and is now simmering past 40 degrees. But something else has been building since then, spreading into the mainstream so fast that by Tuesday it was the centrepiece of the dinner table.

We did not just make small talk about the weather, the World Cup or artificial intelligence in general; people who could not, a couple of years ago, have told a chatbot from a teapot (including yours truly) now hold firm views on who owns these models, who can switch them off and who reads our data once a machine has it. That is sovereignty: control of the AI, the data and the computing power that a country, a company or practically each of us now depends on.

The gap between AI familiarity and true understanding is getting wider. Every week, new buzzwords, terms and concepts seep into ordinary conversation faster than the last was grasped.

On the global stage, sovereignty moved from talk to action on three fronts.

/ The fight over who is even allowed to use the most powerful models turned concrete: a legal-tech company sued the US government over losing access to Fable 5, while Washington accused Dutch firm ASML of letting one of the machines that make the world’s most advanced chips reach China.

/ AI assistants stopped waiting to be accessed in their own app or window, and moved into the apps we already use all day.

/ And underneath it all, the chips that AI runs on drew fresh challengers while the memory beside them ran short, sending the first price rises to everyday devices.

Now, read on…

The kill switch

The right to use the most powerful AI models has become something a government can grant and take away, one company at a time or as a blanket ruling. This week partly revolved around who has access, and who does not.

Ten days before the Five Eyes intelligence alliance (the US, UK, Canada, Australia and New Zealand) warned company boards that AI-driven cyber attacks are “months, not years” away, Amazon had reported to the US government that Anthropic’s latest public model Fable 5 could be coaxed into searching for and finding software flaws. Washington had not waited for the alliance: on June 12 it imposed export controls (national-security rules that limit who outside the US may use a technology), and Anthropic pulled both Fable and the underlying Mythos out of worldwide circulation within about 90 minutes of being notified.

Earlier this year Anthropic had given roughly 150 organisations worldwide early access to Mythos, its most capable model (on which Fable 5 is based), through a vetting programme called Project Glasswing. The Washington Post reported that, a couple of weeks ago, the White House grew alarmed after spotting a South Korean telecoms company it suspected of China ties on that list, and asked Anthropic to cut the firm off (Wired indicated the firm is SK Telecom, which denies any China connection). Meanwhile, Bloomberg reported that two American security firms, Dragos and Cisco, kept their Mythos preview, while ENISA, the European Union’s cybersecurity agency, saw its invitation withdrawn. Legion, a US legal-software firm, sued the government on June 23 over its lost Fable access, calling the harm “immediate, irreparable and existential”.

Washington doubled down: commerce secretary Howard Lutnick accused ASML, the Dutch company that builds the only machines able to etch the most advanced chips, of letting one reach China (ASML denies it). Abu Dhabi, on the other hand, reached for a completely different lever: its state-backed fund MGX raised close to $50bn from outside investors to buy into frontier AI directly, putting capital beside law as a way to have some control over what comes next.

The @ team

The competition in enterprise AI is shifting from which model(s) a company uses to where that model actually works. On June 23, Anthropic launched Claude Tag, an always-on version of its assistant that runs inside Slack, the workplace messaging app. A team member types @Claude into a channel, hands over a task, and the assistant works through it in stages and reports back in the thread, using the tools, files and code the team has connected. Claude Tag works in a channel where the whole team can see it, direct it and pick up where a colleague left off, and it carries the channel’s history rather than starting from fresh each time.

Tencent is making a parallel bet in the consumer market. It has begun testing Xiaowei, an agent (software that performs multi-step tasks for you, rather than only answering) inside WeChat, the messaging-and-payments app that around 1.4 billion people in China use. A typed or spoken instruction can set off a payment, a ride or a message through the app’s mini-programmes.

What separates these developments from the assistants people already know, such as Meta’s AI in WhatsApp or Grok in X, is what they are built to do. Instead of merely asking a question and getting back a reply, an image or a draft, Claude Tag is built to act, taking a delegated job and carrying it out across a company’s own systems, and it is shared rather than kept private. As for Xiaowei, it does the same in daily life: a request comes back as a completed task inside WeChat, an outcome rather than an answer the user must act on.

In both cases the aim is the same: become the assistant inside the place where people already work, produce and spend, rather than a separate app they have to open.

Total Recall

While the market for AI chips is opening up to competition, the memory that runs alongside them is getting more expensive and harder to get. Both power the physical layer under every model a business runs, and both impact the cost of building on AI… and more.

Micron, one of three companies that make the fast memory that AI servers depend on, reported net income of $28.2bn for its latest quarter, roughly 14 times a year earlier (not a typo… I promise), and said its memory is sold out till the end of 2026. The same day, South Korea’s SK Hynix filed to raise about $29.4bn by listing its shares in New York to pay for more capacity; if it prices at the top of its range on July 10, it would be the largest listing of its kind on record, ahead of the $21.8bn Alibaba raised in 2014. The rush to buy that high-end AI memory has left ordinary memory in short supply, and today it reached consumers: Apple raised prices on its Macs and iPads by up to $300, blaming the memory shortage, with chief executive Tim Cook calling it a “hundred-year flood”.

The other half of the story was taking place in the inference space (the work of running a trained model to answer requests, as distinct from training it). Cerebras, which was freshly listed on the stock market, beat expectations on sales, with quarterly revenue of $193.4m, but missed on profit and guided its margins sharply lower, and its shares fell about 14 per cent even as it confirmed a deal worth more than $20bn with OpenAI.

OpenAI and Broadcom unveiled an inference chip of their own, Jalapeño, taking it from design to manufacturing in nine months, which OpenAI called the fastest such effort in advanced chips thanks to using AI. Qualcomm, for years a smartphone-chip company, named Meta as the first big customer for its new data-centre chips and forecast more than $15bn in data-centre sales by 2029, up from roughly $300m this year.

For now these are early challengers rather than rivals: none matches Nvidia at scale, and Cerebras’s thin margins show how difficult the economics still are. After two years effectively locked to a single supplier of AI chips, buyers will at least have somewhere else to look soon. In memory there is no such escape: the three makers who supply it are all sold out. Even Apple, the company Mr Cook built into a master of supply-chain management, was not immune.

Wirelessly yours,

Ziad Matar
Co-founder, Tokenando
Editor-in-chief, The Compute

After Hours

Barcelona’s night was lit up with beach bonfires and the echoes of street fireworks as Sant Joan came on Tuesday night, while parts of this issue were being written. Keeping with the local theme, we have been watching Ravalear, a series about a century-old, family-owned restaurant in Barcelona’s El Raval neighbourhood, resisting corporate greed and gentrification. Lined up for the weekend are two films inspired from the VivaTech stage: Tetris, which tells the story of how the game went global, and General Magic, the documentary about the origins of the smartphone.

The soundtrack to the week was a Mediterranean summer playlist, as it should be: warm evenings, a salty breeze, and an ice-cold beer with the occasional negroni sbagliato or mezcal paloma.