AI Actually
Issue No. 21 — Wednesday, June 24, 2026
This week a rocket company rented out its supercomputer for $6.3 billion, a Nobel Prize winner quit his job, and the company that makes Claude floated the idea of asking for your ID. Four different stories, one quiet theme: the AI boom has turned into a fight over who owns the things you can’t make more of — the chips, the people, the locks, and, as of this week, possibly you. Let’s get into it.
A rocket company is now renting out supercomputers
SpaceX — yes, the one that makes rockets — signed a deal worth up to $6.3 billion to let a startup called Reflection AI borrow time on its supercomputer, nicknamed Colossus. Reflection gets to run its training on rows of Nvidia’s newest chips. SpaceX gets a very large check.
The strange part isn’t the number. It’s the business model. SpaceX doesn’t build AI models. It’s renting out raw computing power the way a landlord rents apartments — and AI companies are lining up at the door, because the single scarcest thing in this entire industry is “enough fast computers to train a modern model.”
We watched the first version of this back in No. 17, when Google was reportedly paying SpaceX close to a billion dollars a month for the same kind of access. The update this week: now it’s a startup writing the rent check, not a tech giant.
Why it matters: The assumed advantage in AI was always supposed to be brains — the best researchers, the cleverest model. It’s turning out to be hardware: who has the most computers, and who’ll let you borrow them. When the company with the most spare supercomputer time is the one that builds rockets, you’re not really looking at a software industry anymore. You’re looking at a real-estate market for electricity.
The AI talent war just claimed a Nobel Prize
John Jumper won the 2024 Nobel Prize in Chemistry for AlphaFold — the AI that predicts the shape of proteins, now used by more than two million scientists to speed up work on vaccines, cancer drugs, and antibiotics. After nearly nine years at Google DeepMind, he’s leaving to join Anthropic, the maker of Claude. He’s taking a break first, which, fair.
Here’s the detail that says the most. Before he left, Jumper had reportedly been moved off science and onto coding tools — the unglamorous product DeepMind has struggled to sell to businesses. A freshly minted Nobel laureate, reassigned to help ship a developer feature. You can see how that conversation might have gone.
He wasn’t the only one out the door. The same week, Noam Shazeer — one of the people who literally co-wrote the research paper that kicked off this whole AI boom — left DeepMind for OpenAI. Google had paid a reported $2.7 billion two years ago partly to get him back in the first place.
Why it matters: DeepMind was famously the place people didn’t leave. Two of its biggest names walked out in a single week — one of them carrying a Nobel Prize with the company’s name on it. If prestige and billions of dollars can’t keep your stars, then money isn’t the lever anymore, and every rival lab now knows exactly where to go shopping.
AI is now picking the locks and selling the deadbolts
OpenAI expanded a program this week called Daybreak, its push to aim AI at cybersecurity. The pitch: its models have gotten so good at finding holes in software that humans can’t patch them all fast enough — so OpenAI is now shipping tools to do the fixing, too. That includes a restricted, souped-up model called GPT-5.5-Cyber and an effort (with cURL, Python, and Go already signed on) to repair the free, invisible software the entire internet quietly runs on.
Notice the flip. For years, finding the security bug was the hard, expensive part. AI made finding bugs easy. The new bottleneck is fixing them all before someone holding the same AI gets there first.
Access is locked down — only “verified defenders” get the powerful version, which raises the obvious question of who gets to decide who counts as a defender. The timing is its own story: this lands while Anthropic’s own cyber-capable model sits frozen, pulled and stuck in government limbo since the spring (No. 18). OpenAI walked straight through the door Anthropic left open.
Why it matters: The same technology that breaks into systems is the technology that defends them — it’s one tool wearing two hats. Whoever controls the powerful, locked-down “good guy” version ends up holding an enormous amount of trust and an enormous amount of leverage. We’re quietly handing the keys to the internet’s locks to a handful of companies, and they’re the ones deciding who’s allowed a copy.
Claude might ask to see your ID
Anthropic said that starting July 8, it may ask some Claude users to verify who they are with a government-issued ID (handled by an outside company called Persona). It says this will only hit a small group — accounts that got flagged for something but weren’t banned outright — and it hasn’t spelled out exactly what trips the check.
Small policy, but a notable line to cross. Talking to a chatbot has felt, up to now, like about the most anonymous thing you can do online — closer to muttering to yourself than to opening a bank account. “Please upload your driver’s license” changes the texture of that a little.
Why it matters: The “type anything to a faceless AI, no questions asked” era is quietly ending. Whether that reads as reassuring or unsettling mostly depends on which end of the flag you picture yourself on. Either way, “show me your papers” is a strange sentence to hear from a robot.
Safe to ignore this week
The “scary” new Chinese open model. It’s genuinely good. It’s also the thing we spent all of Sunday’s issue on (No. 20). Consider yourself caught up.
Alibaba’s video generator climbing to No. 2 globally. A Chinese model is very good at a thing. We are now reporting this roughly twice a week.
A “claude-sonnet-5” label spotted on a website. A slightly newer model name appeared somewhere. It is not a model you can use, an announcement, or even a confirmed fact. It is a label.
Tencent putting an assistant inside WeChat. Useful if you live in China, invisible if you don’t.
Anthropic bringing its Cowork tool to phones. You’ll soon be able to assign your AI homework from the bus. Hold your applause.
Until Sunday
Four stories, one through-line: this week the AI industry spent its energy building fences around the things it can’t simply manufacture more of — computers, brilliant people, security, and now, maybe, your ID. Sunday we’ll slow down and take one of these apart properly over coffee.
As always — if something here confused you, or you think I’ve got it wrong, just reply to this email. It comes straight to me.
See you Sunday.
—
AI Actually is a twice-weekly newsletter explaining what’s actually happening in AI, in plain language, for people who’d rather not read technical papers on a Wednesday.