May 2026: The Month AI Went Public

Transcript

Alex: Picture this. A company nobody outside the chip world could really value walks onto the Nasdaq, and within minutes the market says it's worth ninety-five billion dollars. The same month, an AI quietly finds ten thousand real, exploitable holes in the world's software — and then optimises a drug reaction chemists have been stuck on for decades, in an actual lab.

Sam: Hang on — ten thousand? And a lab? Not a benchmark, an actual lab with beakers?

Alex: An actual lab. And here's the line that ties it all together: this was the month all the parts of AI that used to be hidden suddenly got a public, auditable number stuck to them. And almost every one of those numbers came in bigger than anyone expected.

Sam: Okay. That's a hook.

Alex: Welcome to Dan's AI Intel — the show where we try to make sense of the fastest, most consequential shift any of us is going to live through. I'm Alex.

Sam: And I'm Sam. And this one's a bit different — it's our monthly briefing. We zoom out and ask: across all of it — the labs, the chips, the money, the politics — what actually happened in May 2026, and what does it mean.

Alex: The whole reason this show exists is that the field moves faster than any one person can track. The knowledge horizon — how far ahead you can actually see — is brutally short. So once a month we stitch the chaos into one picture you can carry around.

Sam: So give me the picture. What's the through-line for May?

Alex: One sentence: this was the month the abstractions turned concrete. Capability stopped being a leaderboard score and started finding real bugs and making a real discovery. Capital stopped being a private round and became a listed share price. And the electricity all of it runs on stopped being a line item and turned into a public fight over who pays for the grid.

Sam: And if you had to name the single biggest story?

Alex: AI went public. Literally. A chip company called Cerebras had a blockbuster stock-market debut, and that cracked open a window that OpenAI, Anthropic and a freshly-merged SpaceX-and-xAI are all now lining up to climb through. For the first time, the market got to put a price on the AI build-out directly, in daylight.

Sam: So we're not doing one story tonight. We're doing the month.

Alex: We're going lens by lens. Five of them. We start with the big ideas — the conceptual shifts. Then the frontier labs and what they shipped. Then the infrastructure — the chips and the power. Then the business and the money. And we finish on governance and society — the politics, the law, the jobs. Major stories we'll dig into; the smaller ones we'll flag as we pass.

Sam: Here's what I want by the end: which of these numbers should actually change how I think about where this is going. Not just "wow, big number."

Alex: That's exactly the right question, and it's the one we'll keep coming back to. If you've been enjoying the show, follow us on Spotify or Apple Podcasts so the next briefing lands automatically. Right — let's start at the top, with the big ideas. So, big ideas and horizons. This is the lens where we step back from "who shipped what" and ask "what did we actually learn about what these things can do." And the most important thing that happened in May wasn't a model release at all. It was a result.

Sam: The lab thing. Walk me through it, because "AI made a discovery" gets thrown around a lot and usually means "it summarised a paper."

Alex: Right, and that's exactly the bar this clears. OpenAI, working with a small Polish chemistry startup called Molecule.one, took one of their models and ran it as a near-autonomous agent across ten thousand and eighty real wet-lab reactions.

Sam: Ten thousand and eighty. That's oddly specific.

Alex: It is, and the specificity is the point — these were physical experiments, not simulations. And out of that, the system produced a genuine, independently confirmed improvement to a reaction that medicinal chemists have wrestled with for decades.

Sam: Okay, you're going to have to make the chemistry land for me, because "Chan–Lam coupling" means nothing to me.

Alex: Fair. So think of a drug molecule like a piece of furniture you're assembling. There's one particular joint — a join between a carbon atom and a nitrogen atom — that turns up everywhere in modern medicine. The standard way to make that joint is a method called Chan–Lam coupling. And there's a specific version of it, on a chemical group called a primary sulfonamide, that just... works badly. Low yields. You put your ingredients in and most of it doesn't react.

Sam: And that joint matters because...?

Alex: Because that exact group shows up in more than ninety FDA-approved drugs — cancer drugs, antibiotics, heart drugs. So it's a known choke-point. Chemists hit it constantly in early drug discovery and just live with the bad yields. The AI system found that adding a particular oxidant — something called TEMPO, which chemists had not expected to help — was the unlock.

Sam: TEMPO. Why would chemists not expect that to help? What makes it surprising?

Alex: I'll keep this at the right altitude — the honest answer is that chemistry has a lot of received wisdom about what plays nicely with what, and TEMPO sits in a category you'd assume would interfere rather than help. So no human would have prioritised testing it. The system didn't carry that assumption, so it tried things a trained chemist would have skipped on instinct. Sometimes not knowing the folklore is an advantage.

Sam: So it found a trick the experts had missed — partly because it didn't know it was supposed to be a dead end.

Alex: That's a sharp way to put it. The lack of bias is a feature here.

Sam: But — devil's advocate — couldn't a grad student grinding through ten thousand experiments have stumbled on the same thing?

Alex: Possibly, over years. And that's actually the key. For three years the standard comeback to all the AI hype was: these things just remix what humans already know, they can't generate genuinely new knowledge about the physical world. A reproducible, externally-reviewed improvement to a real reaction is a direct counter-example to that. The news here is the discovery, not the model.

Sam: So why does that change how I should think about it?

Alex: Because it points at where the money goes next. If these systems are research instruments, not just chatbots, then the value isn't in the SaaS app — it's in the multi-trillion-dollar industries where a single validated discovery is worth more than an entire software category. Pharma. Materials. Energy. For years "AI for science" was an essay. In May it showed up with a yield number attached.

Sam: There's a word you keep leaning on. Near-autonomous. What's carrying the weight there?

Alex: Everything, honestly. This wasn't a chatbot suggesting an idea that a chemist then went and tested. It was a system running a closed loop by itself — propose a condition, run the reaction, read the result, propose the next one — thousands of times, with humans supervising rather than driving.

Sam: So it's less "the AI is a genius chemist" and more "the AI never gets tired of trying the next thing."

Alex: That's exactly it, and it's the lesson that generalises way past chemistry. The value didn't come from the model knowing more than a professor. It came from it being allowed to act, in a real environment, fast, many times over, without waiting for a human at every step.

Sam: So anywhere you've got a clear measure of success and a cheap way to take lots of shots at it —

Alex: — drug screens, materials, code, ad copy, logistics — an agent that can run the loop itself starts to out-search human teams. Not because it's smarter per step, but because it never stops taking the next step.

Sam: Okay, but let me push on that, because it sounds almost too neat. If it's just "take lots of shots," why hasn't brute force already solved everything? Computers have been fast for ages.

Alex: Great question, and the answer is the missing ingredient was judgement between the shots. Brute force tries everything blindly. What's new is a system that looks at the result of shot number four hundred and uses it to pick a smarter shot number four hundred and one. It's the difference between a slot machine and a chess player — both take lots of moves, but only one learns from each one.

Sam: Ah. So it's not "try everything," it's "try everything, getting warmer each time."

Alex: Exactly. And that raises this slightly unsettling question: how many of the world's hard problems are secretly just search problems — a clear goal, cheap repeated attempts, judgement between them — sitting there waiting for a tireless searcher who actually learns as it goes?

Sam: Park that, because I think it comes back. There's a second big-idea shift you mentioned — something about the conversation itself changing.

Alex: Yeah. Underneath all the product noise, the serious conversation moved. For a couple of years the question among the people who think hardest about this was "how capable are these models?" In May the operative question became "how much autonomy should we hand them?"

Sam: Why is swapping one word for another a big deal?

Alex: Because they're fundamentally different kinds of question. Capability is something you measure — you run a benchmark, you get a score. Agency is something you grant. And a grant is a governance decision, not a research result.

Sam: Oh. So one's a fact about the model and the other's a choice we make.

Alex: Exactly. And here's why it matters for the whole episode: that single reframing connects three stories that otherwise look unrelated. The AI finding ten thousand vulnerabilities. Enterprise agents crossing into real budgets. The labs walking back their job-loss predictions. They're all secretly arguments about the same variable — how much unsupervised latitude an AI gets inside a high-stakes loop.

Sam: So when you frame it as "intelligence," the policy lever is, what, a test score threshold.

Alex: Right. And when you frame it as "agency," the lever becomes liability, oversight, and the design of the loop the model sits inside. Which is a completely different regulatory conversation. Expect the next year of safety research and regulation to be organised around agency, not raw capability.

Sam: Before we move on — anything smaller worth a mention in big ideas?

Alex: Two quick ones. First, the labs revised their own prophecies — both Sam Altman and Dario Amodei publicly softened earlier sweeping predictions, which is a real tell. We'll get into that properly when we hit jobs at the end. And second, Amodei's essay calling this the "adolescence of technology" stayed a reference point all month — the idea that we're in a dangerous in-between stage to be survived, not a finish line to cross.

Sam: So to recap the big ideas before we move: AI made a real discovery in a real lab, which kills the "it only remixes" argument. And the whole debate quietly shifted from how smart these things are to how much we should let them do on their own.

Alex: Nicely done. And hold onto that second one especially, because the agency question is about to get very concrete. Right — let's move from the ideas to the labs themselves, and what they actually shipped. Frontier labs, models and products. This is the lens where we look at the three big players — Google, OpenAI, Anthropic — and what they actually put out in May. And the fascinating thing is how differently each one played it.

Sam: Different how?

Alex: Google went big and loud at its I/O conference. OpenAI went silent and invisible. And Anthropic changed what it's even selling. Three completely different strategies in one month.

Sam: Start me with Google. What did they ship?

Alex: Two things. The headline was a model called Gemini 3.5 Flash. Now, "Flash" is Google's cheap-and-fast tier — and the news is that this cheap, fast tier now beats their own previous top-end "Pro" model on hard coding and agentic tasks.

Sam: Wait, the budget model beats last year's premium model? That seems like it should be a bigger deal than it sounds.

Alex: It's quietly one of the most important trends of the year. When the cheap tier gets good enough to be your default workhorse, the economics of everything built on top of it shift. But honestly, the more strategically interesting reveal was the second thing: a model called Gemini Omni — billed as a generative world model.

Sam: And a world model is...?

Alex: Okay, so a normal language model learns the statistics of text — given these words, what word comes next. A world model tries to learn the dynamics of an environment — how things actually move and interact. So instead of "what word follows this word," it's "if I push this object, what happens next."

Sam: So one's learned to talk and the other's trying to learn how the world behaves.

Alex: That's a good way to put it. And the reason Google planting its flag there matters: it signals the frontier race is no longer only about building bigger language models. This is the "what comes after the LLM" bet, made on the biggest stage Google has.

Sam: Did they ship anything you can actually use, or is it all "billed as"?

Alex: They did. They wired in something called Managed Agents — one API call spins up a sandboxed Linux environment that an agent can plan and act inside — and Science Skills, hooked into thirty-plus life-science databases. It's a coherent push to be the platform for doing things, not just chatting. As we flagged in last month's frontier coverage, Google's real moat is that it's vertically integrated — it owns everything from the chips, the TPUs, up through the research to the product. I/O was that moat on display.

Sam: Before we leave Google — one thing nags me. They've got the loud model and the world model and the agents. Is that a coherent strategy, or are they just throwing everything at the wall?

Alex: That's the right scepticism, and I'd say it's coherent precisely because of the vertical integration. Think of it like a restaurant that owns its own farm. Most AI companies rent everything — they buy chips from Nvidia, rent data centres, license research. Google grows its own ingredients: its own TPU chips, its own research labs, its own products on top. So when it ships a world model and an agent platform and a cheap fast model in one week, it's not scatter — it's the same supply chain firing at every level at once.

Sam: So the moat isn't any one model. It's that they own the whole stack.

Alex: That's the moat. And it's why I'd watch Google differently than the others — their advantage compounds quietly. Right. Now contrast that with OpenAI, because you said they went silent. This is my favourite story of the lens, because it's the opposite of a stadium keynote. On May 5th, OpenAI made a model called GPT-5.5 Instant the default in ChatGPT, for every single user, including the free tier. Almost no fanfare.

Sam: So they just... swapped the engine while everyone was driving and didn't tell them.

Alex: Pretty much. And the gains are pointed at exactly the right problem. The new default makes 52.5% fewer hallucinated claims than the one it replaced — on high-stakes prompts, things like medicine, law, finance. Plus it's tighter, about 30% fewer words, and it remembers your past chats and files better.

Sam: Fifty-two percent fewer made-up facts on medical and legal questions. That's the bit that matters, surely.

Alex: That's the whole so-what. A default model serving hundreds of millions of people every day is the single largest lever in consumer AI. And halving the hallucinations on exactly the questions where a wrong answer does real damage — that's the unglamorous reliability gain that turns a toy into infrastructure.

Sam: Why does "default" do so much work in that sentence? People can pick a different model if they want.

Alex: They can, but almost nobody does. Think about how few people ever change the default search engine, or the default keyboard on their phone. The default is what the overwhelming majority experience, forever. So when OpenAI changes the default, it isn't offering a better option — it's silently upgrading the actual experience of nearly everyone, including hundreds of millions on the free tier who'd never go looking for a setting.

Sam: So the quiet move reaches more people than the loud one ever could.

Alex: By a mile. That's why I called it the most consequential move nobody saw.

Sam: And the contrast with Google is basically the entire strategy in one move.

Alex: It is. Google sells the spectacle of new capabilities. OpenAI quietly upgrades the thing people already use every day. And here's the thing — both are right. The gap between "the model you demo" and "the model you actually ship to everyone" is now a deliberate product decision, not an accident.

Sam: Okay. Anthropic. You said they changed what they're selling.

Alex: This is the big one. This connects straight back to the agency idea. Under a program called Project Glasswing, Anthropic reported that its Claude "Mythos" system had identified more than ten thousand high- or critical-severity vulnerabilities in widely-used software.

Sam: This is the ten thousand from the cold open. Give me the detail, because a number that round makes me suspicious.

Alex: Good instinct, and the detail is what makes it real. They scanned over a thousand open-source projects and flagged 23,019 issues. Of those, 6,202 were rated high or critical. Then — and this is the part that matters — six independent security firms reviewed 1,752 of them and validated over 90% as genuine. True positives.

Sam: So it's not the AI marking its own homework.

Alex: Right, outside firms checked. And among them, Mythos autonomously found and exploited a seventeen-year-old flaw in FreeBSD — that's a widely-used operating system — that hands an unauthenticated attacker full root access. It also found a certificate-forgery bug in a crypto library called wolfSSL that runs on billions of devices.

Sam: Let me make sure I understand what "found and exploited" means, because those feel like different things. Finding a hole versus actually climbing through it.

Alex: They are different, and the difference is the scary part. Finding a flaw is saying "this lock looks pickable." Exploiting it is actually picking the lock and walking into the building. Mythos didn't just flag the FreeBSD bug as suspicious — it went the whole way and demonstrated it could take full control of the machine. That second step is what used to require a rare human expert.

Sam: So it's not a smoke alarm saying "might be a fire." It's the thing that proves it can burn the house down.

Alex: Exactly that. And it did the proving itself, autonomously, at scale.

Sam: Seventeen years that bug was sitting there. So why is this good news and bad news at the same time? I can feel there's a catch.

Alex: There's a big catch, and it's why this is the story to watch all year. It cuts two ways. Defensively, this is the strongest evidence yet that AI can do real, validated security work at a scale no human team could match — a genuine advantage for defenders who deploy it. Offensively, the exact same capability in the wrong hands is an automated zero-day factory.

Sam: A zero-day being...?

Alex: A vulnerability nobody's patched yet, because nobody knew it was there. The most dangerous kind, because there's no defence ready. Now think about what changed. The world's software is stuffed with these latent bugs — the FreeBSD one sat there seventeen years. Until now, finding them was bottlenecked by how few expert human researchers exist. That scarcity was acting like an accidental safety margin.

Sam: Oh, I see. The bugs were always there. It was just hard to find them, so mostly nobody did.

Alex: Exactly. And an AI that can find and weaponise them at machine scale removes that margin — for everyone, at the same moment. So picture two teams in a race against the same pile of unpatched bugs. The defenders' edge is they can scan their own code first and patch before anyone else looks. The attackers' edge is they only need the bugs nobody got to yet.

Sam: So whoever moves faster wins the window.

Alex: That's the whole game. And it's exactly why Anthropic ran this as a consortium — with AWS, Apple, Google, Microsoft, Cisco, the Linux Foundation, JPMorgan — instead of a press release. The point was to get the patches shipped before the same technique leaks into the wild. May was the starting gun on a global patch-or-be-pwned sprint, and the clock runs the same for both sides.

Sam: And notice — that's your agency point made physical. The question isn't "how smart is Claude." It's "how much should we let it loose inside critical systems."

Alex: You're getting ahead of me, and you're completely right. Capability stopped being a number on a benchmark and became ten thousand actual security holes. One more frontier story before we move on, and it's a people story.

Sam: Go on.

Alex: On May 19th, Andrej Karpathy — OpenAI co-founder, former Tesla AI lead, one of the most-followed teachers in the whole field — joined Anthropic's pre-training team.

Sam: Why do I care where one researcher works?

Alex: Because talent moves at this level are a leading indicator — they tell you where the most interesting work is, before the results show up. Pre-training is the expensive, foundational work that sets a model's core capabilities. And Karpathy choosing to do that at Anthropic — rather than go back to OpenAI, where he'd worked twice before — is a vote on where the frontier action is.

Sam: And there's a twist, isn't there. What's he actually working on?

Alex: Using Claude to accelerate pre-training research itself. So — using Claude to build the next Claude.

Sam: Okay, that's the bit that makes the hair stand up a little. Why does "AI helping build AI" matter more than it sounds?

Alex: Because of what it does to the speed. Right now, progress is gated by how fast a few thousand brilliant human researchers can think. If the AI itself becomes a useful researcher — generating ideas, running experiments, spotting what works — you start to loosen that human bottleneck. Each generation helps build a slightly better next generation, a bit faster.

Sam: Which is the flywheel everyone's either excited or terrified about.

Alex: Both, usually at once. It's the self-improvement loop the whole field is racing toward — and the fact that it's now being staffed by one of the most respected names in the business is the tell that it's stopped being a thought experiment and become an actual research programme. And notice — that loop is the agency idea again, pointed inward. We're not just asking how much we let AI act in the world; we're asking how much we let it act on itself.

Sam: Okay. Quick mentions before we leave the labs?

Alex: Three worth flagging. Chinese open-weight coding models reached Western parity at roughly a third of the cost — several now beat the top US models on a coding benchmark called SWE-Bench Pro. That's a huge story, but it's really a money-and-geopolitics story, so we'll do it properly in the business lens. Google confirmed a Gemini 3.5 Pro in internal use for the following month — keeping the premium tier moving in lockstep. And Microsoft previewed its own MAI model family, including its first reasoning model — which is the product side of Microsoft loosening its grip on OpenAI, again, business lens.

Sam: So the thing that stuck with me from the labs: three players, three totally different moves — Google selling spectacle, OpenAI silently upgrading the thing a billion people use, and Anthropic turning Claude from "smartest chatbot" into "autonomous security researcher that found ten thousand real flaws."

Alex: That's the lens in one breath. And every one of those moves needs two things to actually happen: chips, and power. So that's exactly where we go next. Right — let's move to infrastructure. Infrastructure and compute. This is the physical layer — the chips everything runs on, and the electricity those chips drink. And May gave us one of the strangest market moments of the year.

Sam: Strange how?

Alex: Nvidia — the company that makes the chips basically all of this runs on — posted a record quarter. Record revenue, eighty-one point six billion dollars, up 85% on the year. Data-centre revenue alone was a record seventy-five point two billion. By any normal measure, a company firing on every cylinder.

Sam: That's an absurd amount of money. So the stock went up, obviously.

Alex: The stock fell.

Sam: ...Wait, what? They beat expectations and the stock went down?

Alex: It did. And that, to me, is the single most important market signal of the month.

Sam: Okay, that genuinely doesn't make sense to me. Explain it like I'm not a markets person.

Alex: Here's the way to think about it. When a company beats expectations and the stock still drops, the market has stopped asking one question and started asking a different one. It's no longer asking "can they grow?" — clearly they can. It's asking "who ultimately pays for all of this, and for how long?"

Sam: Ah. So the worry isn't about Nvidia. It's about Nvidia's customers.

Alex: Exactly. And two details under that headline explain the nerves. First, roughly half of Nvidia's data-centre revenue now comes from beyond the big familiar cloud companies — including AI-specific clouds and, crucially, sovereign customers. Nations. Countries buying compute as national infrastructure.

Sam: Whole countries are buying chips like they're buying roads.

Alex: That's exactly the shift — compute as statecraft. It broadens the customer base, but it ties Nvidia's fortunes to geopolitics. Second detail: Nvidia shipped zero of its Hopper data-centre chips into China this quarter, against four point six billion dollars' worth a year earlier — a self-inflicted hole carved out by export controls.

Sam: So the number was huge, but the reaction was the actual story.

Alex: The reaction was the tell. And there's a deeper worry underneath the share price, and it's the one I'd really sit with. It's about circularity.

Sam: Circularity. Unpack that.

Alex: A growing share of Nvidia's demand traces back to companies that are themselves funded by the AI boom. Labs raising tens of billions to buy GPUs. "Neoclouds" — newer cloud companies — borrowing against future AI revenue to buy GPUs. Sovereigns spending strategic budgets to buy GPUs.

Sam: So the money flows in at the top, and it flows out to Nvidia at the bottom.

Alex: Right. And as long as capital keeps pouring in at the top, the revenue keeps landing at the bottom. So the question the falling stock is really asking is: what happens at the bottom if the top ever slows down?

Sam: That sounds — and tell me if I'm being dramatic — a little like everyone's selling each other GPUs with borrowed money.

Alex: You're not being dramatic, that's the worry in plain English. It's the kind of loop where the music's wonderful right up until it stops. And then layer on one more thing: depreciation.

Sam: Depreciation — that's the accounting word for "stuff wears out," right?

Alex: Exactly. When a company buys a GPU for, say, thirty thousand dollars, it doesn't count that as one big loss — it spreads the cost over the years it expects the chip to keep earning. If you assume six years, the yearly hit looks small and the profits look healthy. If the real answer is three years, suddenly you're losing money twice as fast on paper.

Sam: And nobody actually knows which it is.

Alex: That's the unsettling part. These chips age fast, and Nvidia keeps shipping better ones, which obsoletes the old ones. The assumption about how long a GPU earns its keep is one of the least-tested numbers in the entire build-out. If chips wear out — or get out-competed — faster than assumed, the whole sector's economics look very different overnight.

Sam: So when you say the market is making a "macro bet dressed up as a chip stock" —

Alex: — that's exactly it. Investors aren't doubting Nvidia's engineering, which is not in question. They're underwriting whether everyone else stays willing to keep buying. May was the month they started treating it like the macro bet it is.

Sam: Okay. You said compute itself became something you can trade. What do you mean?

Alex: This is the second infrastructure story, and it's a real shift in the unit of competition. Two things happened. Cerebras — the Nvidia challenger — listed publicly, which put a ninety-five-billion-dollar public price tag on compute. We'll do the IPO itself in the money lens. But the other one is wild: Anthropic rented an entire three-hundred-megawatt data centre from SpaceX. A facility called Colossus 1.

Sam: Rented... the whole building? Not some servers in it?

Alex: The whole building. Not a rack, not a row — the entire thing, over two hundred and twenty thousand GPUs.

Sam: Put that GPU number in human terms for me, because two hundred and twenty thousand of anything is hard to picture.

Alex: Think of it this way — a high-end gaming PC has one of these chips, maybe two. A serious company's whole AI setup might have a few dozen. Anthropic just took a single building with the equivalent of a couple of hundred thousand of them, drawing three hundred megawatts — roughly a small city's worth of power — for one lab's roadmap.

Sam: One lab. One building. A small city's electricity.

Alex: One lab. And on top of that, Anthropic pledged to cover any increase in consumer electricity prices that its US data centres cause.

Sam: Hold on, why would a company volunteer to pay people's power bills? That's not normal corporate behaviour.

Alex: It's a tell, and it's the perfect bridge to the third story. But first — the so-what on the lease. A year ago, labs competed on model quality. Now they compete on whether they can even secure a whole data centre's worth of power and silicon. Anthropic taking a building wholesale is what it looks like when compute access becomes the binding constraint on your roadmap. A chip company going public and a lab leasing a whole data centre are the same story from two ends — capital and capacity are now the game.

Sam: So back to the weird bit. Why is Anthropic pre-emptively offering to pay people's electricity bills?

Alex: Because of the third and maybe most consequential infrastructure story: power, not chips, is the real ceiling now. And the bill is going public.

Sam: Meaning what, concretely?

Alex: Microsoft is weighing whether to delay or abandon its 2030 clean-energy target — they call it "100/100/0" — because its AI data-centre build-out is pushing power demand past the assumptions that pledge was built on. They're hedging with on-site hydrogen fuel cells and nuclear deals, including a restart at Three Mile Island.

Sam: Three Mile Island. The one from the famous accident?

Alex: That one. Restarting nuclear to feed AI data centres. And analysts pegged the sector's near-term need at tens of gigawatts of new capacity.

Sam: Give me a gigawatt in terms I can feel.

Alex: Roughly, one gigawatt is about the output of a large power station — enough for something like a million homes. Tens of gigawatts is tens of power stations' worth of new demand, for AI, soon. Here's a number that makes it vivid: a single AI query can draw on the order of a thousand times the power of a regular web search.

Sam: A thousand times. Per query. Okay — that genuinely reframes it for me. I think of an AI question as basically free, like a Google search. You're telling me it's a thousand searches' worth of electricity.

Alex: On that order, for the heavy ones, yes. And multiply that by hundreds of millions of people using it daily, plus all the agents now running loops in the background — remember the agency story, these things don't ask one question, they ask thousands in a row to get something done — and you can see why the demand curve goes vertical.

Sam: Right, the agents are the hidden power hogs. Okay, now the consumer-bill thing makes sense.

Alex: That's the so-what exactly. This is the first AI constraint that lands on ordinary people's bills — which is why it became a public fight rather than a boardroom one. When a hyperscaler quietly walks back a flagship climate commitment just to keep building, and your power bill might go up to fund it, the politics change fast. Anthropic offering to cover the increases is the industry seeing the backlash coming.

Sam: So if chips were the bottleneck of 2025 —

Alex: — power and the grid are the bottleneck, and the political flashpoint, of 2026. Couple of quick mentions to close the lens: Nvidia explicitly flagged sovereign demand as a structural pillar now — compute as statecraft, like we said. And that Colossus 1 lease, the two-hundred-twenty-thousand-plus GPUs, really drives home how a single lab now consumes data centres at building scale.

Sam: So the standout for me from infrastructure: a record-smashing quarter where the stock fell anyway, because the market's now betting on whether everyone keeps buying — and underneath it all, the actual ceiling turned out to be electricity, not silicon.

Alex: That's the lens. And notice the thread — Cerebras listing, Anthropic's lease, the IPO queue — all of it is really about money flooding in. So let's follow that money. Right — let's move to business and markets. Business and markets. If big ideas was about capability and infrastructure was about the physical layer, this lens is about the capital — where the money is moving, and what it's revealing now that it has to move in public.

Sam: And this is where Cerebras finally gets its full story, right? You've been teasing it all episode.

Alex: It's the door the whole month walked through, so let's open it. On May 14th, Cerebras — an AI chip maker, an Nvidia challenger — priced its IPO at a hundred and eighty-five dollars a share. Raised about five and a half billion. Opened near three hundred and fifty dollars and closed up 68%, around three hundred and eleven.

Sam: So it nearly doubled on day one.

Alex: Nearly. That valued the company near ninety-five billion dollars — the biggest US tech IPO since Uber back in 2019. Demand reportedly outstripped supply more than twentyfold.

Sam: Twentyfold. So everyone wanted in. Did it hold?

Alex: It gave back about 10% the next day — froth, meet gravity, as it were. But the day-one number isn't really the point.

Sam: Quick sanity check, though — who even is Cerebras? Because I know Nvidia, I don't know these guys.

Alex: They make AI chips, but with a genuinely unusual approach — instead of lots of small chips wired together, they build one enormous chip the size of a dinner plate, which they argue is faster for certain AI work. The point for tonight isn't the engineering, though. It's that they're an Nvidia challenger, and the market was hungry enough to value the challenger at ninety-five billion on day one. That hunger is the signal.

Sam: Then what is? Because "chip company has a good IPO" doesn't feel like the biggest story of the month on its own.

Alex: Right, and here's the reframe: Cerebras was the door, not the headline. For two years, private AI valuations ballooned with no public-market test. Nobody outside could actually check them. A successful, massively oversubscribed listing is the proof-of-demand that lets everyone else file.

Sam: And "everyone else" being the big names.

Alex: Within weeks the queue formed. OpenAI, Anthropic, and a newly-public SpaceX-with-xAI-inside — together targeting roughly three point six trillion dollars in combined valuation. That turns the back half of 2026 into one of the most consequential IPO windows in market history.

Sam: Three point six trillion. I genuinely can't picture that number, so help me — why does it matter beyond "big"?

Alex: It's not the size that matters, it's the daylight. This is the real so-what. Private markets let AI labs run on narrative — a compelling founder, a scaling chart, a promise about AGI — without ever reconciling that story against an income statement a stranger could check.

Sam: And an IPO ends that.

Alex: An IPO forces what's called an S-1 — a detailed financial filing. For the first time, the public will see how much revenue these companies actually book against the staggering sums they spend on compute. How fast that gap is closing, or widening. And how aggressively they're depreciating hardware that might be obsolete in three years — there's that depreciation question again, from the Nvidia story.

Sam: So this is where we find out if the emperor's wearing clothes.

Alex: That's exactly the stakes, and it cuts both ways. If the numbers are strong, the listings pour rocket fuel on the whole sector and validate the build-out. If they reveal that even the leaders are burning capital faster than the business can justify, the same disclosure becomes the pin that finds the bubble.

Sam: And this connects right back to the Nvidia worry, doesn't it. The circularity thing.

Alex: It's the same anxiety from the other end, yeah. Nvidia's stock fell because the market's unsure who ultimately pays. The IPOs are the moment we actually find out — because the buyers have to open their books. If OpenAI and Anthropic are booking serious revenue against their compute spend, the circular-money fear evaporates. If they're not, the falling Nvidia stock looks prophetic.

Sam: So the IPO filings are basically the answer key to the question the Nvidia chart was asking.

Alex: That's a lovely way to put it. The chip story and the IPO story are the question and the answer.

Sam: So the years-long argument about whether AI is over-hyped or under-hyped —

Alex: — stops being a vibe and becomes a filing. And here's the subtle bit: the order of listing matters. The first lab to print its real economics sets the frame everyone else gets judged against. The bet went from private to priced.

Sam: You mentioned the xAI question got "answered by merger." What was the question, and what's the answer?

Alex: The open question was: what happens to xAI — Elon Musk's AI company — as a standalone thing? The answer: SpaceX completed an all-stock merger that folded xAI in as its AI-and-social division, alongside Starlink and the launch business. The combined entity moved toward a public listing — it eventually priced at a hundred and thirty-five dollars and started trading under the ticker SPCX in mid-June.

Sam: So why bundle them? Why not just float xAI on its own?

Alex: It's a financing-and-narrative move. It lets investors buy "Musk's AI" wrapped inside Starlink's actual cash flows, rather than as a standalone bet on a frontier lab. And it sets up a head-to-head: Anthropic is targeting a listing as soon as October; OpenAI is preparing to file as early as the third quarter, fresh off a funding round that valued it around eight hundred and fifty-two billion dollars.

Sam: So whoever lists first —

Alex: — sets the comparable that prices the rest. Think of it like the first house to sell on a street nobody's bought on in years — that price becomes the anchor every other valuation gets measured against, fair or not. The first AI lab to print its real numbers does that for the entire sector.

Sam: So there's actually a strategic advantage, or risk, in the timing of who goes first.

Alex: Huge one. Go first with strong numbers and you set a high bar that flatters you and pressures rivals. Go first with weak ones and you've handed everyone a stick to beat the whole sector with. The era of the privately-mythologised AI lab is ending. The era of the quarterly earnings call is beginning.

Sam: Okay. Now, earlier you kept deferring the Chinese models story to this lens. Pay it off.

Alex: This is the one hiding in the plumbing, and I love it because almost nobody outside the field noticed. By late May, Chinese open-weight models accounted for more than 60% of all tokens processed on a platform called OpenRouter.

Sam: Tokens being the units of AI text, and OpenRouter being...?

Alex: A platform that routes developers' AI requests to whichever model they pick — so it's a decent proxy for where real, actual usage flows. And 60% Chinese models, up from under 2% eighteen months earlier.

Sam: From two percent to sixty in a year and a half. That's not a drift, that's a landslide.

Alex: It's seismic. The leaders are models like MiniMax M2.5, Moonshot's Kimi K2.6, DeepSeek V3.2, Zhipu's GLM-5.1 — several now matching or beating the top US models on coding, at ten to twenty times lower cost.

Sam: Ten to twenty times cheaper and just as good. So why would anyone pay for the American one?

Alex: For mid-tier work, increasingly, they wouldn't — and that's the first big consequence. It caps what Western labs can charge for mid-tier inference. When an open model is a tenth of the price and good enough, the floor falls out of the pricing.

Sam: And you said there was a geopolitical second effect.

Alex: Yeah, and it's why this recurs in the governance lens. It routes a growing share of the West's actual AI usage through models trained — and sometimes hosted — under another government's jurisdiction. That's a data-sovereignty exposure that simply didn't exist a year ago. The neat way to hold it: the frontier is American, but the workhorse is increasingly Chinese.

Sam: Hang on though — if these Chinese models are open-weight, doesn't that mean you can download them and run them yourself? So is the sovereignty worry overblown? Just host it on your own machine.

Alex: You can, and serious companies do exactly that. But most developers don't — they hit it through a hosted service because it's easier and cheaper, which means in practice a lot of traffic still flows through servers under another jurisdiction. And there's a subtler point: even the model itself was trained by another government's industry, with whatever priorities and blind spots that bakes in. Open weights solve the hosting question if you bother; they don't change who built the thing your whole workflow now depends on.

Sam: So "you could self-host" is technically true and mostly not what happens.

Alex: That's the gap exactly. And it's why this is a genuinely uncomfortable split. Last business story? Two threads that tightened together. First, Microsoft moved visibly to reduce its dependence on OpenAI — building its own MAI model family, led by its first reasoning model, MAI-Thinking-1. Explicitly to cut costs and sidestep royalty payments. In blind tests their reasoning model was preferred over Claude Sonnet 4.6, matched the bigger Opus 4.6 on that SWE-Bench coding test, reportedly at a tenth of the cost of GPT-5.5 on one enterprise benchmark.

Sam: So Microsoft, OpenAI's biggest backer, is quietly building the thing it's been buying.

Alex: That's the headline tension. And second thread: enterprise AI agents crossed an adoption tipping point. Salesforce's "Agentforce" reached something like eight hundred million dollars in annual recurring revenue.

Sam: So the money's actually showing up somewhere. Where?

Alex: In the unglamorous middle. Not foundation-model subscriptions — agents embedded in CRMs and dev tools, doing billable work. That's the so-what. Microsoft building its own models is the clearest signal yet that even the closest partners now see foundation models as a commodity input worth owning rather than renting. The defining alliance of the early LLM era — OpenAI and Microsoft — is loosening into competition. The money's moving from "access to a model" to "work done by an agent."

Sam: Quick mentions to close the money lens?

Alex: Two. The hardware suppliers rode the boom toward trillion-dollar territory — Dell, HPE, Micron, Samsung all swept up in the demand. And a board-level worry emerged: agent token costs. Companies discovered that autonomous agents burn far more tokens than chatbots do — so AI spend started showing up as a nasty budget-line surprise. Which, by the way, is the cost side of the same agency story.

Sam: So the standout from business: Cerebras kicked open an IPO window worth three and a half trillion, which means for the first time we get to actually audit the AI build-out's economics instead of taking the pitch deck's word for it. And under the surface, the cheap workhorse of the whole industry quietly went Chinese.

Alex: Beautifully put — and that "audit in daylight" idea is the through-line of the whole month showing up again. Now, all of this — the capability, the chips, the money — it's running straight into the people whose job is to govern it. So let's finish there. Right — let's move to governance and society. Governance and society. Our last lens. This is where the technology meets the law, the regulators, and the labour market — and May was a month of genuine plot twists here.

Sam: Plot twists in regulation. That's not a sentence I expected.

Alex: It's earned. The biggest one: the EU blinked.

Sam: The EU — as in the people who wrote the big AI law everyone talks about?

Alex: The very same. The EU AI Act was supposed to be the planetary template — the strictest, most ambitious AI law in the world. In May, the EU reached a provisional agreement on something called the Digital Omnibus that postpones the Act's high-risk obligations. For stand-alone high-risk systems, from August 2026 all the way to December 2nd, 2027. For AI embedded in regulated products, to August 2028.

Sam: So they wrote the strict law and then... pushed the strict bit a year and a half down the road.

Alex: Essentially, yes. And the stated reason was pragmatic — the national authorities and the technical standards needed to actually comply with it weren't ready in time.

Sam: Okay, but is that a big deal or just bureaucratic slippage? Deadlines move all the time.

Alex: It's a big deal, and here's why. Europe spent two years positioning this as the global model — there's even a name for it, the "Brussels effect," the idea that EU rules become world rules because everyone has to comply to access the market. Delaying the teeth of it, under industry and competitiveness pressure, is an admission that getting hard rules right is harder than passing them.

Sam: And the signal that sends to everyone else?

Alex: That even the most regulation-forward bloc on earth concluded that moving too fast risked freezing out its own industry. Whether you read it as prudent or as capitulation, it shifts the near-term balance toward deployment over restraint.

Sam: Help me understand the "Brussels effect" a bit more, because that's the part that makes this matter globally rather than just for Europe.

Alex: Sure. The idea is that the EU is a market too big to ignore — so if you want to sell there, you build your product to its rules. And once you've built it that way, it's easier to just ship that version everywhere. So Europe's rules quietly become the world's default, without anyone else voting on them. It's how EU privacy law ended up shaping the cookie banners you click on every website on earth.

Sam: Oh, those are an EU thing? That actually lands it for me.

Alex: That's the Brussels effect in your daily life. So when Europe softens the teeth of its AI law, it's not just a European story — it's the world's most likely default ruleset getting weaker before it ever took hold.

Sam: There's something almost ironic about the timing, though, isn't there?

Alex: That's the part worth really sitting with, and you've spotted it. This delay landed in the exact same month that AI capability stopped being hypothetical. A system finding ten thousand exploitable software flaws. Another running an autonomous chemistry lab. That's precisely the moment the case for guardrails is strongest — and it's the moment Europe eased off.

Sam: So the gap between how fast the tech moves and how fast the law can move —

Alex: — is widening, and the bloc most committed to closing that gap just conceded it couldn't, at least not yet. The competitiveness argument won: the fear that hard rules push AI investment to the US and China outweighed the fear of under-regulating. For deployers, the practical effect is roughly eighteen extra months of lighter-touch operation in the world's second-largest market. For the bigger question — can democratic states regulate frontier AI on anything like its own timescale — May's answer was a quiet, consequential "not yet."

Sam: Right. You said the cyber-threat "got the call." What happened?

Alex: This is the defensive flip-side of the Anthropic Mythos story landing in policy, in the same month. US financial regulators convened banks specifically on AI-driven cyber risk.

Sam: Because if an AI can find a seventeen-year-old root exploit on its own —

Alex: — then in the wrong hands it's an automated attack engine pointed at the most systemically important targets there are. Banks. And the so-what here is the compression of the timeline. AI cyber risk went from conference-panel talk to regulator-to-bank briefings in the span of a single month — because the capability stopped being theoretical the moment Mythos produced thousands of validated, exploitable flaws.

Sam: And this is your agency point becoming actual policy.

Alex: This is the first arena where it does. The question regulators are now asking isn't "how smart is the model." It's "how much autonomous reach should it have inside a critical system — and how fast can defenders deploy the same thing before attackers do." Agency, not intelligence, written into a regulator's agenda.

Sam: Okay, the last one. The jobs story. And you've been promising me this paradox all episode.

Alex: This is the month's sharpest paradox, and it's about work. Sam Altman said he'd been "pretty wrong" about AI's near-term economic impact — that the entry-level white-collar wipeout he warned about in 2025 just hadn't…