Subscribe
Share
Share
Embed
A daily briefing on the AI systems, products, companies, and policy shifts that are just becoming possible.
Want a podcast for your own topics? Join early access: https://www.barelypossible.to/waitlist/?source_path=public_feed&feed_source=rss
Okay kiddos, I'm your boy Tony DeLuca, and we've got a fresh tray of AI morsels straight out of the oven today. Grab your coffee, pull up a chair, and let's have at it.
Here's the thing I want to put at the center of today's show. We've spent this whole week circling the same anxiety — the money, the bubble, the cloud debt, the hyperscaler dependency. And that's real, we'll get to some of it. But there's a quieter story running underneath all of it, and it showed up today in a piece about a nuclear reactor on the shore of Lake Michigan. It's a story about what happens when the "move fast and break things" culture that built Silicon Valley crashes into an industry where breaking things means radiation in somebody's hair. That's our spine today. The AI boom needs power. A lot of power. And the people racing to supply it are learning, the hard way, that physics doesn't care about your IPO timeline.
So let's dig into the reactor that was supposed to be the comeback story.
This is a recent report from Ars Technica, by Jamie Smyth, originally out of the Financial Times, published October twenty-third. And I want to be straight with you on the timeline, because this matters: the warning that's at the heart of this story, and the incidents behind it, those are from this past summer — July and August. So this isn't "a thing that happened this morning." This is a piece recently pulling together events from over a few months into one picture. And the picture is not pretty.
The plant is called Palisades, in Michigan, near South Haven. It first went online in 1971. It was decommissioned back in 2022. And then somebody got the bright idea to bring it back from the dead — the first time anyone's tried to reopen a shuttered nuclear reactor in the United States. The company doing it is Holtec International, out of New Jersey. They've got about two-point-two billion dollars in government support riding on this. It's been sold as the symbol of America's nuclear resurgence.
Now here's where it gets uncomfortable. Let me read you what actually happened, because the raw details tell you more than any summary could.
The US Nuclear Regulatory Commission — that's the NRC, the referees of the nuclear world — they did an inspection at Palisades in July over concerns about safety awareness. And quoting the report: "It followed an incident last year in which a worker fell into the reactor cavity, ingested some water, and was sent off-site for medical attention. The individual was decontaminated by radiation protection personnel but had 300 counts per minute detected in their hair, according to an incident report."
A guy fell into the reactor cavity. Swallowed some water. And they're counting radiation in his hair afterward. Let that sit for a second.
Then, quote: "In a second incident in August, a bundle of nuclear fuel rods tilted over during a refueling operation, delaying what was anticipated to be the first reopening of a shuttered nuclear reactor in the US."
A bundle of fuel rods nearly tipped over. And here's the kicker — the NRC said nine of the fifteen reactor operators at Palisades failed the test for their initial license exam over the summer. Nine out of fifteen. Six got licensed. So the majority of the people they put up to run this thing couldn't pass the test.
And the NRC's main finding, the thing they officially warned Holtec about, wasn't any single accident. It was what they called a "chilled work environment" — and that's the phrase I want you to remember. A chilled work environment. That's regulator-speak for: the workers are scared to raise their hands when something's wrong, because they're afraid of getting punished for it.
Now, who connects the dots here? The source quotes Allison Macfarlane, who was the NRC chair from 2012 to 2014. And she lays it out cleaner than anybody. Quote: "The situation at Palisades is that you have a power plant operator newbie in Holtec trying to restart a nuclear reactor, and they may be under a lot of financial and political pressure to get this done speedily." And then she says — and this is the line, this is the whole show in one sentence — quote: "That perhaps reflects the new culture sweeping through part of the industry — move fast and break things. But to ensure safety, nuclear needs a culture based on going slowly and carefully."
There it is. Move fast and break things. That's the motto that built the app economy. Ship it, patch it later, failure is a feature. And in software, fine — your worst case is the site goes down for an afternoon and somebody writes an angry post. But you cannot patch a tilted fuel rod later. You cannot hotfix a guy who fell in the reactor cavity. The failure mode is permanent, and it's measured in counts per minute.
So why's the pressure on? Why is a company that — and I want to be precise here, the source says this flat out — a company that has never operated a nuclear power plant, racing to restart one on a deadline? Because of us. Because of the AI buildout. The piece says plainly that the Trump administration has promoted nuclear energy as an edge in the AI race against China. They've slashed regulation, poured in tens of billions, and — quoting the article — "promised to reverse a culture of 'risk aversion' at the NRC to accelerate projects."
Reverse a culture of risk aversion. At a nuclear regulator. Think about what that phrase is actually asking for. Risk aversion is the entire job of a nuclear regulator. That's like asking a lifeguard to be less obsessed with people drowning.
And Holtec's on a clock. They've got a March 2027 contractual deadline to start delivering power to customers — customers who, by the way, got an extra one-point-three billion in government money just to buy the power. So the whole arrangement is leaning on this thing coming online, fast.
Now, in fairness — and I always want to give you the other side — Holtec pushed back hard. They told the FT they take the issues seriously and they're addressing them. They called the recovery of that tilted fuel assembly, and I'm quoting, "a remarkable demonstration of technical ingenuity." And they said, quote, "A strong safety-conscious work environment is fundamental to our culture... The assertion that we have rushed has directly been countered by the work that has been accomplished." And the NRC did downplay any immediate risk to the public, saying mitigation measures kept plant safety intact.
So that's the company's story. Ingenuity, not recklessness. And maybe. But you'll notice "look how brilliantly we recovered from the fuel rod nearly tipping over" is a weird flex. The brilliant move is the fuel rod not tipping over.
Here's the part that should really grab the builders and founders listening, though, because this isn't just a safety story — it's a market story. The warning to Holtec landed right as investor enthusiasm for nuclear is cooling off. The piece says more than thirty billion dollars has been wiped off the value of five of the biggest US-listed nuclear energy companies over the past year — Oklo, Centrus, BWX Technologies, NuScale, Nano Nuclear. And last month Holtec had to postpone an IPO that was supposed to raise nine hundred million and value the company at up to ten billion. Postponed. Pulled.
So connect it up. The AI boom created a frenzy for power. The frenzy for power created a nuclear gold rush. The gold rush pulled in operators who've never run a plant, put them on political deadlines, told the regulator to chill out — and now that the AI bubble chatter is getting louder, the money's getting nervous and the IPOs are getting yanked. Meanwhile the fuel rods are still physical objects that fall over.
The lesson for anybody building anything: hype can set your deadline, but it can't do your work for you. And in some industries, the gap between those two things is where people get hurt. Nuclear doesn't have a staging environment.
Alright. Let me shift from the plant to the people betting on the whole machine. We've been living in bubble-talk all week on this show, and I'm not going to re-litigate all of it. But there was one concrete number today that puts a human face on the Anthropic IPO everybody keeps circling.
This is a current piece from TechCrunch by Marina Temkin, published October twenty-third, citing Bloomberg. It walks through what Anthropic's early backers are worth on paper ahead of the IPO. And look, normally a "rich people are rich" story isn't worth your time. But the specifics here are genuinely wild and they tell you something about the shape of this market.
Anthropic's last private valuation was nine hundred sixty-five billion dollars. The IPO's expected this year, possibly valuing it at up to two trillion. So here's who's sitting pretty. Eric Schmidt — he put money into the Series A in 2021, when the company was valued at four hundred sixty-one million. Today he owns about one percent, worth nearly ten billion. One percent stake, ten billion dollars.
Spark Capital — and this is the one that made me whistle — they put in seventy-five million in the 2023 Series C, when Anthropic was valued at five billion. The piece says that position has, quote, "ballooned more than ninetyfold to $7 billion." Ninetyfold. In about three years.
And then there's the late money. Iconiq wrote a six billion dollar check last year when the company was already valued at a hundred eighty-three billion. That's worth seventeen billion now. Sequoia only got in earlier this year, at a three hundred eighty billion valuation, three billion in, already more than doubled.
Here's what I want you to notice, and it ties straight back to everything we've been chewing on. The early folks — Schmidt, Spark — they took a real risk on a weird research lab and got rewarded ninetyfold. That's venture capital doing its actual job. But the late money, the Iconiqs and Sequoias writing multi-billion-dollar checks at hundred-plus-billion valuations — they're not betting on whether Anthropic is a good company anymore. They're betting the IPO prices even higher than the already-insane private round. That's not venture investing. That's musical chairs, and everybody's watching the band.
And that's exactly the nervousness that's got the nuclear IPOs getting yanked and the hedge funds dumping tech at the fastest pace in a decade. When the late money is only worth it if the next guy pays more, the whole thing runs on confidence. The minute confidence cracks, the chairs disappear fast. I'm not calling a crash — I've watched enough of these cycles to know the bubble-callers are wrong more often than they're right. But I'd keep one eye on that Anthropic pricing, because that's the moment the music either keeps playing or stops.
Now let's go from the big money to the actual plumbing of building these agent companies, because there were a few things today that matter if you're actually shipping product.
First one — a company called Daytona hit a one billion dollar valuation. This is from TechCrunch, Julie Bort, and I want to flag the age on it: the underlying Series A reference is from February, so treat this as a renewed-attention piece, not breaking news. But the sixty million dollar Series B round, led by FirstMark Growth, is the fresh news. What's interesting isn't the money — it's what they're building and who's backing it.
Daytona is building a cloud specifically for long-running agents. And the founder, Ivan Burazin, said something in his announcement that's worth sitting with, quote: "Agents don't work like the long-lived services today's cloud was designed for. They spin up thousands of short, stateful jobs nobody planned in advance, and they need a cloud built for that." And his pitch is basically: give every agent its own computer — sandboxed Linux, Windows, macOS, GPUs, all on one platform.
Now the angels here are a who's who — Stripe's Patrick Collison, Shopify's Tobi Lütke, Perplexity's Aravind Srinivas, Datadog's Olivier Pomel, Quora's Adam D'Angelo. When those specific people put their own money into infrastructure, they're telling you where they think the pick-and-shovel money is. And the thesis is one I think is genuinely right: the cloud we built for 2015 web apps — long-lived servers, predictable load — is a bad fit for agents that spin up a thousand throwaway machines you never scheduled. If you're building agentic products, the infrastructure layer underneath you is getting reinvented right now, and that's worth watching more than another chatbot launch.
Related, and this one's just a quiet signal — OpenAI shipped something called the Decisions API. The description is short: "Fast, focused, low-cost decisions for any application." That's it, that's the whole tell. The industry's figuring out that you don't need a trillion-parameter model to answer a yes-or-no routing question. You need something fast and cheap that makes one call. That's a maturity signal. When the labs start shipping small, boring, cheap primitives instead of giant flashy models, that's builders' tools, not demo-day theater.
Okay, now let me connect the agent plumbing to the dark side of agents, because the same week everybody's celebrating agent clouds, the healthcare world got reminded what happens when the attackers move faster than the defenders.
This is Zack Whittaker at TechCrunch — and heads up, the broader context in this piece reaches back to breaches earlier in the year, but the Quest Diagnostics disclosure itself is the fresh item. Quest — the big lab-testing company, the folks who process your bloodwork — confirmed a cyberattack and data breach. They filed an 8-K with the SEC saying they learned of an attack targeting one of their lab systems called Quanum. And Quanum is used by about half a million clinicians across North America, holding what the piece describes as potentially hundreds of millions of medical documents.
They've confirmed patient records were accessed. They haven't said how many. And here's the detail that tells you everything about how companies handle this now — the piece notes that Quest's own security notice page contains "noindex" code. That's the instruction that tells Google: don't list this page in search results. So they disclosed it, legally, as required — and then quietly made the disclosure as hard to find as possible. That's the modern corporate two-step. Tell the truth where nobody can see it.
And this is Quest's second major breach in a decade. They lost eleven-point-nine million patient records back in 2019 through a billing vendor. So here's the through-line to the agent story: we are standing up thousands of autonomous agents, giving them their own computers, letting them log into our systems — and the companies holding the most sensitive data on earth can't keep attackers out with the tools they already have. The attack surface is about to explode, and the defenders are, charitably, treading water.
Let me do a couple quick hits before we land the plane, because there were a few things worth a mention.
The feds are nervous about something else, too — prediction markets. Now, I want to frame this carefully, because the specific incidents here — the arrested soldier, the Polymarket bets — those are from the spring, April timeframe. What's newer is Defense Secretary Pete Hegseth's October twentieth memo, published by CBS. The subject line was, quote, "Do Not Compromise Military Operations for Personal Financial Gain." And his argument is genuinely interesting for anybody in the crypto and prediction-market world. He wrote, quote: "Our enemies watch movements in these markets. When an American Service member... engages in a financial transaction on a military outcome, they create a signal. That signal can be aggregated, correlated, and exploited by hostile foreign intelligence services to anticipate our movements, targets, and timing."
Think about that. A betting market on Kalshi or Polymarket becomes an intelligence feed for foreign adversaries. The context is an Army soldier arrested in the spring for allegedly making four hundred ten thousand dollars in Polymarket bets on the timing of a military operation. And a report found a hundred fifty-two suspicious wallets won a combined eight million on defense markets. For the founders building in prediction markets: this is the regulatory shoe you've been waiting to drop. When the Pentagon says your product is an operational security risk, the light-touch regulatory honeymoon is ending.
Quick one on your inbox — Google updated its privacy policy, and this is a current item from Ars by Ryan Whitwam. The old language said flatly: "We don't show you personalized ads based on your content from Drive, Gmail, or Photos." That firewall language is now gone. The mechanism is a feature called Personal Intelligence, which pipes your app data into their AI products. Google insists ads aren't "based directly" on your Gmail. But as the piece puts it, that's a roundabout way of confirming that your Workspace and Photos data may inform ads if you've turned Personal Intelligence on. It's opt-in — but the piece notes you may have turned it on without realizing, through pop-ups. So if you care: Gemini settings, Personal Intelligence, disable connected apps. A builder lesson in there too — the quiet erosion of a hard promise into a soft "clarification" is the oldest play in the data business, and we just covered Spotify doing its own version with family-plan QR-code enforcement. Same instinct: get them hooked, then tighten.
And one for the gamers, because you gotta have a little dessert. A Valve developer said "HLX" out loud in public. If you don't know, HLX is the widely-reported internal codename for Valve's next Half-Life project. At an audio conference, composer Mike Morasky let it slip while talking shop, and his fellow panelist can be seen leaning over afterward to say, quote, "Good job — you only said HLX once!" It's the first time a Valve employee has uttered the codename in public. Means nothing, confirms nothing, and I love it anyway. Some mysteries are more fun unsolved.
So let me bring it home. We started on the shore of Lake Michigan, with a guy who fell in a reactor cavity and fuel rods that nearly tipped over, all because a company that's never run a nuclear plant is on a political deadline driven by the AI boom. And everything else today rhymed with it. The late money piling into Anthropic betting the next guy pays more. The agent infrastructure racing ahead of the security that's supposed to protect it. Google softening a hard promise into a shrug. A theme connects them, and it's the one Allison Macfarlane handed us: move fast and break things works great right up until the thing you break can't be un-broken.
Software gives you a staging environment. Nuclear reactors don't. Patient records don't. Trust doesn't. The builders who win the next couple years aren't the ones who move fastest — they're the ones who know which of their systems have a "patch it later" and which ones don't.
That's the menu for today, kiddos. Keep your fuel rods upright and your private keys bigger than the number four. This has been Barely Possible — I'm Tony DeLuca, and I'll catch you next time.