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 plate of AI business coming out of the oven today. Grab your coffee, settle in, and let's have a look at what matters and what's just noise.
Let me start with a number that stopped me cold, because it's the kind of number that tells you something about the whole room, not just one company. Anthropic — the Claude people, the ones you've heard me talk about plenty this month — put out an S-1 filing that leaked over to The Wall Street Journal. And buried in that filing is a figure that I've read three times to make sure I wasn't misreading it. Anthropic believes it could spend more than seven hundred fifty billion dollars on cloud contracts. Seven hundred fifty billion. With a B. On renting computers. And — here's the kicker — the filing says that still isn't enough.
Now, I want to slow down on this one because it's the story today that actually tells you how the sausage gets made in this business, and it's the deep dive, so stick with me. Let me read you what the piece lays out, because the details matter.
Anthropic is having trouble securing enough cloud capacity from AWS, Google, and Microsoft. And the reason is the part that should make every founder sit up: those three hyperscalers increasingly see the Claude-maker as a competitor rather than a customer. Think about that relationship for a second. You're a company. You rent your entire operation — the thing that makes your product work — from three landlords. And those same three landlords are also trying to build the exact thing you build. They have their own models. They have their own ambitions. You are paying them rent, and they are watching you grow, and at some point the math in their heads flips from "nice customer" to "that's the guy eating my lunch."
So what does Anthropic do? The filing says they're leasing their own data centers and working with third-party data center operators — the newer players, the ones outside the big three. But — and this is a direct worry from the filing — those operators have, quote, "a limited operating history," which could make them less reliable than the hyperscalers. So you're caught in a vice. The reliable suppliers see you as the enemy. The suppliers who'd love your business are untested. Pick your poison.
And then there's the money underneath all of it. The filing reports an eighty point seven billion dollar net loss in the first half of the year. Eighty billion in the red, first six months. Against sixteen point three billion in revenue over that same stretch. Now, the revenue growth is genuinely enormous — last week the company told prospective investors that third-quarter revenue alone has reached sixteen point eight billion, up forty-five percent from the previous quarter. So this is not a company that can't sell. People want Claude. The demand is real and it's roaring.
But here's the detail that a builder should really chew on: only two customers account for about a fifth of sales. One-fifth of your revenue, two accounts. That's concentration risk that would make any seasoned operator nervous. If one of those two gets a better deal, or builds in-house, or just has a bad quarter, you feel it immediately.
And then the filing does something you don't usually see spelled out this plainly. Building on earlier catastrophic-risk disclosures, it warns that Anthropic could be held liable for damages caused by harmful AI behavior — to a dollar amount that would exceed available insurance coverage. Let me translate that from legalese. They're telling prospective investors, in writing, "if our product does something terrible, the damages could be bigger than any insurance we can buy." That's an honest disclosure. It's also a flashing hazard light bolted to the front of a November IPO.
So let me tell you what I think this story actually is, because it's not just "big company spends big money." We've circled the AI-bubble conversation plenty around here, and I'm not going to pretend I know whether the music stops. But what this filing shows you is the specific shape of the dependency. The frontier AI business is not really a software business in the way we used to understand software. Software was cheap to run and you kept almost all the money. This is a business where your single largest cost is rent on machines, your landlords are also your rivals, your revenue leans on a tiny handful of whales, and your legal exposure is, by your own admission, bigger than you can insure. That's not a software company. That's closer to an airline that leases its planes from its three biggest competitors.
For you, the builder listening, the lesson isn't "Anthropic's in trouble." They might do just fine. The lesson is about supplier power. If you're building on top of these models — and most of you are — understand that the people selling you the picks and shovels are not neutral. They have their own goldmine. The moment your product looks like something they want to own, the terms change. Anthropic is the biggest, best-funded example of this squeeze, and even they can't fully get out from under it. So when you're writing your own architecture, ask yourself: who's my landlord, and does my landlord want to become me?
Now, that question of "the house always has an angle" runs right into the next story, because the house we're talking about here is Elon Musk's house, and the angle is wild.
So, framing this correctly — this is an older report resurfacing, from back in November of last year, the SpaceX results discussion. But it's worth sitting with because of what it tells you about where the compute money is going. SpaceX — the rocket company, remember that part — reported that its AI business is now fifty-two percent of its income. More than half. The rocket company makes most of its money from artificial intelligence now. They renamed the whole line of business "Super Intelligence," and it pulled in seven and a half billion in revenue off the acquired Cursor coding tool, the Grok bot, and most importantly a neocloud business running one point seven gigawatts of compute.
And here's the line from Musk in that discussion that connects straight back to the Anthropic story. He said, quote, "People are starting to realize that the constraint on deployment of super intelligence is more in power and cooling and infrastructure than it is in chips." He compared SpaceX's ability to build out power infrastructure to — his words — "the New York Yankees playing in Little League. It's not fair." Now, I'm a Bronx guy, I'll let the Yankees comparison slide, but you hear what he's saying. The bottleneck isn't the GPUs anymore. It's the electricity and the cooling to run them. And he gave a concrete example — turbine makers are facing twelve-to-eighteen-month delays getting the metal blades and vanes they need, so SpaceX is using its rocket metal-casting to build those parts itself and jump the line.
So put the two stories side by side. Anthropic, drowning in cloud costs it can't secure, dependent on landlords who hate them. And then SpaceX, vertically integrated down to the metal casting, turning a rocket company into a power-and-compute company. That's the real divide in this industry right now. It's not who has the best model. It's who controls the physical world underneath the model — the land, the power, the cooling, the metal. The companies that own the dirt and the turbines have leverage the pure-software AI shops simply don't.
Alright, let's shift from the infrastructure giants to something that actually matters to how you use these tools day to day. Let's talk about Claude Haiku 5.5, which Anthropic put out in a recent report a couple weeks back — I'm framing it as a recent announcement, not breaking-today news, because it dropped on the seventh.
Here's the short version, and I'll keep it short because I'm not going to drag you through benchmark tables. Haiku 5.5 is Anthropic's cheapest, fastest small model. And the pricing is the headline. It costs around seventy-five percent less to run than the previous Haiku. For requests up to a hundred thousand tokens — which Anthropic says is about ninety percent of what people actually send — it's ninety percent cheaper than the last version. Input tokens dropped to a dime per million. Output tokens, fifty cents per million. Those are the kind of numbers that change what you can afford to build.
And they didn't stop at the small model. They halved the price of cache reads on Sonnet 5.5, which they say makes Sonnet about twenty percent cheaper on most agentic work. And they're handing Max and Team subscribers monthly API credits — a hundred bucks, two hundred bucks, up to five hundred for teams — specifically to go experiment with building agents.
Now why does a founder care about a cheap small model? Because of the architecture everybody's converging on. You don't run your big expensive model for everything. You run the big one — Opus or Sonnet — as the lead, the one that plans and does the hard reasoning. And then you farm out the grunt work to the little cheap fast one. The summaries, the database lookups, the classification, the "go into the ten-K and pull me the one revenue line" kind of task. One of the companies quoted in the announcement, Rogo, put it exactly that way — a big model builds the deck, and a Haiku subagent goes into the filing and grabs the number the deck needs. Accurate enough to trust on that narrow job, cheap enough to run a thousand times.
So the practical takeaway: the cost of the boring, high-volume layer of your product just dropped through the floor. If you shelved an idea six months ago because running it at scale was too expensive, it might pencil out now. That's the real news here, not the leaderboard positions.
Now let's get to the story in today's pile that honestly reads like a spy novel, because it is one. This is a current story and it's a doozy. The FBI's case against a former OpenAI engineer, a fella named Abdulrahman Alfozan, accused of bugging a top-secret OpenAI conference room.
Here's what the filing lays out. A court on Monday denied this guy's release — he's staying locked up while the case proceeds. The government's argument is that he planted a hidden recording device in a conference room, and that the device was in place during a top-secret meeting with US government officials. In the days around when that bug was installed, OpenAI executives were meeting with government officials in that very room. So this is not corporate gossip. This is national-security-adjacent.
And the details the FBI dug up are something. On his iPhone, they found he'd accessed a folder called — I'm not making this up — "Saudi Gov Docs." They found photos on his phone of what appear to be additional concealed recording devices and the input-output panels of TVs and electronics in various OpenAI buildings, including another conference room where sensitive meetings happened. They found WhatsApp logs going back to 2025 where a member of the Saudi royal family seemed to be recruiting him. In one exchange, the royal messages that "Saudi needs you more than the US," and Alfozan responds with — quote — "100."
Now here's the part that should make every one of you building with these tools pause. According to the complaint, Alfozan allegedly used an OpenAI productivity agent — one of their own internal AI tools — to research which conference rooms might be best to exploit for his spying. And he used a ChatGPT account to research how to modify his recording devices so they'd auto-record when they detected speech and allow remote access. And after opening that "Saudi Gov Docs" folder, logs show he used OpenAI's productivity agent to ask, quote, "who from OpenAI could be suitable to attend a high level informal meeting with ministers in Saudi Arabia? List 10 people and why each. Ideally high levels."
Think about that. The alleged spy used the company's own AI assistant as a planning tool for the espionage against the company. The logs became the evidence. There's a bitter irony in there — the surveillance machine kept a perfect record of the surveillance. And for you, the operator, it's a reminder worth tattooing somewhere: every agent you deploy inside your company is also a witness. It logs everything. That's a liability and a safeguard at the same time, depending on which side of the subpoena you're standing on.
The government's calling him a flight risk — a one-way ticket to Riyadh allegedly paid for by the Saudi government, two bank accounts in Saudi Arabia, his wife and kid already gone. He faces up to five years on the recording charge, but the government signaled more serious charges may be coming. We covered the arrest itself when it first broke; this is the new filing with the folder, the WhatsApp logs, and the detention ruling. I'll keep an eye on where it goes.
Now that espionage case is really a story about guarding secrets, and wouldn't you know it, Anthropic put out a piece on exactly that — guarding their secrets from a very different kind of thief. This one's framed as an older story from September that resurfaced, so I'm not going to present it as breaking, but the subject is genuinely interesting for builders: distillation attacks.
Here's the plain-English version. Distillation is when somebody takes a big, expensive, smart model — in this case Claude — and uses its outputs, especially its reasoning, the step-by-step thinking it writes out before answering, to train a smaller, cheaper model to mimic it. It's copying the homework, basically. And Anthropic says it's been happening at industrial scale, with networks of fraudulent accounts — one network they describe controlled more than twenty thousand fake accounts at once — to pull the reasoning transcripts out and train competitors.
And they name names. They say a company called Zhipu distilled the cyber capabilities of leading US models while building something called GLM-5.3, and that a US standards body has since assessed that model as the most cyber-capable open-weights model released to date. They also describe a case where requests to another lab's models were secretly forwarded to Claude as part of a distillation effort, and one of those requests appeared to come from someone affiliated with China's military trying to analyze surveillance footage of a specific person.
Anthropic says its new defenses — classifiers, controlling where the reasoning can travel, banning accounts fast — have cut the reasoning that distillers can extract by a factor of ten to a hundred. They say they've turned away trillions of tokens of paying traffic from distillers. Let that sink in: they're voluntarily refusing paying customers because those customers are copying the product.
Now why does this matter to you beyond the geopolitics? Two things. One: if you're building on an API, know that the provider is now running aggressive classifiers that can, in rare cases, block or misidentify your legitimate requests — and Anthropic admits they now charge you even for blocked requests in these categories. So if your integration ever starts getting weird refusals, this is part of the landscape now. And two: the deeper point is that the reasoning trace — the thinking — is the crown jewel. It's what everybody's trying to steal and everybody's trying to hide. That tells you where the actual value lives in these systems. Not the final answer. The path to it.
Alright, let's come up for air and talk about something lighter but genuinely important for anybody building consumer products — the download numbers on Meta's AI companion app, Muse. Now, I want to flag the framing honestly: the underlying milestone comparison references Threads from back in 2023, so this isn't a brand-new-today event, it's data that resurfaced. But the Muse numbers themselves are the story.
Muse has hit ten point nine million downloads since its September debut, ten point two million of those in the US. That puts it in a club of only a dozen apps that have ever hit ten million US downloads in their first forty days — a club that's mostly mobile games like Fortnite and Pokémon GO, plus Threads and Disney+. For context, ChatGPT did three point seven million US downloads in its first forty days. Grok did five hundred thousand. Claude did two hundred thousand. Muse blew past all of them domestically.
But here's where I want you to be skeptical, because the download number isn't the interesting number. Meta is buying a lot of this. The piece says Meta put thirty-nine percent of its total advertising spend in October behind Muse, up from twenty-three percent the month before. Muse jumped from the seventy-third largest brand by ad impressions to the twenty-first. So some of this rocket is just Meta pointing the biggest marketing cannon on Earth at its own app across three and a half billion daily users. Anybody can buy downloads. I've seen it a hundred times — the install chart goes vertical and then the app's a ghost town in a month.
So what makes me actually pay attention here is retention, not installs. Muse is holding thirty percent of users after seven days and twenty-three percent after fourteen. Its rivals average nineteen percent and sixteen. That's a meaningful gap. And its daily active users hit two and a half million, up fifteen percent week over week. Retention and daily actives are the honest metrics — those are the ones you can't fully buy with ad spend. People are coming back.
And the thing driving the retention, by most accounts, is personality. Muse has a mascot, it's got a character, it's fun to talk to. That's the bet Meta's making — that once everybody's models are smart enough, the thing that keeps you is whether you enjoy the conversation. For the builders out there sweating over benchmark scores: the market may be telling you that past a certain capability threshold, charm beats IQ. That's an uncomfortable truth for the engineering-brained among us, but the numbers are the numbers.
Let me rattle through a few quicker ones, because there's a theme hiding in the pile.
OpenAI's been busy. They announced locked chats in ChatGPT — you can now put a four-digit PIN or Face ID on a conversation, so when you hand your phone to your mother-in-law to show her a photo, she can't scroll into whatever you were planning. Useful, modest. But read the fine print, because OpenAI did, to their credit, spell it out plainly: locking a chat does not mean end-to-end encryption, it does not create a separate encrypted vault, it does not change how long they retain the chat, and it does not remove what the model already remembered. So it keeps your cousin out. It does not keep OpenAI out. Know the difference between privacy from the person next to you and privacy from the company holding the data. Those are not the same thing, and the feature only does the first one.
OpenAI also acquired a company called Orchestra Bio, to push into life-sciences R&D and drug development. And here's the continuity note worth drawing — that same OpenAI, the one whose own AI tools allegedly got turned into an espionage planning assistant in the story we just covered, is now buying its way deeper into biotech. The breadth of this company's ambition is getting hard to track. Coding, consumer chat, now drug discovery.
Google, meanwhile, is pushing a more personal Gemini into your car — pulling from your Gmail, Calendar, YouTube, Maps, and search history so it can tell you your flight's delayed while you drive to the airport. The feature itself rolled out earlier in the year; this is an expansion. And I'll just say what the piece says: that level of enmeshing, Google's own reporter notes, might make some people uncomfortable. You opt in, you can toggle it off. But the direction of travel is clear — these assistants want to know everything about you so they can be useful, and "useful" and "surveilling" start to rhyme if you're not careful.
Mistral — the European outfit — shipped Voxtral 3, which is genuinely relevant if you're building anything with voice. It's a full voice-agent stack: transcription in fifty-plus languages, natural speech out, tool-calling, the whole pipeline, with each turn completing in under a second. And the pitch that matters for a lot of you is sovereignty — it runs on Mistral's EU-hosted API or on your own infrastructure, so you control where the audio goes and whether it's retained. For anybody building contact-center agents or voice copilots who's been nervous about shipping customer audio off to someone else's cloud, that's a real option on the table now, and the pricing's aggressive.
And over in the physical world — a VC named Amulya Vishwanath, who spent nearly seven years at Nvidia before starting a firm focused on physical AI, told TechCrunch's Equity that robotics is in its "LEGO era." Her point: everybody keeps waiting for the "ChatGPT moment for robotics," but we're actually earlier than that — people are still building the basic blocks, the hands, the actuators, the foundational software, that'll eventually snap together. I think that's the right framing, honestly. It's a cold-water correction to the robot hype. We're assembling the Legos, not building the castle yet. And her firm's whole pitch is that talent is the real moat, because so much of it is getting vacuumed up by the Anthropics and Nvidias of the world. Which, if you were listening to the first half of this show, is a theme you're now hearing for the third time: in this business, the scarce resources are power, metal, and people — not ideas.
Couple of real-economy notes to close the loop. GM is closing a battery assembly plant south of Detroit — the Brownstown plant that made packs for the Cadillac Celestiq and the Corvette E-Ray. But credit to the reporting, the headline tells you not to jump to conclusions: the company says it's not an EV-strategy retreat, it's just a lease running out, and the workers move to other facilities. Though the piece does note the broader context — the One Big Beautiful Bill Act repealed EV incentives last year, and GM's gotten quieter about electric since. So, not a crisis, but a weathervane.
And Rivian's CEO RJ Scaringe was out at TechCrunch Disrupt dreaming about how full autonomy will let him build "dramatically different" vehicle designs — rooms on wheels, radical form factors. And look, the reporting does the honest thing here, which I appreciate: it reminds you we heard this exact pitch a decade ago, when Mercedes and Toyota and Hyundai — Hyundai literally showed a car with legs — promised a Cambrian explosion of crazy vehicle shapes that never left the prototype stage. The difference now, maybe, is there are real robotaxis on real streets. But Scaringe himself admits ninety-five percent of US miles are driven in personally owned vehicles and that's not changing much even with full autonomy. So the dream is a dream, and he's "wildly excited," but the thing he's selling is a car that doesn't exist yet called R4 that he says, and I quote, "is going to be cool, it's going to be great." That, my friends, is the sound of a man who needs you to stay patient.
So let me tie the bow on this one. If there's a single thread through today's pile, it's that the AI business has graduated from being about clever software to being about who controls the hard, physical, boring stuff underneath — the power, the cooling, the metal, the data centers, the scarce engineers, even the legal exposure. Anthropic can sell all the Claude it wants and still get squeezed because it doesn't own its landlords. SpaceX turned a rocket company into a power company because it figured out the bottleneck was electrons, not algorithms. The robotics VC says we're still assembling Legos. And the model makers are now so protective of their reasoning traces they'll turn away trillions of tokens of revenue to keep them. The intelligence is getting commoditized. The stuff that's hard to copy — the dirt, the watts, the people, the trust — that's where the leverage is moving. Build with that in mind.
That's the menu for today, kiddos. Watch your landlords, read your fine print, and remember that every agent you turn on is also keeping notes. This has been Barely Possible. I'm Tony DeLuca — take care of yourselves, and I'll catch you next time.