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 welcome back to Barely Possible. It's Halloween, the kids are gonna be banging on your door in a couple hours, and I've got a menu of delicious little AI morsels for you before the sugar crash. Grab your coffee, pull up a chair, and let's have at it.
Here's the thing I want to lead with today, and it's not a costume gag, though we'll get to those. It's a number. One number that tells you more about where this whole AI thing sits right now than a hundred earnings calls. Alphabet lost half a trillion dollars in market value in a single day. Five hundred billion. Gone. Worst day in the company's history. And they didn't do anything wrong, exactly. They reported that their AI is working. That's the part that should make your neck hair stand up. Their AI is working, and that's precisely what spooked everybody.
So let me walk you through this, because this is the real story today, and everything else kind of orbits around it.
There's a piece from TechCrunch by Rebecca Bellan, published the thirtieth of October, headline is "Cracks are emerging in the AI spending boom." And the setup is simple. Over the past two weeks, Alphabet, Amazon, Meta, and Microsoft have all laid out plans to spend more than seven hundred forty billion dollars on capital expenditures this year. Seven hundred forty billion. That's data centers, that's chips, that's the physical guts of the AI buildout. And here's what changed. For the last couple years, when these companies announced they were gonna spend like drunk sailors, Wall Street cheered. They read it as confidence. More spending equals more faith in the future, buy buy buy.
Now? The mood flipped. The same spending that used to be a green light is suddenly a red flag. And the reason is this phrase you're gonna hear a lot: negative free cash flow. Now I know, "free cash flow" sounds like something your accountant mumbles. Let me put it in kitchen-table terms. It's the money left over after a company pays for everything it needs to keep the lights on and keep growing. For years, these giants had so much of it that it was basically falling out of their pockets. Now they're spending more cash than they're bringing in. They're in the red.
And Alphabet's the clearest example. Google's parent reported negative free cash flow of nine point one billion dollars in the third quarter. Second quarter in a row in the red. Meanwhile their quarterly capex ballooned to fifty-one point four billion. They expect to spend up to two hundred five billion dollars this year, and they're saying that number keeps climbing into 2027.
Now here's the part that's actually, genuinely interesting, and this is the thing I want you builders to really chew on. The thing that's eating Google alive is its own success. Google's got this thing called AI Mode. It's the AI chatbot feature built right into Search. And people love it. It's a hit. But every time somebody uses AI Mode instead of clicking through a bunch of blue links with ads next to them, Google makes less money. Their search advertising business, which is roughly half of all their revenue, saw growth slow to twelve percent. That's the weakest pace in five quarters. Sundar Pichai basically admitted it on the earnings call last week — the strong growth in AI mode queries is hitting them because AI mode monetizes worse than the old way.
So think about the trap here. Google is spending two hundred billion dollars building the AI technology that is actively undermining the search business that generates the two hundred billion dollars. That's not a strategy, kiddos. That's a snake eating its own tail and calling it lunch.
And it's not just Google. Amazon reported negative free cash flow of forty-two point four billion over the past twelve months. Year before that, they were positive fourteen point eight billion. That's a sixty billion dollar swing in the wrong direction. Their cloud business, AWS, grew thirty-nine percent — that's great, that's genuinely strong — and the stock still dropped seven percent after earnings. Meta's burning cash too, negative two point one billion in free cash flow in the quarter, guiding to as much as a hundred forty-five billion in capex this year. Zuckerberg even admitted that Muse, their new free AI agent, is a money pit — "very compute-intensive to run" is the quote. Their ad revenue grew twenty-six percent and the stock still took a ten percent haircut.
The only one who came through clean was Microsoft. Eighteen point six billion in positive free cash flow. But — and this is a wicked little twist — the reason they looked good is they actually spent less on capex than they projected. Forty-five point two billion instead of the fifty billion they planned. And Microsoft had to come out and say, no no no, we're not pulling back on AI, we just couldn't get the capacity from our third-party providers in time. And what happened? The shares of the smaller companies that rent out AI computing power — the ones Microsoft would've bought from — those sank. Because if Microsoft isn't renting, who is?
Now let me zoom out, because here's what makes this a builder story and not just a stock-ticker story. The anxiety spilled all over the market on Friday. The Nasdaq fell almost three percent. Nvidia dropped four percent, Broadcom more than five, Qualcomm eleven percent. And where's all the money for this buildout actually coming from? Not just profits. According to S&P Global Market Intelligence, Microsoft, Alphabet, Amazon, Meta, and Oracle raised a combined three hundred twenty-six billion dollars in debt in 2026. Triple what they raised last year. Triple.
So that's the picture. The demand for AI compute is real. It's red hot. Nobody's disputing that. But the legacy businesses that have always paid for everything — search ads, retail, the boring reliable cash cows — those are the ones getting squeezed, sometimes by the AI itself. And the gap is getting filled with borrowed money.
Which brings us to the part that should matter most to you if you're building a company: Anthropic. We've been talking about Anthropic's IPO all week on this show — the S-1, the risk factors, the spend-reporting tools. I'm not gonna re-run yesterday's tape. But there's a genuinely new wrinkle today, and it ties right into this spending story. The Financial Times reported Friday, and TechCrunch's Lucas Ropek picked it up, that Blackstone — which is the anchor investor in a sixty-three billion dollar deal to finance Anthropic's chips — told investors it now expects Anthropic to get a "high junk" rating from the credit agencies after its IPO.
Now stop and sit with that for a second. "Junk" rating. The polite term is "speculative grade." We're talking the double-B range, which is the top tier of junk, but it's still junk. And here's why it matters: Blackstone had previously said the opposite. They'd told people Anthropic would get an investment-grade rating — low risk, cheap money — before 2027. Now they've reversed. And a junk rating isn't just a label. It makes it harder and more expensive for Anthropic to borrow the money it needs to keep funding its spending spree. Which means they lean even harder on the hyperscalers — the Googles and Amazons — to fund their infrastructure. The same hyperscalers who, as we just discussed, are themselves bleeding free cash flow.
You see the daisy chain here? Anthropic depends on compute it can't fully afford, borrows to get it, and the lenders are getting nervous, which pushes Anthropic back toward the giants, who are also stretched thin and borrowing. That's a lot of IOUs stacked on top of each other. I'm not calling it a bubble — I leave the crystal-ball stuff to people who get paid to be wrong confidently. But if you're a founder and your whole business plan assumes cheap, abundant compute forever, this week is the week the market reminded everybody that nothing is forever, and nothing is free.
Now let me connect this to a story that's almost poetic in its timing. Same day, TechCrunch's Amanda Silberling runs a piece about Tyler Shultz. Now, Tyler Shultz, for the kids in the back, was the whistleblower at Theranos. His first job out of college, he's working at Elizabeth Holmes' blood-testing company, and he's the one who figured out the machines didn't work and blew the whistle. His grandfather, George Shultz, was a Secretary of State and a Theranos board member and investor, so this was not an easy thing for the kid to do. Family dinners got real quiet, I imagine.
And there's a documentary out right now about Holmes' final weeks before prison. It's become one of the highest-grossing documentaries of all time. But Shultz didn't need the movie. He's got thoughts, and they're sharp. He said, and I'm quoting him here: "the only reason we're talking about Elizabeth Holmes right now is because she wants us to be talking about Elizabeth Holmes right now."
But the line that stopped me cold, the one that earns its place on this show, is this. Shultz said: "I think the days of having a thirty-minute coffee with someone and sending them a hundred thousand dollars should be over. Elizabeth took that model and scaled it to a hundred million." And then he brings it home for the AI era. He said: "A huge danger with AI is that I think it is really hard to do due diligence on it. It's really hard to tell the difference between what is something truly unique and new versus like, a ChatGPT wrapper. If you want to deceive people, I think this is a great opportunity to go deceive people."
Now I want to be careful here, because I'm not drawing a line from Theranos to any specific AI company. That's lazy and it's unfair. Nobody's accusing anybody of fraud. But the structural point Shultz is making is dead on, and it rhymes with the spending story we just walked through. When money is flying around at trillion-dollar valuations, when companies that have never turned a profit are raising capital like it's confetti, the discipline of actually kicking the tires — asking "wait, does this thing actually work, and is it actually different from the forty other things that look just like it" — that discipline gets expensive and unfashionable. And the one thing that makes AI particularly slippery, per Shultz, is that it's genuinely hard to tell real innovation from a thin layer of polish over somebody else's model.
For you builders, there's a double lesson in that. One, if you're raising, the era of easy money on a good story might be tightening — see the whole first half of this episode. And two, if you're building something real, the ability to actually demonstrate that it works, that it's not a wrapper, that it has a moat beyond vibes — that's about to be worth a lot more than it was six months ago. The tide's going out a little. You wanna have your trunks on.
Alright. Let me shift gears, because not everything today is doom and debt. Some of it is just gloriously weird. Let's talk about space, because we got three stories today about putting computers in orbit, and they range from "actually happening" to "charmingly delusional."
Start with the charming one. Google. TechCrunch's Tim Fernholz reports that Google's first data center satellite — part of a project called Suncatcher — is online and communicating with the ground. It launched earlier this month on a SpaceX Falcon 9, alongside more than a hundred other spacecraft. This little guy carries four of Google's Tensor Processing Units — their custom AI chips — and generates up to one kilowatt of electricity. The whole idea is to eventually build networks of satellites flying close together, doing big AI inference workloads up in orbit where there's endless sunlight and no neighbors complaining about the power bill.
And when the engineers sent their first prompt to the model loaded onto the satellite — that's Google DeepMind's Gemma 4 — they asked it to write a haiku about being in space. And I'm gonna read it to you, because it's Halloween and we could all use a little whimsy. The satellite wrote:
"Blue marble below. Data streams in silent grace. Starlight feeds my core."
Not bad for a box of chips in the vacuum of space. Though I'll note — the power demands of the chip are so high it can only turn on one TPU for about twenty minutes at a time. So it's a poet, but it's a part-time poet. Now, the real reason this matters, and the reason it's not just a cute stunt: they're testing how these chips hold up against space radiation. Radiation can cause what they call "bit flips" — it literally flips a stored value in computer memory, and that can corrupt data without tripping any of your normal alarms. Ground testing suggested this would be minor for inference but a real problem for training. Now they get to see it for real. And look — this is a classic Google moonshot. They're openly saying the economics only make sense in the 2030s. There was even a peer-reviewed analysis saying SpaceX would need to fly its Starship rocket eighteen hundred times to get the cost reductions needed to make this pencil out. Eighteen hundred times. So file this under "fascinating, long way off."
Now here's the connective tissue, and it's a little grim. This beautiful space-poetry moonshot comes from the same company that lost half a trillion dollars in value this week. Fernholz actually points it out in the piece — getting to this 2030s payoff "may depend on how the rest of the company responds to the pressures and opportunities of AI in the near term." Translation: it's easy to fund the dreamy orbital stuff when your core business is a cash geyser. It gets a lot harder when your free cash flow goes negative and your search ads are slowing down. The poet in orbit lives and dies by the accountant on the ground.
Second space story, and this one's scrappier. TechCrunch's Dominic-Madori Davis writes about a startup called Satlyt, founded by a guy named Rama Afullo who previously worked at Google and SpaceX. Satlyt isn't building satellites — they're building software to turn existing satellites into a shared data-processing network. After thirteen months of work, they just ran a successful test with the University of Houston on a NASA-funded research project. The software figures out how to spread computing tasks across multiple satellites based on what compute and communication is available. Their software is already on three satellites in orbit.
And I love this story for a very specific reason. Afullo's totally honest about the warts. He calls the finding "a trade-off" — one approach worked best but needed more computing power. They even found a bug in how their software was time-stamping delays — it was recording when something arrived instead of how long it took to arrive, which is a totally different number. And he just... says that. Out loud. To a reporter. That's a founder who understands that "it mostly worked and here's what broke" is a more credible pitch than "everything's perfect." In a week where we're talking about the difference between real and a wrapper, that kind of candor is a signal worth noticing.
And the third space item, I'll keep brief because it's older, resurfaced context — Ars Technica had a piece from Stephen Clark about China's secret spaceplane. The underlying report, a Defense Intelligence Agency threat assessment, traces back to observations from 2024 and 2026, so I'm not gonna present it as breaking news. But the one durable nugget worth your time: there's now an emerging field called "non-Earth imaging," where commercial Earth-observation satellites point their telescopes at other satellites in orbit to inspect them. Corporate and international espionage, moved up to the high frontier. If you're building anything that goes to space, know that your competitors — and governments — can now take a good hard look at your hardware while it's up there. Privacy's dead down here, and apparently it's dead up there too.
Now let me pivot from the heavens back down to earth, because there's a story today that's a genuine sign of where AI is heading in professional services, and it's got that Wall Street drama you can't make up.
TechCrunch's Julie Bort reports that an AI-first law firm called Crosby has put in a bid to buy some of the assets of Weil, Gotshal and Manges. Now, for those of you who don't follow Big Law — Weil Gotshal was a storied white-shoe Wall Street firm, top twenty in the country, they handled the Lehman Brothers and Enron bankruptcies. And this week they confirmed to the New York Times that they're preparing to vote on dissolution. Winding the whole thing down. Why? The reporting says it was an exodus of top lawyers, a failed succession plan — the heir apparent turned down the job — and a merger that fell through. Important: nobody's blaming AI for the collapse. This was old-fashioned human drama.
But here's where it gets interesting for builders. Crosby is a two-year-old startup. It's raised about eighty-five million dollars, backed by Lux, Index, Sequoia, Bain Capital Ventures. And crucially — Crosby isn't selling AI tools to law firms like the Harveys and Legoras of the world. Crosby IS a law firm. An actual one, with clients, that uses its own AI to do the legal work. And they've bid to buy Weil Gotshal's trademark, its brand, and — this is the key part — some of its internal data. Plus they want to hire some of the firm's twelve hundred lawyers.
Now why would a two-year-old AI startup want a dead law firm's data? And this is the insight. A hundred-year-old firm like Weil Gotshal isn't just an office building full of lawyers. It's an archive. Decades of precedent documents, contract templates, workflows — the institutional playbook for doing incredibly specialized, incredibly expensive legal work like large corporate bankruptcy. Stuff that historically showed up as zero on a balance sheet because all the value supposedly walked out the door with the partners. But if you're building an AI that needs to learn how elite legal work actually gets done? That archive could be worth tens of millions. It's training data with a pedigree.
And to be clear — this isn't about selling private client records. That'd be wildly illegal. It's the firm's own precedent and process. There's no guarantee Crosby wins the bid. But the move itself tells you something. The next frontier for AI in white-collar work might not be building a better model. It might be buying the dusty institutional knowledge that the models have never seen. The moat isn't the algorithm — it's the fifty years of filings in the basement.
And it fits a pattern you can see all over today's menu. OpenAI's own feed had a piece about Ropes and Gray developing AI for transactional due diligence, and another about Credit Saison, a financial company, claiming a nine hundred fifty-four percent return on investment from a ChatGPT Enterprise pilot. Now, I always take a vendor's own ROI number with a boulder of salt — nine hundred fifty-four percent is the kind of figure that comes from a press release, not an auditor. But the direction is unmistakable: the real action in enterprise AI right now is in the deeply specialized, document-heavy, expertise-gated professions. Law, finance, diligence. Not chatbots. Playbooks.
Alright, let me do a quick lightning round of the smaller stuff, because there's some good texture here for builders.
First, a cautionary tale with a twist. TechCrunch's Zack Whittaker reports that Applebee's — yeah, the two-for-twenty-dollar place — exposed its customers' personal data because somebody left a database password hardcoded right inside the mobile app. For the non-engineers: "hardcoding credentials" means you literally write the password into the app's code. And anyone who downloads the app — over a million downloads on Google Play, by the way — can crack it open and pull that password out, then walk right into the customer database like they own the place. Names, emails, dates of birth, phone numbers, zip codes, even veteran status and the IP address of your last login, which gives away roughly where you are. Applebee's has at least eight and a half million people in its rewards program.
Now here's the twist, and it's the part that matters for you. The researcher, a UK fellow named Robert Wiggins, used AI to find that credential — buried in thousands of lines of source code. So AI is now a tool on both sides of the fence. It finds the vulnerabilities, and it finds them fast. If you're shipping an app, the lesson is older than AI itself — never, ever put a secret in your client code — but the enforcement just got a whole lot more aggressive, because now the scanners are tireless. And one more detail that should chill any founder: when TechCrunch asked Applebee's parent company if they even had server logs to check whether anyone had accessed the data, the spokesperson wouldn't say. You can't investigate a breach if you weren't keeping records in the first place. Log your stuff, people.
Second quick one — and this is a nice counterpoint to the hardcoded-password disaster — there's a startup called Zest. TechCrunch's Sarah Perez covered it. It's a restaurant recommendation app, been around since late 2024, backed by Alexis Ohanian among others. And their new move is interesting: instead of making you open an app, their AI "restaurant concierge" lives in your text messages. You text it, it recommends places based on where you've actually eaten — it links to your credit card data through Plaid to learn your real tastes — and then it books the table for you through OpenTable or Resy. Drop it in a group chat and it blends everybody's tastes and even runs a vote. The founder's line is the one to remember: "The agent is the product now." That's the shift. The app isn't the interface anymore. The conversation is. If you're building consumer software, chew on that — the thing people interact with might not be your beautifully designed screen. It might be a text thread.
Now, because it's Halloween, I've gotta give you a little treat. TechCrunch ran a list of last-minute costumes for tech nerds, courtesy of Amanda Silberling, and a couple of them are genuinely great commentary dressed up as jokes. My favorite: dress up as an AI usage limit notification. Grab a piece of cardboard, print out those "you've reached your usage limit" screenshots, and go spook some vibe-coders. Because, as the piece puts it, for some people there's nothing scarier than running out of Claude tokens. And honestly? After the week we just had with the spending story, running out of compute is the real horror movie. The costume's funnier than it has any right to be.
Let me close with one more thing, because it ties everything together. There was a piece on OpenAI's feed this week with a title that stuck with me — "The future of the firm." The subhead is about how AI might enter the org chart — "beneath, beside, and between us." It's a recent reflection, not a product launch, so I'm not gonna over-read it. But it sits right at the intersection of everything we talked about today. The law firm getting bought for its data. The agent becoming the product. The spending that's betting the whole farm on AI entering not just your workflow but your actual company structure.
And here's what I keep coming back to, kiddos. This week gave us both sides of the coin at once. On one side, a satellite in orbit writing poetry about starlight feeding its core — the pure, optimistic, moonshot version of this technology. And on the other side, half a trillion dollars evaporating in a day, junk ratings, negative cash flow, a Theranos whistleblower warning that it's never been easier to deceive people. Same week. Same industry. Same technology.
The truth is usually somewhere in the middle, and the people who build carefully through moments like this — who can actually show their thing works, who keep their logs, who tell the reporter about the bug instead of hiding it, who own the data and not just the demo — those are the ones still standing when the tide comes back in. And it always comes back in.
That's the menu for today. Keep your passwords out of your source code, keep your due diligence sharp, and if some kid shows up at your door dressed as a usage-limit notification, give him the full-size candy bar. He's earned it.
This is Tony DeLuca. Happy Halloween, watch your free cash flow, and I'll catch you on the next one.