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 the coffee's hot, the menu's loaded, and we've got a real spread of AI morsels to chew on today. Pull up a chair. There's cloud money the size of small nations, there's a surveillance company telling cops that the people protesting them aren't real people, and there's a quiet little announcement from Anthropic about enterprise spend reports that tells you more about where this whole industry is heading than any trillion-dollar valuation.
Let me set the table. The through-line today, the thing I keep bumping into no matter which story I pick up, is this: the AI business has officially stopped being about whether the models are impressive and started being about whether anybody can prove the money is worth it. The receipts are coming due. Microsoft wants to show you the receipts. Anthropic wants to help you read your own receipts. Google is quietly admitting its old receipts are shrinking. That's the show. Let's eat.
Let's start with the one that'll make your CFO sit up, because this is the most consequential story today for anybody actually building a company.
Anthropic, on Wednesday, launched a new suite of reporting tools designed to show companies exactly what their AI spending is buying. And on the surface that sounds like dull enterprise plumbing. It is not. It's a tell. Here's the setup. The financials Anthropic shared with investors this week suggest Claude Code, their coding tool, is the business they wanted it to be. They recorded an operating profit in the second quarter. And they posted something called 900 percent net revenue retention. Now let me translate that into kitchen-table English, because the number is wild. Net revenue retention measures how much the same customers spend year over year. A hundred percent means they spent the same. Anthropic's number means spending from the same customers in the third quarter of 2025 grew nine times by the third quarter of 2026. Same customers. Nine times the spend. That is not growth, that's a vertical wall.
So what do these new tools actually do? Using internal analytics and employee surveys, a manager can look at a team's work over a date range, see what the team produced, how much it cost, and where the workflows broke down. The reports even offer recommendations on how to improve. And Anthropic says, importantly, it won't use this data for model training by default. They're also rolling out pooled group budgets so teams can manage token spending together.
Now here's why I care, and why you should. Tim Fernholz, who wrote this up, put his finger on the real story. Businesses have been hunting for exactly this kind of information. Because while the venture-backed crowd is spending like drunken sailors on tokens, most of the Fortune 500 are not seeing major gains from deploying AI. Let that sit. Most of the Fortune 500. Not seeing major gains. The spend is going up and the payoff is murky. That's the whole ballgame right now.
And there's a number in here that should scare and excite you in equal measure. Jellyfish, a company that makes tools for software developers, put out a report based on agentic coding tool usage across 276,000 of its customers. Claude Code and other tools built on Anthropic's models are the most widely used, responsible for nearly a third of all proposed code changes submitted for review. A third. And spending on those tools has grown 10x since the start of the year, now equal to about 2 percent of total engineering cost. Nicholas Arcolano, the chief scientist at Jellyfish, said it plainly: "Even if you're saying 2% is still small, that 10x growth is growth that most CFOs didn't plan for." And then the key line: "it's not enough to just spend a bunch. You now have to optimize it, make it work."
That right there is the pivot of this entire era. We've gone from "look what the model can do" to "show me the ROI or the budget's getting cut." And Anthropic, smart as they are, saw that wind blowing and built the weathervane. Because here's the strategic genius of it. These reporting tools aren't really a feature. They're a retention moat.
Nikolas Albarran, the AI product lead at Jellyfish, said something I want you to hear. He said OpenAI's Astra model this summer looked like a real threat, and on X "everyone in the Valley is like, 'Oh, Codex is better.'" But in the actual usage data, Anthropic barely slowed down. Why? Because, in his words, "Anthropic was just there with great enterprise motions, and they had an incumbent weight of their own — this stuff just worked and worked well, and a lot of people just put a lot of effort into standardizing workflows around it. And there's not a big enough reason to switch."
That's the oldest story in enterprise software, folks. Once you've got your hooks in, switching costs do the selling for you. Albarran pointed to all the labs investing in forward-deployed engineering teams, these are consultants they embed inside your company to get you up and running. And he called these new reporting tools a kind of "FDE-lite," a lighter version of that hands-on help. "Once you're in, it's going to be harder and harder to switch because you've spent the effort." The cost to switch, he said, is not zero.
So here's my read for the builders listening. We spent the last two years arguing about which model is smartest on which benchmark. That argument is becoming a bar-stool argument. The real competition now is over who makes it cheapest to prove value and most painful to leave. Anthropic just made a very deliberate move to be sticky, not just smart. And given that most of the Fortune 500 isn't yet seeing the payoff, the company that helps the nervous CFO see the payoff — even a little — is the company that keeps the contract. Watch for OpenAI and Google to ship their own versions of this inside a quarter. The ROI dashboard is about to become table stakes.
Now, that's the demand side. Let me flip you over to the supply side, because the cloud giants reported earnings this week and the numbers are frankly absurd, in a way that's worth understanding.
Microsoft posted its tenth straight quarter of record-breaking revenue. Ten in a row. This quarter: 92.3 billion dollars, up 19 percent year over year. And it's profitable — net income of 36.6 billion, up 32 percent. The engine is Microsoft Cloud at 63.9 billion. And here's a small thing that's actually a big thing: Microsoft, for the first time ever, told us Azure's actual revenue in dollars. 33.2 billion for the July-to-September quarter, up 48 percent. They've never done that before. For years they'd only give you a growth percentage, some vague milestone talk. Now they're putting the dollar figure on the table so investors can line it up directly against Amazon Web Services and Google Cloud.
Why disclose now? Because the number is good enough to brag about. Azure is now Microsoft's largest and fastest-growing product line, roughly 36 percent of total revenue. And here's the race angle. Alphabet, reporting the same day, said Google Cloud brought in 30.7 billion. So Azure's slightly ahead. But — and this is the part that matters — Google Cloud roughly doubled its revenue from a year earlier, growing about twice as fast as Azure. So Microsoft's bigger today, Google's climbing faster. The transparency isn't generosity. It's a competitive flex with a clock on it.
Here's the detail I want you to hold onto, though. Microsoft said its "commercial remaining performance obligation" grew 73 percent to 677 billion dollars. That's a mouthful, so let me unpack it. That's future revenue already committed under existing contracts — a big chunk of it the AI labs promising to buy cloud capacity. 677 billion in the backlog. But to deliver on that, Microsoft is spending like crazy. Capital expenditures topped 45.2 billion in the quarter. Another all-time record. So the backlog is enormous, and the spending to fulfill it is enormous. That's the AI economy in miniature: gigantic promises on one side, gigantic build-out costs on the other, and everybody's praying the promises are real.
And here's the piece that stitches the two companies together into one nervous little marriage. Microsoft, in a regulatory filing, disclosed it recorded 10 billion dollars in revenue from OpenAI in the quarter ended September 30th. That's nearly 11 percent of Microsoft's total quarterly revenue. From one customer. For the full 2026 fiscal year, revenue from OpenAI hit 24.1 billion, and at the current pace Microsoft's pulling 40 billion a year from OpenAI-related arrangements. Now this is background we've touched before — the two companies have been loosening their once-exclusive partnership since late 2025, OpenAI's allowed to work with other cloud providers now, and in exchange OpenAI committed to buying an additional 250 billion in Azure services. Microsoft's ownership stake in OpenAI ticked down slightly, from about 25 percent in June to 24 percent in September.
Here's what I want you to take from that. These two companies compete for the same customers — they're rivals in the agent and model game — and yet 11 percent of Microsoft's revenue runs through OpenAI, and OpenAI runs a mountain of its compute on Azure. They're competitors who can't afford to let go of each other. If you're a builder choosing a stack, understand that the "neutral platform" you're building on is also in business with, and in competition with, the model company you're renting. The lines are not clean. They never were.
Now let me connect that cloud story to the other side of Google's earnings, because it's the quiet warning inside all the record numbers.
Alphabet reported 127.6 billion in revenue, up 25 percent. Big number. But the stock fell 8 percent after the results. Why? Because the core business, search advertising, is decelerating. Google Search revenue grew 12 percent year over year — which sounds fine until you see the trend. It was 17 percent the quarter before, 19 percent the quarter before that. Three straight quarters of slowing. That's a deceleration you can set your watch to.
So what's Google's answer? Put ads inside the AI. Philipp Schindler, Google's chief business officer, told investors they're "scaling high-quality ads within AI overviews in AI mode to more users and queries," and — here's the news — "taking our monetization footprint global for the first time." They're expanding ads in AI Mode into international markets. And they claim it's working. Schindler said click-through rates for sponsored product listings in AI Mode are already comparable to organic listings.
Sundar Pichai threw out some numbers that are genuinely interesting for anyone building search-adjacent products. Search queries grew at their fastest year-over-year rate in five years. Queries in AI Mode are three times longer on average than traditional searches. One in five AI Mode searches globally involves voice, images, or documents. And Google's doing over 50 billion visual searches a month.
Here's my read, and it's a little skeptical. Google is in a race against its own success. The thing people love about AI answers — you ask, it tells you, you don't click anything — is the exact thing that threatens the ad model that pays for all of it. For 20 years the deal was: Google gives you ten blue links, you click one, somebody pays. AI Mode collapses that. The answer is right there. So Google has to re-learn how to sell ads inside a conversation, and they have to do it before the old search cash cow slows down too much. The 8 percent stock drop is the market saying, we're not sure you can thread that needle in time. For builders, the lesson is the same one we keep circling: the agentic web is eating the click, and everybody who made money on the click — publishers, affiliate shops, SEO firms — has to figure out what they are in a world where the machine answers first.
And speaking of that agentic web, Google made a developer move the same week that's worth two minutes. They launched something called the Connected Apps Platform. Up to now, connected apps let you link a third-party service to Gemini or Search — connect Canva to design a flyer, connect YouTube Music to build a playlist. Now Google's opening the doors so any business or developer can submit their own connected app and monitor how it performs. You integrate with a cloud-hosted MCP server — that's the Model Context Protocol, the plumbing that lets agents talk to tools — register through the Google Connected Console, and your app can show up across Search and Gemini. After it's live you can track traffic, latency, error rates.
Sarah Perez made the comparison that lands: this is like SEO for the agentic web. Remember how a whole industry grew up around getting your website to rank in Google search? This is the opening bell for the same scramble, except now you're optimizing to get your service picked by the AI when a user asks it to do something. Google already has over 50 connected apps in Gemini — Adobe, Peloton, Experian — and they're coming to Search over the next few weeks. If you run a consumer service, the distribution question is shifting from "how do I rank" to "how do I get the assistant to call my tool." That's a new game, and the early movers are going to learn the rules while the rest of us are still reading the announcement.
Alright, let me shift gears hard, because the next one isn't about money, it's about power. The kind with a camera on it.
We've talked about Flock before on this show — a couple weeks back I walked you through the CEO's slippery TechCrunch interview while police surveillance abuse scandals were piling up. There's a fresh development worth your time, and I want to flag the timing honestly: the underlying protest wave and some of this reporting traces back through the summer and into recent weeks, but there's genuinely new material here, and it's ugly.
Here's the setup. A Florida county — St. Lucie County — did a countywide takedown of Flock automated license plate reader cameras. During the sweep, they found a clump of unpermitted cameras. Mystery cameras. Twelve of them. Five were eventually claimed by city law enforcement. But seven — seven cameras — nobody would own up to. And the county has now decided it won't investigate who owned them. The cameras have since disappeared. A resident named Hannah Galindo, 31, who'd petitioned to get the Flock cameras removed, told The Washington Post she was worried "that data could be anywhere." And when she heard the county was dropping it, she said: "It sounds like when you do something wrong, and you don't want to get caught, and you kind of just want to sweep it under the rug."
The county's spokesperson, Erick Gill, said the goal was to remove the cameras, not identify the owners, which is why they gave owners 30 days to take them down without facing any questions. "Our goal was to get the cameras out of the right-of-way within 30 days," he said. "We accomplished that goal." Great. Mission accomplished. Who was watching you, and what did they do with it? Shrug. Not our department.
Now here's a detail that tells you how the money works. Over three years, the Port St. Lucie Police Department's Flock contract ballooned from 22,000 dollars to 470,000 dollars by 2025. From twenty-two grand to nearly half a million in three years. That's the growth curve of a surveillance network getting its hooks in — same sticky dynamic we talked about with enterprise AI, just pointed at your license plate instead of your codebase.
And Flock's own response is something. A Flock spokesperson told Ars Technica that Flock does know who owned those cameras — they just won't tell the residents. Quote: "The cameras have been removed and belonged to several different customers. Flock does not share specifics regarding individual agency contracts, camera counts, locations, or deployments." So the company knows. The residents don't get to know. Think about that arrangement. A private company tracking your movements, and when your own neighbors ask who was watching, the answer is: that's proprietary.
But here's the part that got me, and the part I think is genuinely new and genuinely telling. 404 Media got leaked audio from a private police conference where Flock's CEO, Garrett Langley, told a room full of cops that the nationwide anti-Flock protests "are not real." He floated a disproven conspiracy theory that Russian bots fueled the viral anti-Flock posts. He said, and I'm quoting the leaked audio here: "Social media chatter on Flock is down about 95 percent from July. It went down 50 percent last week when Ukraine bombed a Russian data center with a drone. Literally 50 percent overnight." And then: "it's anonymous bots, and so it is not real... when I look at protests across the country, it is the same people with the same signs and the same message. That's not normal. That's not organic at all."
Let me just sit in that for a second as a guy who grew up around real protests, real kitchen-table arguments, real people showing up with real cardboard signs. The CEO of a surveillance company is telling the police — the people deploying his cameras — that the citizens objecting to being surveilled are fake. They're bots. They're paid. They're not real. That is a remarkable thing to tell armed government agents about the people they serve. Your constituents are complaining, and the vendor's advice is: ignore them, they're not real.
And it's demonstrably not just bots. The ACLU built a toolkit to help communities "DeFlock" — their word — and on the very day that leaked audio came out, the ACLU updated its "Get the Flock out" resource, warning about the lack of oversight for Flock's camera database. Real organization. Real lawyers. Not a Russian server farm.
Now here's where it gets to be a builder story, and a national security story, and it's the part I most want you to walk away with. Flock told the Texas Tribune in August that there are about 120,000 of its cameras used by law enforcement nationwide. But a cybersecurity researcher named Joshua Michael found something different. He discovered a Flock access token — an exposed credential — that let him retrieve device locations without logging in. He notified Flock of the vulnerability. And he also built a map of all the devices. Matt Busel, founder of a software studio called Tensorust, analyzed that map and found not 120,000 devices — but 214,073. More than two hundred thousand. Nearly twice what Flock says publicly. And the roughly 100,000-camera gap, Busel believes, is Flock's private camera network — operated without regulatory oversight. In his words: "It is bigger than anything Flock says out loud."
And how did Flock respond to a good-faith security researcher who found an exposed credential and reported it? They sent him a cease-and-desist. They demanded he pull down the map, the data file, and his research paper. They demanded he name anyone he shared the data with. They threatened to sue for injunctive relief plus statutory and punitive damages. And — this is the kicker — Michael points out that in the same email where Flock branded him a "malicious actor" and said it had notified law enforcement, the company conceded it found no indication of unauthorized access to its data when it closed the issue back in December. Which, as Michael says, is exactly the monitoring failure his research was warning about. He confirmed the gap, and they confirmed it right back while calling him the bad guy.
Michael's statement is worth hearing because he frames it as national security, and I think he's right: "My research argued that Flock Safety could not detect exports of their data and that because they hold data on nearly every American and their pattern of life, they are a national security threat to our country and people." He's self-funding his legal defense, taking donations, and Proton is hosting a mirror of his map. He says: "I will not put Flock's quarterly profits ahead of the security of the United States."
Here's my take, builder to builder. If you run a company and a researcher hands you an exposed credential they found lying around, you have two paths. Path one: thank them, fix it, maybe pay them a bounty, come out looking responsible. Path two: lawyer up, call them a malicious actor, and try to delete the evidence. Flock picked path two. And path two never, ever makes the problem smaller. It makes the researcher dig in — Michael says flatly, "I'm not backing down" — and it tells every other researcher that your company is the kind that shoots the messenger. When you're sitting on a pattern-of-life database covering nearly every American, that posture isn't just bad PR. It's a genuine argument that the concentration of this data, in a company that can't detect when it's been exfiltrated and won't be transparent about its own footprint, is a liability for the whole country. The protests, Mr. Langley, are real. The cameras are realer. And the hundred thousand you don't talk about out loud are the realest thing of all.
Let me connect that to one more surveillance-adjacent story, because the week had a theme whether anyone planned it or not: the data always ends up somewhere it shouldn't.
A hacker going by the handle "Department of Peace" published personnel data on more than 40,000 U.S. Department of Homeland Security employees. Names, job titles, work emails, cell numbers, office locations — including border and immigration agents, cybersecurity staff, and senior officials. The hacker claims they pulled it from an Okta single sign-on system DHS uses to let employees log into its central network. More than half the dataset — about 21,000 records — relate to Customs and Border Protection. Another 10,000 to ICE and its Homeland Security Investigations unit. Just under a thousand at CISA, which — I'll note the irony — is the federal cybersecurity defense agency. The leaked records appear to have been exfiltrated around mid-August. That's about one-fifth of DHS's entire 220,000-person workforce. Okta, for its part, said: "Okta is secure, and we have no evidence our data or systems were compromised."
Now, why does this belong in a show for founders and builders, beyond the obvious "patch your SSO" lesson? Because the dataset doesn't just expose people — it exposes the tech stack of the surveillance state, and the vendors in it. The leaked data shows user-group memberships in DHS's Active Directory. Most employees had access to Microsoft 365. Fine, normal. But then it gets specific. Around 4,500 users, including ICE agents and deportation officers, were in a group called "Headspace_users" — the meditation and therapy app. Headspace told TechCrunch it hasn't provided services to DHS or ICE since 2023, and the old contract was for mental-health software. Dozens of other user groups reference Penlink, a web-surveillance company whose technology it claims can track location data of hundreds of millions of mobile devices. One group references Penlink PLX, which captures and analyzes real-time intelligence from phone calls, texts, web activity, and social media. More than 200 CBP users are tied to Mobile Fortify, a facial-recognition app. And several employees, including senior DHS officials, had ChatGPT access, plus groups tied to investigations into the China-backed hacking crews known as Salt Typhoon and Volt Typhoon.
Here's the thing to notice. A single misconfigured or breached identity system became a map of every surveillance vendor the agency does business with. Researchers and transparency advocates are already saying the email addresses and phone numbers are gold for FOIA requests — Ryan Shapiro of Property of the People said they're "indispensable tools" that make the difference between a request that gets stonewalled and one that gets results. DDoSecrets' Emma Best called the leak a "catalyst for investigations."
So two lessons braided together. For the builder: your identity layer is your whole security story. If someone gets into your SSO, they don't just get passwords, they get the org chart, the tool inventory, the vendor list — the complete shape of your operation. And for everybody: we are living in a moment where the surveillance apparatus and the people running it are both getting surveilled, by hackers and by their own leaks. The cameras point every direction now. That's not a comfortable equilibrium. It's just the one we've got.
Okay. Let me lighten the load, because not everything today is dystopian. Some of it's just the normal churn of an industry running hot.
Quick hits on the enterprise land-grab, because the partnership announcements came fast this week and they all tell the same story: everybody's racing to be the layer your agents live in. OpenAI expanded its partnership with Oracle to bring business apps and AI agents into ChatGPT. It expanded with ServiceNow, pushing OpenAI models into ServiceNow workflows and ServiceNow workflows into ChatGPT and Codex. There's a case study on Legora building what it calls an agentic operating system for legal work on OpenAI's models. There's Retell, powering 100 million monthly phone calls with OpenAI's voice models, claiming a 1.6-million-dollar annual revenue run rate per employee — that per-employee number, if real, is the kind of leverage that makes traditional staffing models look like horse-and-buggy stuff. And Anthropic expanded its partnership with the GSA — the government's buying arm — to give federal, state, and local agencies a 40 percent discount on Claude, estimated to save the government a billion dollars. Roughly 500 government organizations now use Claude, and weekly government active users are up about eightfold since February.
Notice the pattern in all of it. Legal, government, customer service, enterprise workflow. Nobody's selling "look how smart the model is" anymore. They're selling "we plug into the software you already run and we'll show you it saved you money." Same theme as our deep dive. The whole industry has turned toward proving value inside the systems people already use. The model is becoming the engine, and the engine isn't the product anymore. The integration is.
One more fun one for the developers, because Simon Willison flagged it and it made me smile. He wrote this week that he thinks October represents another inflection point for coding agents, and his specific evidence was that Claude Opus 5.5 can now build video games that are actually fun. That's a lower-stakes benchmark than SWE-bench, but honestly it might be a more honest one. "Fun" is hard. Fun requires taste, pacing, feel — the squishy human stuff. If the coding agents are clearing that bar, it says something about where we are. I'd just pair it with the Jellyfish data from earlier: the agents are genuinely good enough to write a third of your proposed code changes, and genuinely good enough to make a game you'd actually play. The capability is real. The open question — the one the whole show's been circling — is whether the spend tracks the value. For a weekend game jam, who cares. For a Fortune 500 engineering budget, that's the only question that matters.
And since we're on the consumer edge, two hardware notes I'll keep short because, fair warning, some of this is older product news that resurfaced. Apple launched a touchscreen MacBook Pro — starting at 2,899 dollars, with a stabilizing hinge and a Dynamic Island-style cutout, after years of Steve Jobs famously calling the idea "ergonomically terrible." Times change, so do companies. And over in glasses-land, Xreal launched the Aura, a wired smart-glasses headset powered by Android XR, tethered to a little processing puck, part of Google's partner strategy to chase Meta in the face-computer race. Google's doing with Android XR what it did with Android phones — rally the partners, flood the zone. Whether anybody actually wants a puck in their pocket is the 1,099-dollar question.
Let me bring it home.
If you line up today's menu — Anthropic's ROI dashboards, Microsoft finally putting a dollar figure on Azure, Google scrambling to monetize AI answers as the old search cash slows, the partnership land-grab, even Flock's surveillance empire that's twice as big as it admits — there's one honest sentence underneath all of it. The era of impressive is ending. The era of accountable is starting. For two years the question was "can it do the thing." Now the question is "is the thing worth what it costs, and who's checking." Anthropic is betting the company that helps you check wins. The cloud giants are betting the backlog is real enough to justify 45 billion a quarter in spending. Google's betting it can re-learn advertising inside a conversation. And a surveillance company is betting it can tell the cops the critics aren't real — which is the one bet on this whole menu I'd put money against.
If you're building, the move is clear. Instrument your value. Make the ROI legible to the nervous person holding the budget. Because the money's getting cautious, the receipts are coming due, and the companies that make the receipts easy to read are the ones that'll still be standing when the drunken-sailor spending sobers up.
That's the spread for today. Thanks for pulling up a chair with me. I'm Tony DeLuca, keep your credentials off the public internet and your skepticism switched on, and I'll see you tomorrow on Barely Possible.