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. Pull up a chair. Today we've got a menu that runs from a hundred-fifty-million-a-month compute lease to a guy who decided to take a percussion massage gun to his own eyeballs. Yeah. We'll get to the eyeballs. But the spine of today's show is something quieter and bigger than any single deal, and it's this: who actually pays the bill when AI scales. Not the marketing bill. The literal bill. The power, the water, the workers, the chips, the rent on a launch pad. We've spent weeks on this show talking about talent moving between labs and models leapfrogging each other on benchmarks. Today I want to follow the money down to where it touches the ground, because that's where founders make and lose real decisions. Let's have at it.
Let me start with the number that made me put my coffee down. SpaceX inked a compute deal with Reflection AI, which is an open-source AI lab. This is reporting from Kirsten Korosec at TechCrunch. Here's the shape of it. Reflection AI is going to pay a hundred and fifty million dollars a month, starting July first of this year and running through 2029, for immediate access to Nvidia's latest GB300 chips and the supporting hardware, sitting inside SpaceX's Colossus 2 data center near Memphis, Tennessee.
Let that sit for a second. A hundred and fifty million a month. Multiply that out over three-plus years and you're north of five billion dollars in committed spend, from a lab whose entire pitch is open-weight models. And the landlord here is SpaceX, running a data center called Colossus 2. So let me untangle a couple of things for you, because the names matter.
What this tells you about the AI economy in the middle of 2026 is that compute is now a lease business, and the lease terms are brutal. Reflection isn't buying a building. They're not buying chips outright. They are renting frontier compute on a multi-year fixed commitment, the same way an airline leases jets it can't afford to own. A hundred and fifty million a month is rent. It comes due whether your model trains well or not, whether your revenue shows up or not, whether the open-weights thesis pays off or not. That is an enormous amount of fixed cost for a company giving its core product away.
Now, I'm not here to tell you Reflection is doomed. The bet underneath this is that open-weight frontier models become valuable enough — through services, through enterprise deployment, through being the model everybody post-trains on — that five-plus billion in compute pencils out. That's a real thesis. We've talked on this show about the application-layer value pattern, the Elastic-Deductive AI deal from a few days back, where the money was in the thing built on top, not the model itself. This is the opposite end of the barbell. This is somebody making a colossal infrastructure bet at the very bottom of the stack and hoping the value flows up to them anyway.
For you, the builder, here's the takeaway that actually matters. When a lab signs a hundred-fifty-million-a-month lease, that cost has to be recovered somewhere, and "somewhere" is increasingly the price you pay per token, or the terms you get when you depend on them. The compute crunch we keep circling isn't abstract. It's a landlord-and-tenant economy now, and the rents are setting the floor under everything you build on top. If your product assumes cheap inference forever, look at who's paying a hundred and fifty million a month and ask yourself who eventually covers that.
And notice the shape of the players. The compute landlord is SpaceX. The chips are Nvidia's. The tenant is an open-source lab. The picture of who owns the picks and shovels in this gold rush keeps getting more concentrated, even as the model layer supposedly gets more open and more competitive. Hold that thought, because it comes back.
Now, that hundred-and-fifty-million-a-month rent doesn't conjure itself out of thin air. It runs on power. And power is where the bill gets really ugly, really fast.
Microsoft and Chevron are planning one of the largest gas-powered data center projects in the US. This is Tim De Chant at TechCrunch. Microsoft inked a twenty-year power purchase agreement with Chevron, locking in decades of carbon emissions from a new natural gas power plant to feed data centers. Twenty years. Read that again. In an era where everybody's quarterly slide deck has a net-zero pledge on it, Microsoft just signed a twenty-year commitment to a brand-new gas plant. That's not a hedge. That's a conviction bet that AI demand is going to be so big, so durable, that you build fossil-fuel generation and run it for two decades.
Here's what I want you to sit with. The clean-energy story that the hyperscalers told for years — the solar farms, the wind, the carbon offsets — that story is quietly buckling under the weight of AI load. When you need gigawatts on a deadline, and renewables can't deliver firm power fast enough, you go to gas. And once you've signed a twenty-year PPA, the emissions are baked in. You can buy all the offsets you want; the plant is built and it's running.
And it pairs darkly with another item from the same reporter. Nvidia announced a new cooling system that cuts water use inside the data center, and the headline everybody ran with was "Nvidia tackles AI's water problem." But Tim De Chant's piece makes the sharper point: cutting water inside the data center does almost nothing about AI's biggest water use, which is the fossil fuel power plants generating the electricity in the first place. Power plants are thirsty. You cool a turbine with water. So Nvidia trims the water at the rack and then the gas plant down the road drinks a reservoir to keep the lights on.
I'm not doing an environmental sermon here. I'm doing an honesty audit. Because there's a version of the AI sustainability story that's mostly press releases — the efficient cooling, the green data center — and there's the real ledger, which is twenty-year gas contracts and thirsty power plants. As a builder, the reason you care isn't guilt. It's that energy is becoming the binding constraint on this whole industry, and the companies that locked in cheap, dirty, reliable power early are going to have a structural cost advantage over the ones that didn't. The water and the carbon aren't side issues. They're the cost structure wearing a costume.
Now let's move from the power plant to the factory floor, because the same question — who pays — has a human answer that's a lot harder to look at.
GM installed robots at its flagship EV factory after laying off thirteen hundred workers. This is Jeremy Hsu reporting. The framing in the piece is blunt: the US autoworkers union is warning about robot automation as what they're calling a "dark factory" future looms. A dark factory, if you haven't heard the term, is a plant that runs with so few humans you don't need to turn the lights on. Lights-out manufacturing. The robots don't need to see.
And this isn't happening in a vacuum. Same day, same feed, we get Lucid laying off fifteen hundred workers — the second big cut of the year for them. The CEO says it's part of a plan to "simplify the company." Simplify. That's a word that's doing an awful lot of work in a sentence about fifteen hundred people losing their jobs.
Here's the pattern, and it's not subtle. GM lays off thirteen hundred, then installs robots. Not after. The sequence reads like cause and effect even if the company would phrase it gently. And the union is saying the quiet part: the automation isn't a complement to the workforce, it's a substitute. We have been told for years that AI and robotics would augment workers, free them up for higher-value tasks, all the comforting language. What the auto sector is showing us in real time is the less comforting version, where the headcount goes down and the robot count goes up and the official explanation is "simplification."
Why does a software founder care about robots on a car line? Because this is the leading edge of a story that's coming for knowledge work too. The auto industry is just further along because the capital equipment is more mature. When people ask whether AI takes jobs or makes jobs, the honest answer from a flagship EV factory this week is: it's taking these specific jobs, right now, and the company is being pretty matter-of-fact about it. If you're building AI products that automate work, that's your product working as designed. Just don't kid yourself about what the design does. The person who paid for that efficiency gain is the person who used to stand at that station.
That's the human side of the bill. Now let me show you the security side of it, because here's where it gets genuinely interesting for anyone building agents.
Sam Altman posted that the full version of GPT-5.5-Cyber is here, with what he calls state-of-the-art performance on CyberGym, which is a cybersecurity benchmark. And here's the framing that caught my ear: he says OpenAI wants to help all companies be secure, working with the US government and the security ecosystem. And then this line — they've got tools called Patch The Planet and Codex Security that will, in his words, help solve security problems instead of just finding them.
Solve, not just find. That's the pivot. For a while now the AI security story has been about models that are scary good at finding vulnerabilities. We covered, a few days back, the whole saga around export controls and the claims about Anthropic's model and cyber capability — the reporting from Lorenzo Franceschi-Bicchierai on the long history of export controls that never actually stopped anyone. The fear in all of that is offense: an AI that compresses weeks of expert exploit research into hours. What Altman is putting on the table here is the defensive mirror image — a model that doesn't just find the hole, it patches it.
Now I want to be careful and skeptical here, because this is a CEO post, not a peer-reviewed result. "State of the art on CyberGym" is a benchmark claim, and benchmark claims are marketing until somebody independent confirms them. But the direction is the thing to watch. If you're building anything that touches security — and increasingly every agent that can run code touches security — the question of whether the same model that finds the bug can be trusted to fix the bug is going to define a whole product category. And the explicit "working with the USG" framing tells you OpenAI is positioning this as the responsible, government-aligned counterweight to the export-control panic we've been living in. They want to be the lab that hands the patches to Washington, not the one Washington puts a kill switch on.
Which brings me to the deep dive, and it's the story I think every founder in this audience should chew on for a few minutes. Anthropic says Claude may want to see your ID.
This is Zack Whittaker at TechCrunch, and here's the core of it, straight from the reporting. Claude's chatbot may ask to verify your age and identity — quote — "in certain circumstances," such as with a passport or driver's license, according to a privacy policy change. Anthropic updated its privacy policy to say that, under some conditions, Claude can ask you to prove who you are with government ID.
Let me slow down on this, because it's small in the news cycle and large in what it signals.
Start with the surface. An AI chatbot asking for your passport. Five years ago that sentence would've sounded insane. You went to a website, you typed, the thing answered. The idea that the model would card you at the door like a bouncer — that's a genuine shift in the relationship between you and the software. And it didn't come from nowhere. It comes from the same pressure cooker we've been reporting on all week. The White House has been leaning hard on Anthropic over safety and jailbreaks — there's reporting from Wired about the White House wanting Anthropic to block all jailbreaks, which experts say may not even be technically possible. When you can't perfectly control what the model outputs, what's your fallback? You control who's allowed to ask. Age and identity verification is the pressure valve when output filtering hits its limits.
Now here's why this matters to you specifically, the builder. Think about what ID verification does to a product. It is the single highest-friction thing you can put in front of a user. Every checkout flow, every signup funnel, every growth team on earth spends its life trying to remove steps, not add them. And here's Anthropic adding the heaviest possible step — show us a government document — to a consumer product. They are not doing that because they want to. They're doing it because the regulatory and liability environment is forcing identity to the front of the line.
And the second-order effect is the one I'd circle in red. The moment a major AI lab normalizes "the assistant may ask for your ID," it becomes the template. Regulators love a precedent. Once Claude is checking passports in certain circumstances, the question every other consumer AI company faces shifts from "should we ever do this" to "why aren't you doing this yet." If you're building a consumer-facing AI product right now, you should be designing as if age and identity verification is a feature you will eventually be required to ship, not one you get to opt out of. Budget for it. Architect for it. Because the big lab just made it respectable.
Third, and this is the uncomfortable one — the privacy ledger. When a chatbot collects your passport, that data now exists somewhere. We'll get to a data breach in a minute that should make your skin crawl about exactly this. The same companies asking to verify your identity are, by definition, building a honeypot of identity documents. Every "show us your ID" feature is also a "now we hold your ID" liability. We covered the UK using flawed facial age-estimation on asylum seekers a couple days back — the broader move toward age and identity checks is everywhere, and the technology underneath it is consistently less reliable and more dangerous than the policy language admits.
So here's how I'd frame the whole thing. Anthropic asking for ID looks like a privacy-policy footnote. It's actually the moment the AI industry stopped pretending these are anonymous tools you just talk to, and started treating them like regulated services with a velvet rope. For founders, that's not a side detail. That's a change in the cost of doing business in consumer AI, and the bill — verification vendors, compliance, breach liability — lands on you. The same question keeps coming back: when the system scales, who pays? Here it's the user handing over a passport, and the company holding the bag if that passport leaks.
And it leaks. Let me show you how.
Tata Electronics — a major tech supplier to Apple and Tesla — confirmed a data breach. This is Jagmeet Singh at TechCrunch. The reporting notes the incident comes as Tata Electronics expands its role in global technology supply chains. So here's a company that's becoming a bigger and bigger node in the supply chain for the most valuable hardware companies on earth, and it just confirmed somebody got in.
The connective tissue to the ID story is right there. We are simultaneously building a world where more companies collect more sensitive data — identity documents, supply-chain secrets, manufacturing details — and a world where breaches at major suppliers are routine enough to be a Monday news item. You can't separate the "show us your ID" trend from the "supplier confirms breach" trend. They're the same coin. The more identity and sensitive data we centralize to satisfy safety and compliance, the bigger the prize for whoever breaks in.
And on the breaking-in theme, there's a hardware story worth a quick beat. A new unpatchable flaw in Apple chips opens the door to an iPhone jailbreak. This is Lorenzo Franceschi-Bicchierai again. A European offensive cybersecurity company called Paradigm Shift released details of a flaw, and a technique to exploit it, that lets people unlock and break into older iPhones. The word that matters is "unpatchable." When a flaw is in the silicon, Apple can't push a software update to fix it. The affected hardware stays vulnerable for its whole life.
Now it's older iPhones, so don't go panic-selling your phone. But the principle is the one I keep coming back to: the security guarantees we build products on are softer than the marketing says. "Patched" is a promise about software. Silicon-level flaws are a reminder that some of the foundation is poured concrete you can't re-pour. If your threat model assumes the device is a trusted endpoint, this is a small note that says: not always, not forever.
Let's shift from breaking into things to a different kind of building — the creative side. Google DeepMind is betting seventy-five million dollars on AI's future in Hollywood with an A24 deal.
This is Dominic-Madori Davis at TechCrunch. Google DeepMind and A24 — and if you don't know A24, they're the prestige indie studio behind a pile of films cinephiles obsess over — are teaming up to build AI filmmaking tools. Seventy-five million dollars.
What's interesting here isn't the dollar amount, which is modest by AI standards. It's the pairing. DeepMind has the video-generation models. A24 has taste, and credibility, and a brand that means "we don't make slop." That's a deliberate marriage. The biggest knock on AI video has been that it produces soulless, uncanny garbage — slop, as everybody calls it now. By partnering with A24 specifically, Google is trying to buy its way past the slop reputation and into the room where serious filmmakers actually work.
For builders in the creative-tools space, the signal is this: the frontier of AI media isn't the raw generation model anymore. Everybody has a video model. The frontier is taste, integration, and the workflow that working professionals will actually adopt. Google didn't need A24's compute. It needed A24's credibility and A24's understanding of how a real production pipeline functions. If you're building AI tools for any creative field, that's your lesson — the model is table stakes, and the defensible thing is being genuinely useful inside an expert's existing process. The same thing we keep finding in coding tools, in enterprise tools, now in film.
Now a couple of quick ones that round out the picture, and then I've got two human stories I can't resist.
Groq confirmed a six hundred and fifty million dollar raise, and is re-staffing after Nvidia's twenty-billion-dollar — and I love this phrase — "not-acqui-hire" deal. This is Julie Bort. So the backstory: Nvidia did one of those deals where they take the talent and license the tech without technically buying the company, leaving the original company a hollowed-out shell with a check. The question Julie poses is the right one: what does a company do after one of those? And Groq's answer is — raise fresh money, lean into its neocloud business, hire new executives, and keep going. Six hundred and fifty million says somebody believes there's a real business in being an alternative compute provider even after Nvidia took a twenty-billion-dollar bite. It also tells you how the chip world is consolidating: Nvidia doesn't even have to buy you to absorb your best people. They just write a not-acqui-hire check. Tie that back to the SpaceX-Reflection lease, the Microsoft gas plant — the infrastructure layer keeps concentrating around a small number of giants, and everyone else is renting, raising, or rebuilding around them.
Amazon is testing Alexa+ in India with Hindi support. Ivan Mehta's reporting. Amazon's pushing its new conversational AI assistant, Alexa+, into India and inviting users to test a Hindi-language version. Straightforward, but worth a nod for one reason: the next billion AI users aren't English speakers, and the companies racing to ship real, native-language conversational AI into India and similar markets are playing for a much bigger prize than the crowded English-language assistant space. If you're a founder thinking about where the open territory is, it's not another English chatbot. It's the languages the frontier labs haven't bothered to do well yet.
And one for the developers — Simon Willison released the first release candidate for sqlite-utils version 4, adding a migrations system and support for nested transactions. Small, unglamorous, genuinely useful. After a whole episode about hundred-fifty-million-a-month compute leases and gas plants, there's something grounding about a solid tool that makes your database migrations less painful. Not everything has to be a civilization-scale bet. Sometimes it's just a clean migration system, and that's a good day. Link's in the show notes.
Now. The two stories I promised.
First, the one that's just genuinely fun. Pikes Peak this year was a battle of propulsion, and a 1,250 horsepower hybrid Corvette shattered the production car record on that mountain. Tim Stevens at Ars Technica. Pikes Peak, if you don't know it, is this insane high-altitude hill climb where the air gets thin near the top and engines that breathe air start to gasp. Which is exactly why electrification wins up there — electric motors don't care about thin air. They make the same torque at the summit as at the base. So you've got a hybrid Corvette, twelve-fifty horsepower, setting a production record, and an EV winning outright. The mountain that used to punish engines now rewards the cars that don't depend on oxygen. It's a neat little proof that sometimes the new technology wins not because of hype but because of physics. Thin air doesn't lie.
And then. The eyeballs. I told you we'd get here. A man used a massage gun on his tired eyeballs, and it went exactly as well as you would expect. Beth Mole at Ars Technica. The man had retinal tears and bruises from squishing his own eyeballs with a percussion massage gun — one of those things you're supposed to use on a sore calf. He pointed it at his eyes. Retinal tears. I bring this to you not to be cruel but as a public service and, frankly, as a metaphor. We are living through a moment where everybody's got a powerful new tool and an irresistible urge to point it at something it was never meant for. The massage gun is great. On your shoulder. The man pointed a precision impact device at the most delicate tissue in his body because he was tired and it was right there. Be the person who reads the instructions. In your products, in your life, and especially before you aim anything percussive at your own face.
Let me bring this home. Everything today rhymed, even the stuff that looked unrelated. A lab signs a hundred-fifty-million-a-month compute lease. Microsoft locks in twenty years of gas. GM swaps thirteen hundred workers for robots and calls it nothing. Anthropic starts carding users at the door, and a major supplier confirms the data leaks anyway. Underneath all of it is the same arithmetic we don't like to do out loud: AI at scale has enormous, concrete costs — in power, in water, in jobs, in privacy, in liability — and those costs don't disappear. They move. They land on a worker, a user handing over a passport, a reservoir behind a gas plant, a tenant paying rent on chips. As a builder, your edge is being honest about where the bill goes in your own product, because the people who pretend the costs vanish are usually the ones who end up surprised when they come due.
That's the show. I'm Tony DeLuca, this has been Barely Possible, and please — keep the massage gun below the neck. Talk soon.