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. Today we got a menu that tells you something real about where the money is moving in this AI business, and where it's hitting a wall. We're gonna talk about a compute deal so big it sounds like a typo, an index that just slammed the door on three of the most hyped companies on Earth, and a database company that doubled to ten billion dollars in eight months riding the vibe-coding wave. Plus a couple of stories about the bill coming due — the literal token bill, and the literal electricity bill. Grab your coffee, settle in, let's have at it.
Let me start with the one that made me do a double take, because it ties two stories together in a way that's almost poetic. There's a deal where Google is going to pay SpaceX nine hundred and twenty million dollars per month for compute. Per month. That's reporting from Sean O'Kane over at TechCrunch. Read that number back to yourself slowly. Almost a billion dollars a month, flowing from one of the biggest companies in the world to a rocket company, for computing power. In a statement, a Google representative described the deal as a result of unexpected demand for its recently launched AI products.
Now let's just sit with the shape of this for a second, because the shape is the story. Google — Google, the company that practically invented the modern data center, the company with its own custom silicon, its own global fiber, its own everything — is going outside to buy compute. And they're buying it from SpaceX. Elon's rocket outfit. Why does a rocket company have compute to sell? Because SpaceX has been building out enormous infrastructure for Starlink and its satellite operations, and apparently has capacity and power access that other people now desperately want. The phrase that matters here is unexpected demand. When a company the size of Google says its own buildout couldn't keep up and it had to go rent from a neighbor, that's not a press release flex. That's an admission. The hunger for compute right now is outrunning even the people who are best in the world at building it.
And here's the continuity thread, because we've been circling this all week. Yesterday I told you about the Waymo getaway car. The day before, Uber capping its coding agents at fifteen hundred dollars a month per employee per tool, with Simon Willison reading that cap as a tell about real value. And before that, the whole token efficiency drum we've been banging for two weeks. Every one of those stories is the same animal seen from a different angle. There is more demand for AI compute than there is supply of AI compute. When that's true, prices go up, caps come down, and even Google goes shopping at the neighbor's house. So when you hear nearly a billion a month from Google to SpaceX, don't file it under gee-whiz big number. File it under: the constraint is real, the constraint is physical, and the constraint is now showing up on the balance sheets of the biggest players alive.
Now here's the kicker that makes today's menu sing. On the same day we learn SpaceX has compute so valuable Google is renting it by the truckload, we also learn that SpaceX got told no by Wall Street. Let me explain.
There's a piece by Jeremy Hsu at Ars Technica with the headline that the S&P 500 rejected SpaceX, and in doing so, also blocked the door for OpenAI and Anthropic. Here's what happened. The S&P 500, that's the index that an enormous chunk of America's retirement money passively tracks. When you're in the index, all those index funds automatically buy your stock. It's a firehose of money from people who aren't even looking at your fundamentals, they just own the index. Getting in is a very big deal. And the committee that runs the index won't waive its rules for these companies. The core rule is about profitability — you need a track record of positive earnings to qualify. SpaceX won't get easy access to billions of dollars from passive investors. And because the same profitability gate applies, the report frames it as blocking the on-ramp for the unprofitable AI firms too — OpenAI and Anthropic in the same breath.
Let me tell you why this matters for you as a builder, because it's not just trivia about rich companies. This is the establishment financial system drawing a very old-fashioned line in the sand. The line says: we don't care how revolutionary you are, we don't care how big your valuation is, we want to see you actually make money before we let the schoolteacher's pension fund auto-buy your shares. In an era where the whole AI story has been growth at any cost, tokenmaxxing, burn the cash, change the world — the index committee is sitting there with its arms crossed going, show me the earnings. And these are companies, OpenAI and Anthropic, that are reportedly heading toward public markets. Anthropic, as we covered last week, filed confidentially. So the question of how the passive-money machine treats unprofitable AI is not academic. It's going to shape how much capital these companies can pull in and at what terms.
So put the two SpaceX stories side by side. On one hand, its compute is so precious Google's paying nearly a billion a month for it. On the other hand, the index says you're not profitable enough to come to the grown-ups' table. That's the whole tension of this moment in one company. Enormous strategic value, genuine physical scarcity, and a financial system that still wants to see a bottom line. Both things are true at once, and a smart founder holds both in their head at the same time.
Now let's move from the compute crunch to the bill that comes with it, because there's a piece that ties this whole thing together for anybody actually building product.
Rebecca Bellan at TechCrunch has a piece called The token bill comes due, with the subhead about the industry scramble to manage AI's runaway costs. And there's one line in there that I want to read you because it captures the vibe shift better than anything I could write. Quote: the whole conversation shifted from tokenmaxxing and go fast to we need guardrails, how do we control this? End quote.
That is the entire mood of 2026 in one sentence. For a couple of years the gospel was speed. Throw tokens at it. Use the biggest model for everything. Don't think about cost, think about capability. And now the bill is showing up, and the conversation in every serious company has flipped from go-fast to how-do-we-control-this. And we've got a perfect little data point on exactly how that control is being imposed, from Simon Willison again. He posted clarifying a detail people were getting wrong. The two hundred dollar a month subscription plans — those flat-rate all-you-can-eat plans — are no longer available to large enterprise customers like Uber or Brex. They now have to pay the full listed API price for all of their tokens.
Now think about what that means. The labs were effectively subsidizing heavy users with those flat plans. You pay two hundred bucks, you hammer the thing all month, and the lab eats the difference because they wanted adoption and market share. That was the go-fast era. And now, quietly, the big enterprise accounts are getting moved off the buffet and onto metered pricing where every token costs what it costs. That's the guardrail. That's the control. It's not some grand policy — it's a billing change. And it's exactly why Uber turned around and slapped a fifteen-hundred-dollar cap on its own people, which we talked about a couple days ago. The lab stops subsidizing, the cost flows downhill to the customer, and the customer puts up a fence.
Here's the builder takeaway, and it's not abstract. If you're building anything on top of these models, your unit economics are not stable. The subsidy that made your demo look cheap may not survive contact with scale. The companies that win the next phase are the ones who got token-efficient before they were forced to — who routed cheaper models to cheaper tasks, who built memory so they're not re-feeding the same context every turn, who measured cost per outcome instead of cost per token. If your whole business model assumes flat-rate buffet pricing forever, you are building on sand. The buffet is closing for the big eaters first, and it'll work its way down.
Let me shift now from the cost of building to a story about what AI building is actually producing, because there's a genuine success story in today's pile and it deserves its moment.
Supabase doubled its valuation to ten billion dollars in eight months. That's Julie Bort at TechCrunch. Supabase, for those who don't live in developer-land, is an open-source backend platform — databases, auth, storage, the plumbing under your app. Eight months ago they were valued at half this. Now, ten billion. And the reason given in the reporting is right there: Supabase is an example of an open source project becoming a fast-growing company, and it has greatly benefited from AI tools like Claude, Codex, and the other vibe-coding platforms.
Now why does this one matter beyond the headline number? Because it tells you where the value actually lands in the AI build-out, and it's not always where you'd guess. When millions of people start vibe-coding — describing an app in plain English and having an AI assistant scaffold it — they need a backend. The AI writes the code, but the code has to talk to a real database, real authentication, real storage somewhere. And Supabase positioned itself as the default answer. So every wave of AI-assisted app creation washes more users into Supabase. They're not selling the AI. They're selling the foundation the AI builds on top of.
That's the pick-and-shovel play, and it's a beautiful one. Everybody's mining for gold with these coding agents, and Supabase is selling the shovels — and in this case the shovels are open source, which got them adoption and goodwill, and then the hosted version is where they make the money. For a founder, the lesson is to ask: when AI dramatically lowers the cost of building software, what does all that newly-cheap software still need? It still needs a place to live. It still needs data, auth, deployment, monitoring. The infrastructure layer underneath the vibe-coding boom is quietly becoming very, very valuable, and Supabase just printed the proof at ten billion dollars.
Now I want to do our deep dive, and I'm going to anchor it on a piece that I think is the most consequential thing in today's pile for anyone who builds — the token bill story, but pushed all the way through to what it actually means for company strategy. Let me read you that core framing again and then let's really dig.
The piece quotes that shift — quote, the whole conversation shifted from tokenmaxxing and go fast to we need guardrails, how do we control this. End quote. I want to take that seriously as an operating reality, not a headline, because I think it changes the job description of a whole bunch of people listening.
For the last two years, if you were the person responsible for AI at your company, your job was basically enablement. Get people access. Knock down the barriers. Convince the skeptics. Roll out the tools. The hard part was adoption, and the metric was usage. More usage, good. People typing into the box, good. And the labs were rooting for you, because they were subsidizing the whole thing to win the land grab.
What changed is that usage stopped being free, and in the agent era, usage became effectively unlimited. Here's the mechanism, and it's worth understanding precisely. When AI was a chatbot, a human typed a question, read an answer, typed another. There's a natural governor on that — a person can only type so fast. But agents don't have that governor. You give an agent a goal and it churns. It calls the model over and over, it spawns sub-tasks, it retries when it fails, it reads huge piles of context. One human kicking off one agent can burn through tokens at a rate no human typist ever could. So the same move that makes agents powerful — that they run autonomously — is exactly what makes them financially dangerous. The thing that does the work is the thing that runs up the bill, and it does it without a human watching every step.
Now layer in what we just learned from Willison: the flat-rate plans are gone for the big customers. So you've got unlimited consumption meeting metered pricing. That's a bill that can go vertical overnight. And that is precisely why the conversation flipped to guardrails. Sam Altman, by the reporting that's been floating around the enterprise events, basically admitted that cost went from something that never came up early in the year to a huge issue for some companies. When the guy selling it says budgeting suddenly became a huge issue, you know the buffet's closing.
So what does the new job look like? The person responsible for AI at a company is no longer just an evangelist. They're now also a cost engineer. And the companies adapting well are doing a few specific things, and I want to lay them out because they're concrete and you can act on them.
First: routing. Stop using the most expensive model for every task. A one-line fix and a doc update do not need your most powerful, most expensive model. They need a cheap, fast one. The discipline is matching the level of intelligence to the actual difficulty of the job. The teams winning right now have a routing layer that sends each task to the cheapest model that can actually do it, and only escalates to the expensive model when the task genuinely demands it. The phrase I keep coming back to is: using the priciest model for everything isn't a quality strategy, it's a laziness tax. You pay extra for nothing.
Second: memory and context. A huge amount of token spend is just wasted re-explaining. Every time your system has to re-feed the same background, the same instructions, the same prior work into the model because it didn't remember anything, that's tokens spent before you even get to the actual task. Build systems that learn once and reuse, instead of paying the same exploratory cost over and over. This is why the memory upgrades the big labs are shipping aren't a consumer cuteness — they're a cost lever. A system with real persistent memory doesn't keep paying to relearn what it already knew.
Third, and this is the strategic one: measure cost per outcome, not cost per token. The per-token price was always a trap. The real number is tokens times price times the number of times you had to retry to get it right. A cheap model that rambles, that overthinks, that takes ten tries to converge, can cost you more in total than a pricier model that's terse and nails it on the first pass. So the cheapest model on the price sheet is routinely the most expensive per result. If you're a builder, the metric that matters is dollars per closed ticket, dollars per shipped feature, dollars per resolved support case. Price the way your customer thinks — per result, not per token. Because your customer doesn't care how many tokens you burned. They care whether the job got done and what it cost.
Now here's why I'm hammering this so hard. Bellan's piece, and the whole drumbeat around it, points at a structural truth: the second half of 2026 is going to be defined not by the flashy capabilities that got unveiled at the start of the year, but by the unglamorous work of putting those capabilities into production cost-effectively. The demo era is over. The deployment era is here. And deployment is an economics problem as much as it's a technology problem. If you're building in AI for the enterprise right now, whether you've named it this way or not, you are in the token efficiency business. Every layer of the stack is going to compete on it. The model makers compete on intelligence per dollar. The app makers compete one level up, on dollars per outcome. And the company that ignores this and just assumes compute stays cheap and subsidized forever is going to get a very rude invoice.
That's the deep dive. Now let me connect it forward, because the cost crunch isn't just about software billing — it runs straight into the physical world, into electricity and dirt and angry neighbors. Let's go there.
There's a story from Ashley Belanger at Ars Technica with a headline quote that I love for its honesty: we pissed off a lot of people. A giant data center plan got cut fifty percent amid protests. The developer felt, in their own words, beaten up, with no choice but to shrink the data center. Now I want to be careful here and stick to what the piece actually says, which is that a big data center project got cut in half because of community opposition. People in the area pushed back, hard, and the developer caved and halved the plan.
This is the other side of the compute crunch, and it's the side that doesn't show up in the token bill. All that demand we've been talking about — Google renting from SpaceX, the runaway costs, the whole thing — it all has to land somewhere physical. It lands as enormous buildings full of chips that drink electricity and water and sit next to where people live. And those people get a vote, not always a formal one, but a real one. They show up, they organize, they make the developer feel beaten up, and the plan shrinks by half.
For a builder, here's why this is more than a NIMBY footnote. The bullish case for AI assumes the infrastructure gets built on schedule. More data centers, more power, more interconnect, all coming online fast enough to feed the demand. But infrastructure doesn't get built in a spreadsheet. It gets built in a county, with a zoning board, with neighbors who don't want a humming warehouse and a substation in their backyard. And we're now seeing concrete evidence that the social license to build is itself a binding constraint. The bottleneck on AI might not be the model, and it might not even be the chips. It might be a planning meeting in a town you've never heard of where people decide they've had enough.
And that connects directly to power, which is the next thing on the menu. Because if data centers need electricity, the question of where that electricity comes from is suddenly red hot. Two stories speak to it.
First, from John Timmer at Ars Technica: the Trump administration is trying again to revive the dying coal industry. The plan would put money toward keeping existing coal plants open and building the first new coal plants in over a decade. Now I'm not here to do politics, I'm here to tell you what it means. Coal has been dying for years on pure economics — gas got cheaper, renewables got cheaper, coal couldn't compete. But here's the thing that's changed the conversation: AI demand is so hungry for power that suddenly anything that generates electrons looks strategic again. The fact that there's a serious push to keep old coal plants alive and build new ones tells you how desperate the demand for baseload power has become. When you're willing to revive a dying industry to keep the lights on for data centers, that's the power crunch talking.
And on the other end of the energy story, also from John Timmer, something genuinely new: a small modular nuclear reactor reached criticality in its first test. The reactor's from a startup called Antares, and to be clear, it isn't ready to generate power yet — criticality means the nuclear reaction became self-sustaining, which is a milestone, not a finished power plant. But it matters because small modular reactors are exactly the kind of thing people keep pointing to as the clean, dense power source that could feed the AI build-out without reviving coal. So you've got these two energy stories pulling in opposite directions on the same problem. The desperate, look-backward answer is keep the coal plants running. The hopeful, look-forward answer is little nuclear reactors you can build in a factory. Both are responses to the same pressure: AI needs power, lots of it, soon, and the grid we have wasn't built for it. Which path actually scales in time is one of the most important questions in tech right now, and it's being decided in places that have nothing to do with model architectures.
Let me shift the energy off heavy infrastructure and onto something a little lighter, but it actually rhymes with our whole theme today. There's a TechCrunch video piece arguing that the most interesting startups right now want to get you off your phone.
The framing is that while the AI fundraising machine keeps breaking its own records, some founders are deliberately building in the other direction. Mirror founder Brynn Putnam raised money for a startup called Board, focused on bringing people together through in-person games and social experiences. And there's a whole cyberdeck movement — people building whimsical do-it-yourself computers that, in the words of the piece, literally encourage users to touch grass. The argument is this isn't just reactionary backlash against AI. It's a real, fundable thesis that there's value in the analog, in presence, in being a body in a room with other bodies.
Why do I think this belongs on today's menu? Because it's the human-scale version of the same scarcity story. When something becomes infinitely cheap and infinitely abundant — and AI is making digital content and digital interaction approach exactly that — the thing that becomes scarce and valuable is its opposite. Attention you can't synthesize. A real game night with real people. Something handmade. As founders pour record money into making the digital world denser and more automated, a smart minority is betting that the premium product of the next decade is the un-automated experience. I'm not telling you to go build a board game company. I'm telling you that whenever a wave of abundance hits, the scarcity it creates on the other side is where some of the most interesting opportunities hide. Keep an eye on it.
Now let me do a quick run through the security desk, because there are two stories here that any founder running a real company should hear, and one of them is genuinely creepy.
From Lorenzo Franceschi-Bicchierai at TechCrunch: Google and the FBI are warning about a ransomware group that sends fake IT workers to hack victims in person. The gang's known as the Silent Ransom Group, and the technique is exactly as old-school and as effective as it sounds. They send people pretending to be IT support employees right into law firms' offices, where the criminals steal data using USB drives or remote access tools. Let that sink in. We spend all this time worrying about sophisticated remote cyberattacks, AI-powered exploits, the whole digital threat surface — and these guys just put on a polo shirt, walk in the front door, say they're from IT, and plug in a thumb drive. The most advanced security stack in the world doesn't help you if your receptionist waves the bad guy past the lobby because he's holding a clipboard and looking confident.
The lesson for builders, especially as you scale and start signing enterprise customers who care about this stuff: your security perimeter includes the physical lobby and the human being at the front desk. Social engineering at the door is having a moment, and it works precisely because everybody's looking at their firewalls instead of their front door.
And in the same neighborhood, from Dan Goodin at Ars Technica, a wild one: a highly-reviewed USB-connected speaker, the Sound Blaster Katana V2X, can be hacked over the air to infect connected devices. The piece walks through how a USB-connected speaker can infect a PC without anyone ever physically touching it. And the part that'll make you laugh and then wince — the seller doesn't consider the behavior a vulnerability. So you've got a device that can be compromised wirelessly and then used as a beachhead into the computer it's plugged into, and the company's official position is, that's not a bug. The takeaway is the boring-but-true one: every device you plug in is part of your attack surface, including the dumb speaker on your desk, and you can't count on the vendor to agree that their problem is a problem.
Quick whistleblower note in the same vein, also from Lorenzo Franceschi-Bicchierai: a former cybersecurity executive has filed a lawsuit accusing IBM of covering up several data breaches. The suit alleges IBM and two of its subsidiaries were breached back in the mid-2010s, and that IBM didn't disclose it and actively covered it up. Now this is an allegation in a lawsuit, not a proven fact, so I'm holding it at arm's length. But the reason it's worth a mention is the pattern it points at — disclosure is where a lot of security stories actually live or die. The breach is one thing. What you do, or don't do, about telling people, that's where the legal and reputational dynamite sits. If you're running a company, the cover-up is almost always worse than the crime.
Let me shift one more time, from security over to the product desk, because there's a small review that makes a big point about AI product design.
Ryan Whitwam at Ars Technica reviewed the Fitbit Air, and his verdict is a useful one for anybody bolting AI onto a product. The Air, he says, succeeds as a minimalist, reliable fitness tracker. Good hardware, does its job. But it's weighed down by a chatty AI coach — Google's AI Health Coach — that feels unnecessary. And his specific complaint is sharp: the AI is too nice to be your coach. It's so agreeable, so relentlessly encouraging, that it's useless at the actual job of a coach, which is sometimes to tell you the truth you don't want to hear.
I love this because it's a perfect little parable about the AI-everything reflex. There's enormous pressure right now to staple a chatbot onto every product. Got a thermostat? Add an AI. Got a watch? Add an AI coach. And what this review surfaces is that a bolted-on AorI that's tuned to be pleasant and safe and non-confrontational often makes the product worse, not better. A coach that only ever says you're doing great is not a coach, it's a cheerleader, and you didn't buy a cheerleader. The builder lesson: don't add AI to look modern. Add it where it does a job the user actually wants done, and be willing to let it have an edge if the job requires an edge. A health coach that can't ever push you is a feature in name and a nuisance in practice.
And there's a nice rhyme here with a writing piece floating around — an essay by Sam Kriss, the gist of which, from how it's being discussed, is a fierce stance against letting AI do your writing for you. I'll be honest, the discussion around it is mixed; some readers feel the examples meant to be obviously AI-generated actually read like decent writing, which is sort of the whole anxiety in a nutshell. But the thread connecting it to the Fitbit coach is the same: there's a flatness, a too-agreeable smoothness, that creeps in when you let the machine do the part that's supposed to carry a human point of view. Sometimes the value was the friction. Sometimes the value was somebody being willing to say the hard thing, or write the sharp sentence. Strip that out for convenience, and you've got a product that's technically functional and somehow hollow.
Let me close out the menu with a couple of quick founder-flavored items before we wrap.
There's a piece from Julie Bort about founders sharing VC horror stories on X this week, and some are naming names. A massive viral conversation, she writes — some of the stories are weird, some are infuriating. I'm not going to repeat specific accusations because they're anecdotes flying around social media and I'm not in the business of laundering unverified claims about named people. But the phenomenon itself is worth noting. When founders feel safe enough, in a down-ish funding environment, to start publicly airing how they got treated by their investors, that's a power shift. For years the founder-VC relationship had a strong code of silence, because you might need that person's check again, or their network. The fact that the silence is cracking tells you something about where leverage sits right now and how a generation of burned founders is feeling. If you're raising, the cultural takeaway is simple: reputations are becoming more transparent in both directions. Behave accordingly, on both sides of the table.
And one more from the bootstrap-versus-VC file, because it's a clean little counter-narrative. Kirsten Korosec reports that as venture-backed e-bike startups went bankrupt, a bootstrapped company called Lectric grew. Lectric says the U.S. market is ripe for competition and choice, and it's launched three new brands in the past six months. I love this one precisely because it cuts against the religion of this whole town. The VC-funded e-bike darlings raised big, scaled fast, and went bust. The company that took no rocket fuel, that grew on its own cash flow, that stayed disciplined — that's the one still standing and expanding. It's not an argument that venture is bad. It's a reminder that capital is a tool, not a strategy, and that in a real physical business with real margins, the discipline that bootstrapping forces on you can be the thing that keeps you alive when the funded competitors flame out. Worth keeping in your back pocket the next time somebody tells you the only way to win is to raise the biggest round.
And just to bookend the AI-builder world, a quick note off a swyx post: the AI Engineer World's Fair is about three weeks out, the big sponsor tiers are sold out, and if you were going to go, the room block discount was on a tight clock. I mention it only because the fact that an AI engineering conference sells out its presenting, model-lab, platinum, and gold sponsorships tells you the money and the attention in this space are still very much pointed at the people actually building, not just the labs. The center of gravity is the engineer in the trenches. That's the audience this show is for, so it's nice to see it have its own marquee event.
So let me tie the bow on today. The through-line, if you squint, is the bill coming due — in every form. The token bill, where flat-rate buffets are closing and metered pricing is forcing everybody to get efficient or get burned. The compute bill, where Google's paying SpaceX nearly a billion a month because demand outran even Google's own buildout. The Wall Street bill, where the S&P 500 says show me the profits before you get the passive-money firehose. The electricity bill, where AI demand is reviving coal and pushing tiny nuclear reactors to criticality and getting data centers cut in half by angry neighbors. And the human bill, where a fistful of founders are betting the scarce, valuable thing in an automated world is getting people off their phones and into a room together. Two years of go-fast, and now the invoices are arriving. The builders who win the next phase are the ones who read the invoice carefully, route the cheap tasks to cheap models, measure cost per outcome, and never assume the subsidy lasts forever.
That's the show. I'm Tony DeLuca, this has been Barely Possible, and I appreciate you spending a little of your day with me. Be good to your neighbors — turns out they get a vote on your data center. Catch you next time.