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, the show where we sit down at the kitchen table, push the hype to the side, and figure out what actually moves the needle for people who build things. Big menu today, so grab your coffee, get comfortable, let's have at it.
We're going to start with a number that almost nobody is talking about correctly, because it sits underneath everything else in AI right now. A US memory chip company just reported that its revenue quadrupled to a little over forty-one billion dollars compared to the same period a year ago. And the profit — this is the part that'll make you spit out your coffee — went from one point eight-eight billion to twenty-eight point two billion year-over-year. That's not a typo. Profit went up roughly fifteen times. Kirsten Korosec reported this one, and the framing in the piece is that the memory chip crunch is paying off for this company. I want to spend a minute on it, because it tells you something real about where the money in this whole AI boom is actually piling up.
Here's the thing. Everybody talks about Nvidia. Everybody talks about the model labs, OpenAI, Anthropic, Google. But the AI data center buildout doesn't just need GPUs. It needs an enormous amount of high-bandwidth memory sitting right next to those GPUs, because the models are hungry, and feeding them is a memory problem as much as it's a compute problem. When demand for that memory spikes and supply can't catch up, prices go vertical, and the handful of companies that make the stuff print money. Twenty-eight billion in profit is what a shortage looks like when you're on the right side of it.
Now why should a founder care? Two reasons. One, if you're building anything compute-heavy, your input costs are downstream of this. When memory is scarce and expensive, the people renting you GPU capacity are paying more, and eventually that finds its way into your bill. Two, it's a tell about the cycle. When the picks-and-shovels suppliers are posting numbers like this, it means the buildout is still in full sprint, not slowing down. Whether that's a healthy sprint or a bubble inflating — well, we've got more on that later in the show, because there's a leaked-financials story floating around that points the other direction.
But hold that thought, because the memory story connects directly to the next one. If the supply side is minting cash, the demand side — the companies actually consuming all those tokens — is starting to feel the pinch in a way that's almost funny.
Lucas Ropek had a piece with maybe my favorite line of the week in it: "The tokenmaxxing era was brief. We now appear to be entering the era of token rationing." Companies are scrambling to stop their own employees from maxing out the AI budget on small tasks. Picture it. You give your team access to the good models, the expensive ones, and somebody's burning a fortune in tokens to reformat a spreadsheet or rewrite a two-line email. The bill comes in, and finance has a heart attack.
This is one of those stories that sounds like a joke but is actually the whole ballgame for AI inside companies. For about a year, the message was: use it for everything, the more the merrier, get your people comfortable. That was the tokenmaxxing phase. And now the same companies are putting up guardrails, rationing access, routing cheap tasks to cheap models and saving the premium reasoning for premium problems. That swing — from "use all you want" to "hey, watch the meter" — is the enterprise AI adoption curve growing up in real time.
For the builder, here's the takeaway. If you're selling AI features to businesses, the buyer's question has shifted. It's no longer just "can it do the thing." It's "can it do the thing without blowing up my unit economics." The teams that win the next eighteen months are going to be the ones who are smart about routing — small model for small jobs, big model for big jobs — and who can show a customer exactly where the money's going. The pricing conversation and the capability conversation are now the same conversation. If you can't talk cost-per-task, you're not really in the room.
Now let's shift from who's paying the bills to who's leaving the building. There's a talent story that's getting heavy, and I want to handle it carefully because it's adjacent to something we covered earlier this week.
Amanda Silberling and Lucas Ropek reported that two more top AI researchers, Jonas Adler and Alexander Pritzel, are leaving Google for Anthropic. And that's on top of the departures we already knew about — Noam Shazeer heading to OpenAI, and John Jumper, the Nobel laureate, going from DeepMind to Anthropic, which we talked about last week as a talent-flow signal. So this is a genuine continuation, not a rerun. The new names are new. The pattern is the story.
And I want to be honest with you about how to read this, because there are two ways to overreact. One side says: Google is finished, the researchers are voting with their feet, the model race is lost. The other side says: it's just two people, happens all the time, ignore the noise. And the truth is somewhere in the uncomfortable middle. Individual departures don't tell you much. But a sustained outflow of senior, named talent toward two specific rivals — that's not nothing, especially when the market is watching. We saw earlier this week that Google's stock got hit hard on the earlier departures, shedding a chunk of market cap on the read that they're losing the war for talent at the frontier.
Here's my read as a guy who's skeptical of headlines. The danger for Google isn't any one researcher. It's the perception loop. Once the story becomes "the best people are leaving," it gets harder to recruit the next best people, which makes the story more true, which makes recruiting harder still. That's the trap. And Google has the deepest pockets, the most data, and a custom chip stack nobody else has, so I would not bet the farm against them. But if you're a founder watching where the brains are flowing, the brains are flowing toward Anthropic and OpenAI right now, and that matters for where the next capabilities show up first. Keep your eye on it. Don't write the obituary.
And speaking of jobs and who's safe and who isn't, there's a piece of data that cuts hard against the doomer narrative, and I love a good counterpoint, so let's get into it.
Marina Temkin wrote it up: AI was supposed to kill engineering jobs, but new data suggests they're actually the most resilient. According to SignalFire data, while AI dominates the layoff headlines, engineers are making up a larger share of new hires, not a smaller one. Let that sit for a second, because the conventional wisdom has been the exact opposite. The story everyone tells is: AI writes code now, so we need fewer engineers. The data is saying: companies are hiring more of them, proportionally.
Why would that be? My take is it's the oldest pattern in technology. When you make a thing cheaper and more powerful, you usually don't use less of it — you use a lot more of it, on more problems. AI makes each engineer more productive, which makes engineering more valuable, which makes companies want more leverage from more engineers, not fewer. The layoffs you read about are often in other functions, or they're companies that overhired in the boom years cleaning up. But the people who can actually build and ship and wrangle these systems? They're getting more in demand, not less.
Now I'd put a small asterisk on this. "Engineers as a share of new hires" going up doesn't automatically mean total engineering headcount is exploding everywhere — it can also mean other roles are getting cut faster. But directionally, for the builder listening to this who's been told a robot is coming for their git commits, the early data says: relax a little. The skill that's getting squeezed isn't building software. It's the routine knowledge work around it. If you can architect systems, integrate the messy parts, and take responsibility for what ships, you are on the resilient side of this.
Okay. Let's pivot to the hardware floor, because there's a robotics story that deserves real attention from anyone thinking about where the physical-world money is going.
Agility Robotics, the humanoid robotics startup that spun out of Oregon State University back in 2015, is planning to go public via SPAC in a two-and-a-half-billion-dollar deal. Kirsten Korosec reported it. They expect to generate around six hundred and twenty million dollars in proceeds. Now, you hear SPAC and a lot of you flinch, and fair enough — the SPAC boom of a few years back produced a graveyard of companies that went public before they were ready and got punished for it. So let's be clear-eyed.
Humanoid robotics is one of the hottest narratives in tech right now. The pitch is enormous: general-purpose robots that can work in warehouses, factories, anywhere humans do repetitive physical labor. Agility is one of the more serious names in that space; they've actually had robots doing pick-and-pack work in real facilities, not just dancing in demo videos. So the company has substance. The question is whether the public markets are the right place for a hardware company that's still scaling, still burning cash, and still proving the unit economics of a robot that costs real money to build.
For the founder, here's why I flag it. Two reasons. First, a SPAC at this valuation is a temperature check on how much investor appetite there is for physical AI — robots, not chatbots. The money is clearly hunting for the next leg of the story beyond software. Second, watch what happens after they're public. If a credible humanoid company gets out there and the numbers hold up, it opens the door for the whole category. If it gets out there and the stock craters because the economics aren't there yet, it slams that door shut for everybody behind them. This one's a bellwether. And remember, earlier this week we saw GM installing robots at its flagship EV factory after layoffs — the demand for this stuff in industrial settings is real and rising. The open question is always whether the supply side can hit a price that makes the math work.
Now let's talk about a company doing the opposite of overpromising. Let's talk about a cheap truck.
Ars Technica got a hands-on ride in the Slate electric pickup — the twenty-four thousand nine hundred and fifty dollar one. Roberto Baldwin wrote it up, and the headline framing is "underpromise, overdeliver," which, in an industry that does the reverse constantly, is refreshing to hear. Two hundred and five miles of bare-bones range. Bare-bones is the whole pitch. This is a deliberately stripped-down, affordable EV in a market where the average new car price has gotten genuinely insane.
And there's a companion piece from Tim De Chant about why Slate changed the battery chemistry in that truck — the short version is the decision had been building for years, it wasn't some panicked last-minute swap. I'm not going to drag you through battery chemistry on a builder's podcast. The reason I mention Slate at all is the strategy, because it's a lesson that travels way beyond cars.
In a hype-soaked market, there is enormous room for the company that says, "We're going to do less, charge less, and actually deliver the thing." Everybody else is promising a thousand-mile range and full self-driving and a spaceship. Slate is promising a basic, honest truck for under twenty-five grand. That's a positioning play, and it's one I'd tell software founders to study. When your whole category is over-promising — and AI is over-promising right now at a level that would make a used car salesman blush — there is a real, durable customer for the team that under-promises and ships something that works. Trust is a moat. Slate is building one out of restraint.
Let's stay on cars for one more beat, because there's a regulatory story with teeth. The NTSB — the National Transportation Safety Board — has launched a probe into a fatal Tesla crash in Texas. Sean O'Kane reported it. The safety board, which is known for genuinely thorough, slow, careful investigations, is digging in alongside the National Highway Traffic Safety Administration. We touched on the underlying crash earlier this week when NHTSA opened its look at it. The new piece here is the NTSB joining, and that's meaningful, because the NTSB doesn't move fast but it moves deep. When they get involved, the resulting report tends to actually shape policy and design.
For anybody building in autonomy or selling "hands-off" anything, this is the world you live in now. The technology runs ahead, and then the investigators show up after a tragedy and start asking exactly what the system was doing, what the driver was told it could do, and where the gap was. If you're building agents that take real-world actions — and a car is the most literal version of that — the lesson is the same as it always is: the moment your software does something irreversible in the physical world, you inherit a level of scrutiny that pure-software founders never face. Build for the investigation you hope never happens.
Alright, let me bring you the deep dive, because this is the one that I think matters most for builders, and it's got some real reporting behind it.
Let's talk about the era of token rationing for a second more, but through a wider lens, because there's a leaked-financials story circling that gives the whole "is this a bubble" question some actual numbers to chew on. Ed Zitron, who runs the newsletter Where's Your Ed At, published what's being passed around as an exclusive on OpenAI's financials. Now I want to be careful here. I have not independently verified these numbers, the framing comes from a writer who is openly, famously bearish on AI economics, and the item I'm working from is a pointer to his piece, not the full audited books. So I'm not going to read you specific figures as gospel. What I can tell you is the reaction it's generating, and the reaction is the story.
The person who flagged it summed up the mood in one line — the numbers, they said, read almost as badly as a certain space company's IPO filing. And whether or not you trust Zitron's exact accounting, here's what's actually going on, and why it connects to everything else in today's show. We have, on one hand, a memory chip company posting twenty-eight billion in profit because demand for AI infrastructure is white-hot. We have SPACs valuing humanoid robot companies at two and a half billion. We have researchers commanding the kind of pay that moves a trillion-dollar company's stock when they leave. The supply side and the talent side are screaming "boom."
And on the other hand, we have the demand side — the companies actually buying and using all this — rationing tokens because the bills are scary, and we have a steady drumbeat of reporting questioning whether the biggest model labs actually make money at the scale they're operating. That's the tension. That's the whole thing. Massive, confident capital expenditure on infrastructure, meeting genuinely uncertain unit economics at the application layer.
Now here's why I'm framing the deep dive this way instead of just dunking on OpenAI or just cheerleading the boom. Because the interesting question for a founder isn't "is it a bubble, yes or no." Bubbles and real revolutions look identical from the inside — the internet was both at the same time. The interesting question is: if the cheap capital dries up, who survives? And the answer, every single time in history, is the companies whose customers can't live without them and whose costs are under control. That's it. That's the whole survival kit.
So connect the dots. The token rationing story isn't a cute aside — it's the early warning system. When enterprises start metering AI usage, they're telling you they're not yet convinced the value exceeds the cost on every task. They believe in some of it. They're skeptical of the rest. And the labs are spending like every task is going to convert. If Zitron's bearish read is even directionally right — and again, I'm holding that loosely — then the gap between infrastructure spending and proven, profitable demand is the thing that pops if anything pops.
What do you do with that as a builder? Three things. First, get religious about your own unit economics now, while capital is still cheap, so you're not scrambling if it stops being cheap. Second, build the kind of product where the customer rations everything except you — be the load-bearing tool, not the nice-to-have. Third, watch the suppliers. When the memory and GPU guys start reporting softening demand instead of fifteen-x profit jumps, that's your signal the music's changing tempo. Until then, the buildout is real, the money is real, and the skepticism is also real, all at once. Hold both. That's the honest position.
Okay, let's come up for air and hit some shorter ones, because there's a bunch worth knowing.
Google has started lowering Play Store fees, making good on the Epic Games settlement. Ryan Whitwam reported it. A few additional markets get the lower fees this year, ahead of a global rollout in 2027. This is the long tail of the Epic Games versus Google fight finally landing in developers' actual pockets. If you ship a mobile app, this is money — the cut Google takes coming down is real margin back to you. It's slow, it's phased, it doesn't hit everywhere at once, but the direction is set. The era of the unquestioned thirty-percent app-store tax is grinding toward its end, and that's good news for anyone building on these platforms.
Next up, Mistral released Mistral OCR 4. Optical character recognition — turning documents and images into machine-readable text. Now, I'd normally breeze past a model release, but OCR is quietly one of the most useful unglamorous building blocks out there. Every company drowning in PDFs, scanned contracts, invoices, forms — that's an OCR problem before it's an AI problem. A strong, current OCR model from Mistral is the kind of thing that just makes a hundred boring-but-valuable workflows possible. Worth a look if you're building anything that has to read the real world's messy paperwork. Link in the show notes.
On the security front, there's a one-two punch in a global operation that disrupted what investigators called a cybercrime "assembly line." Dan Goodin reported it — "Operation Endgame" simultaneously disrupted two widely used crime tools. This is law enforcement getting smarter about hitting the infrastructure of cybercrime rather than just chasing individual bad guys. For founders, the reminder is that the threat landscape your product lives in is an arms race with actual cops on the field now. Good to see the defenders landing punches.
Related, and useful: a new website is naming and shaming companies that still don't offer passkeys. Lorenzo Franceschi-Bicchierai wrote it up. According to the new site, twenty-four percent of the most popular websites in the world don't support passkeys, which are considered the most secure way to log in. If you're building a product and you're still doing passwords-only, you're now publicly on a list of laggards. Passkeys aren't exotic anymore. They're the baseline. Get them in.
There's also an FCC story worth a quick mention: the FCC is planning an ID mandate that could block anonymous use of prepaid burner phones. Jon Brodkin reported it. Privacy advocates and domestic violence groups are saying it's a big mistake, and their argument is concrete — anonymous prepaid phones are a lifeline for people fleeing dangerous situations. This is one of those policy moves where the stated goal, cutting down on crime, runs straight into a real human cost. Worth watching how it shakes out, because the trend toward identity-verification-for-everything keeps marching, and it has knock-on effects for anyone building communications or privacy products.
And in the "AI meets the creative world" bucket, a couple of quick ones. Deezer, the music streaming service, rolled out a feature that lets fans remix songs with artist consent. Lauren Forristal reported it, and the framing is that Deezer's taking a contrarian approach to AI — consent-first, artist-in-the-loop. That's the interesting wrinkle. Most of the AI-and-music fights have been about scraping and using artists' work without permission. Deezer's trying to build the opposite: a sanctioned remix lane where the artist opts in. Whether fans actually want it is a different question, but the consent-first design pattern is the one to steal. Meanwhile, Facebook is rolling out an AI companion app for creators, currently in testing with select creators, built around Facebook's AI creator assistant. Aisha Malik reported that one. Meta keeps pushing AI deeper into the creator workflow — that's a durable strategy from them, wrapping tools around the people who make the content that fills their feeds.
Let me close the loop with a couple of space and science notes, because they're genuinely interesting even if they're not strictly builder fuel.
Eric Berger had a report justifying NASA's recent program cancellations — an analysis found the exploration programs that got canceled were running way, way late and way over budget. The line that jumps out: contract values for these efforts ballooned from nearly two point eight billion dollars to five point nine billion. Thirteen years and five hundred million dollars for a stage adapter, per the headline. I bring this up because the contrast with the private sector is stark. Earlier in the week we saw reporting that Kennedy Space Center isn't even ready for the era of super-heavy rockets, with SpaceX talking about launching Starship every eight days. You've got legacy government programs taking thirteen years for a part, and private operators talking about a launch cadence measured in days. That gap — institutional pace versus startup pace — is the same gap we keep seeing everywhere in tech. It's why founders can still beat giants. Speed is a weapon.
And one pure delight to send you out on: an experimental wine bottle that tracks oxygen moving through the cork. John Timmer wrote it up. Researchers built a bottle that lets them watch oxygen and other chemicals move in and out through that little pocket of air. Why do I love this? Because it's a reminder that careful measurement of a thing everybody assumed they understood — how wine ages — can still surprise you. That's the scientific temperament, and honestly, it's the builder temperament too. Don't assume you know what's happening inside the black box. Instrument it. Measure it. Watch the oxygen move.
So let me tie the bow on today. The supply side of AI is printing money — twenty-eight billion in memory-chip profit, multi-billion-dollar robot SPACs, researchers worth a stock-moving fortune. The demand side is rationing tokens and quietly asking whether the value covers the cost. Both of those things are true at the same time, and the leaked-financials chatter around OpenAI is just the loudest version of that same tension. Your job, if you're building, is to live in that contradiction without picking a comfortable side. Be the load-bearing tool. Know your unit economics cold. Watch the suppliers for the first crack. And under-promise, like that cheap honest little truck, and then deliver.
That's the menu for today. I'm Tony DeLuca, this has been Barely Possible, and I appreciate you spending a little of your time at the table with me. Be good to each other out there, and I'll catch you on the next one.