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, pull up a chair, grab whatever's in the fridge, I'm your boy Tony DeLuca and this is Barely Possible, where we take the day's pile of tech noise, shake out the confetti, and see what's actually worth your morning. We've got a menu today that runs from a knife fight over smartphone memory chips to a couple of billionaires trash-talking each other about data centers in outer space, plus a guy who thinks your AI should help you plan a murder, and I mean that literally. So buckle up, there's a lot to chew on.
Let me start with the one that hits your wallet and your product roadmap at the same time, because it's less glamorous than the celebrity feuds but a lot more consequential if you build anything that ships in hardware. There's a report out from Ryan Whitwam that smartphone shipments have hit historic lows, and the driver is a memory shortage. Now the headline reads like a bad-news story, phones are down, everybody's hurting. But read the actual angle and it flips on you. The biggest smartphone makers, Apple and Samsung, they're the ones benefiting. When components get scarce and expensive, the giants keep on trucking, as the piece puts it, because they've got the buying power, the supply contracts, the balance sheet to eat the cost or pass it along. It's the little guys that get squeezed out of the memory market.
And I want you to sit with that for a second, because it's a pattern we've been circling on this show for weeks now without naming it directly. A couple weeks back we talked about South Korea's trillion-dollar bet on memory fabs. We've talked about SK Hynix raising twenty-six and a half billion in that giant US IPO. All of that is the same story from a different window. Memory is the chokepoint right now. The AI buildout is vacuuming up high-bandwidth memory, the same fabs that make the chips for your phone are being told to prioritize the data-center stuff, and so downstream, the consumer device world is getting starved. And when supply gets starved, scale wins. Whoever's biggest, whoever has the longest-term contracts, whoever can lock up capacity — they win, and everybody else fights over scraps.
Why should you care if you're a founder? Because if any part of your product touches physical hardware — a device, a wearable, an edge box, anything with RAM in it — your bill of materials is not stable right now, and it's not going to be stable through 2027. The chairman of one of these memory makers, in comments we referenced last week, said flat out he expects 2027 to be the worst year for memory supply. So if you're pricing a hardware product on today's component costs and projecting them flat, you're building on sand. Plan for the squeeze. And if you're a software shop feeling smug about not touching hardware — remember your cloud inference costs ride on this same memory market too. The scarcity doesn't skip you, it just reaches you through a longer pipe.
Now let me pull us over to a fight that's a lot more fun to narrate, even if it's dumber. Sam Altman and Elon Musk are back at each other's throats, and this round is about data centers in space. Here's the setup. Musk accused Altman of being a scammer, and Altman fired back — and I'll read it clean — quote, "homeboy you're the one selling public market investors on short-term space datacenters." End quote. Now Tim Fernholz makes the point in the piece that I think is the actual news buried under the schoolyard stuff: Altman's dig, the trash talk, is basically what most experts already believe. The idea of putting data centers in orbit in the near term — cooling them, powering them, servicing them, getting the compute up and down — most serious people think that's a fantasy on any short timeline. So when Altman calls it a short-term space data center pitch to public market investors, he's not just being petty, he's saying the quiet part: somebody's selling a story to the markets that the engineering doesn't back up yet.
And look, I don't want to spend your morning refereeing two of the richest men on Earth clowning on each other. There was more of this over the weekend, the parole officer cracks, the whole bit. But here's the founder takeaway, and it's a real one. When the biggest names in the field are publicly accusing each other of selling vaporware to the public markets, that tells you the capital-raising environment has gotten frothy enough that even the insiders are calling foul. Space data centers, whatever fusion timeline of the week — there's a lot of story being sold right now on the strength of AI demand. When you hear a pitch that leans on a decade-out miracle to justify today's valuation, be the skeptic in the room. The people at the top of this thing are, so you can be too.
And speaking of stories sold to public markets, let me hand you a clean case study. There's a report from Tim De Chant that General Fusion started trading on the Nasdaq — now this is framed as an older story, the merger was announced back in January, so I'm not going to pretend this all happened this morning, but it's worth flagging as background because it's such a perfect illustration. They became the first publicly listed fusion power company, and on debut the stock rallied, up around forty percent from twelve dollars and change. Investors wanted in. Sounds great, right? Here's the part that matters. This was a de-SPAC deal, and like most de-SPAC deals, it got hammered by redemptions. Without redemptions, they could've added as much as two hundred thirty million to the balance sheet. After redemptions and fees, one estimate puts it at less than thirty million. Thirty. They cobbled together about a hundred fifty million in cash total with a private raise on top.
Now here's the human reality behind the shiny debut. This is a company founded in 2002 — twenty-four years old — that has raised over six hundred million dollars from private investors over the years, laid off at least a quarter of its staff last year, had to run a pay-to-play round to keep the lights on, and their breakeven milestone, the actual proof that the physics works, has slipped to 2028 or later. First power plant, quote, "by approximately 2035." I say that not to dunk on the science — fusion is genuinely hard and genuinely important — but to point at the gap between a stock popping forty percent on day one and a business that's spent two decades and six hundred million chasing a milestone it still hasn't hit. The market enthusiasm and the operational truth are living in two different zip codes. That's the same thing Altman was needling Musk about with the space thing. Keep your eye on the daylight between the story and the ledger.
Let me shift us from the money to the machines, because there's a story here that I think is the most genuinely important thing in today's pile for anybody thinking about where product value actually lives. It comes wrapped in the dumbest possible headline, so bear with me. TechCrunch ran a piece by Russell Brandom with the title, and I'm not making this up, "Should AI help you get away with killing your spouse?"
Now before you spit out your coffee — this is a real debate about AI alignment, and the murder framing is the writer poking at a genuinely serious argument. Here's what happened. There's an alignment policy paper going around called AI 2040: Plan A, from the AI Futures Project, and its big idea is a world where researchers collectively agree to slow down AI development for fourteen years for the good of humanity. And George Hotz — the Comma AI founder, longtime jailbreaker — he read that and basically said, absolutely not. Hotz's whole position is that AI progress shouldn't be managed for some collective good. His preferred approach to safety is locally controlled models that are tightly aligned to the interests of their individual user. Not a model in a data center that answers to a lab's rulebook — a model that's yours, that does what you want.
And here's where the murder thing comes in. Hotz compares a user-aligned AI to a gun. A gun, he says, doesn't complain if you use it to kill your stepmom. A truly aligned AI, in his framing, would order the meth-lab equipment off Amazon Prime and show you how to use it if that's what you asked for. He says, quote, "We either live in a world with freedom or we don't." End quote. And he says he'd die to defend that principle.
Now Brandom pushes back, and I think correctly, that this is a little adolescent. His line is that any structure with a lot of people in it — societies, marketplaces, companies — requires binding individual needs into a network of accountability, and that anybody shipping mass-market products should take seriously the interests of, in his words, the as-yet-unmurdered spouses and stepparents of the world. Which, funny, but also true. The freedom Hotz enjoys is a space of possibilities made possible by collective enterprise. Cute philosophy fight.
But here's why I'm giving this real airtime instead of laughing it off, and here's the part that matters for you. Strip out the murder theatrics and Hotz is pointing at a real architectural fork in the road, one that has been under all our stories this week without anybody quite naming it. Right now almost all the AI you use is centrally managed — Claude, ChatGPT, the big hosted models. Brandom himself notes there are infrastructure reasons for that: it's expensive to host frontier models, and most people don't use them heavily enough all day to justify a truly personal, local AI. But — and this is the key — those factors get weaker as the technology matures. The economics that force central hosting today may not hold tomorrow. And the question of whether the model that runs your life answers to you or answers to a lab in San Francisco — that's not a philosophy-seminar question, that's a product-strategy question. That's a moat question. That's a data-ownership question.
And here's the beautiful thing about today's stack of stories — that exact fork shows up again, from a very different mouth, wearing a suit. Because Satya Nadella went out this week with what one write-up called a shocking warning to companies using AI. Now I want to be careful, the fresh item I have here is short on the full text, so I'm going to give you the shape of it rather than pretend I've got every word. The framing is that the worry rippling through Silicon Valley right now is that the giant labs selling proprietary models are, in effect, acting like Trojan horses. The argument — and this echoes things Nadella and others have been saying in longer form — is that when you pour your company's proprietary knowledge into somebody else's hosted model to make it useful, you're paying twice. Once with money, and again with the knowledge you have to hand over. And every correction you make, every eval you run, every time you fix the model's mistake — that's institutional know-how, and it can leak, trace by trace, into the thing that belongs to the lab, not to you.
Now connect the dots. Hotz, the jailbreaker provocateur, says own your model or you don't have freedom. Nadella, running one of the biggest software companies on the planet, says own your learning loop or the economic value flows to the guy who owns the infrastructure, not to you. Different vocabulary, same fork. And that's the through-line I want you carrying out of today's episode: the fight over whether AI is a service you rent or a capability you own is quietly becoming the fight of this year. If you're building a company on top of a hosted frontier model, ask yourself the uncomfortable question — what are you teaching it, and who ends up owning what it learned? That's not paranoia. That's due diligence.
Let me stay in the AI world for a minute but bring it down to earth, because while the philosophers argue about ownership, the labs are just quietly moving product. A recent report — this is from about five, six days back, so I'm not going to call it hot off the press — OpenAI put out a batch of stuff. There's GPT-5.6, pitched as more intelligence per token and stronger performance per dollar. And there's ChatGPT Work, which they describe as an agent that can take action across your apps and files and, quote, "stay with a project for hours if needed," turning a goal into finished work. That "stay with a project for hours" line is the tell. That's the whole industry pivot in one phrase — from a chatbot that answers a question to an agent that grinds on a task while you go do something else.
And here's a nice piece of self-awareness buried in the same batch, and this one's genuinely useful for builders. OpenAI put out an analysis, in an essay called Separating signal from noise in coding evaluations, poking holes in SWE-Bench Pro — one of the popular coding benchmarks everybody cites — raising concerns about its reliability and accuracy for evaluating models. Now think about who's saying that and why. A frontier lab whose models get measured on these benchmarks is telling you the benchmark is noisy. That could be defensive, sure. But it's also just true, and it's the right instinct for you as a builder: do not buy a model because it topped a leaderboard. The leaderboards are leaky, the benchmarks get gamed and contaminated, and the number that matters is how the thing performs on your workload, on your evals, in your product. Build your own test set. That's the ownership theme again, in a different suit — own your evals or you're trusting somebody else's scoreboard.
There's an international wrinkle worth thirty seconds too. Anthropic started localizing Claude pricing for India, rolling out rupee-denominated subscription plans. India is described as their biggest market after the US. That's not a splashy story but it's a meaningful one — it says the frontier labs are done treating the world as a monolith priced in dollars and are starting to fight for developing-market users on local terms. If you're building for a global audience, that's a signal about where the next wave of AI users is coming from and how much they'll pay. Watch the pricing, not the press release.
Now let me pivot hard, because not everything today is about who owns the robot brain. Some of it is about people getting their networks quietly hijacked, and this one's a genuine to-do item for you and your ops team. There's a fresh advisory — Dan Goodin wrote it up — where the US government, through CISA, is warning that Russian state hackers are going after home and small-office routers. Specifically it's the FSB, Center 16, and the crews get tracked under a menagerie of names — Berserk Bear, Energetic Bear, Dragonfly, and my personal favorite for a threat group, Static Tundra. The advisory got co-signed by allied governments — Australia, Denmark, New Zealand, the UK.
Here's the mechanism, and it's less exotic than the spy-novel names suggest. They scan the internet for routers running old versions of SNMP — that's the network management protocol — that are sitting there with default or common credentials. Versions one and two of SNMP don't even encrypt the password. They get in, they turn your router into an exit node, and then they route their attacks against real targets — energy, defense, finance, government — through your benign-looking, trustworthy IP address so the firewalls on the other end don't flinch. Your router becomes the getaway car for somebody else's crime. And the piece is honest that the government's countermeasures so far have been whack-a-mole — they knock down a botnet, the operators just stand up a new one.
The fix is boring, which is exactly why nobody does it. Disable SNMP versions one and two, use version three or turn SNMP off entirely if you don't need it, kill Cisco Smart Install, use real passwords, update your firmware. If you run a startup out of a home office or a small space with a consumer-grade router humming in the corner, this is your reminder that the humming box is an attack surface. Go poke at it this week. It's the least sexy security homework you'll get all year and it's probably the highest-return.
Now I told you we'd get into the geopolitics, and there are two threads here worth pulling, briefly, because they both bear on the same thing: how much the physical world still runs your digital ambitions. First — and I want to be straight with you about the framing — there's a piece by Jeremy Hsu about Ukrainian drone strikes shutting down Russian shipping in the Sea of Azov. The reporting describes strikes on more than a hundred Russian tankers and ships over about a week, forcing Russia to halt shipping through a vital corridor and further isolating Crimea. Now this item carries an older event reference from December of last year folded into it, so I'm not going to present the whole thing as breaking-this-morning news — treat it as a recent report on an ongoing campaign rather than a bulletin.
But the durable insight is worth your attention, and it's this: a country with essentially no traditional navy managed to blockade a maritime corridor using cheap one-way attack drones. As one expert quoted in the piece put it, you're not sinking the ship, but it's a mission kill. The piece notes a similar dynamic in the Strait of Hormuz. Why does a founder care about drone warfare? Because the entire logic of cheap, distributed, autonomous systems overwhelming expensive centralized assets is the same logic reshaping software. Cheap and many beats expensive and few, over and over, in domain after domain. And because when shipping corridors get disrupted, grain prices move, oil prices move, and — tying to a small item Yann LeCun reshared — oil was up around four and a half dollars after a diplomatic breakdown with Iran. The physical supply chain and the energy market are the ground your data centers stand on. You can't abstract your way out of them.
Which brings me to a quieter policy story that's going to matter more than it looks. There's a report that Colorado is going to put a question to voters about whether to add a "right to natural gas" to the state constitution — an amendment that would restrict building codes that promote electrification. Small item, big signal. As AI data centers eat more and more power, the fights over how we generate and deliver energy are going to get sharper and more local and more constitutional. Meanwhile California just created a thirty-five-hundred-dollar rebate for new EV buyers and a smaller one for used, both with price caps. You've got states pulling in opposite directions on the energy transition at the exact moment the compute industry is becoming one of the biggest new electricity customers on the grid. If your business plan quietly assumes cheap, abundant, clean power in five years, understand that assumption is now a political football being kicked around fifty different statehouses. That's not settled. Watch it.
Let me give you two quick business-world hits before I bring it home. The states-versus-Paramount story resurfaced — twelve state attorneys general, led by California's Rob Bonta, suing to block the hundred-ten-billion-dollar Paramount-Warner Bros. Discovery merger. Now the underlying filing traces back to earlier in the year, so this is renewed attention on an existing fight, not a brand-new lawsuit dropping today. The claim is the combined company would control something like twenty-seven percent of US film distribution, thirty percent of blockbuster distribution, and a big chunk of basic cable. Bonta's line, quote, "America has no kings in government or our economy." The wrinkle worth noting for anybody watching regulatory risk: the Trump administration's DOJ already cleared this deal, saying it wasn't likely to harm competition, and now a coalition of states is trying to block what the feds waved through. That state-versus-federal split on antitrust is the thing to file away. If you're building in a consolidating industry, the approval you get from Washington may not be the last word anymore.
And a small founder-flavored one I liked: the TV-tracking app TV Time is shutting down, and its founder is turning right around and building a successor called Bingers that'll let users import their watch histories and keep the community intact. I flag it because it's a nice reminder of a real pattern — when a product dies, the community that formed around it doesn't necessarily die with it, and there's a founder move in catching that community mid-fall and giving it a new home. Data portability as a feature, community as the moat. Not every story has to be a hundred-billion-dollar merger.
Let me bring it back around before I let you go, because I want to connect today's pile to the conversation we had yesterday and the day before. We spent yesterday on Reed Jobs and his biotech fund betting on AI to chase the pharma patent cliff — a story about AI as a genuine tool inside a hard, physical, regulated business. And a couple days back we dug into that proficiency report, the one where sixty-nine percent of organizations had "taken action" on AI agents but only sixteen percent of workers actually used one. Different stories, but they rhyme with what ran through today.
Because here's the thread I keep coming back to. Whether it's Hotz screaming about freedom, or Nadella warning you're paying twice, or OpenAI shipping an agent that grinds for hours, or memory scarcity handing the win to whoever's biggest — the real question under all of it is the same one. When intelligence gets cheap and abundant, what's actually yours? Not what you rent. Not what you pass through. What do you own — your data, your evals, your learning loop, your relationship with the customer, the humming box in the corner that you actually locked down? The value in this next stretch is going to accrue to the people who can answer that question with something concrete. Everybody else is just feeding the model and hoping.
That's the menu, kiddos. Go lock down your router, question your benchmarks, and read the fine print on anybody selling you a data center in space. I'm Tony DeLuca, this has been Barely Possible, and I'll see you back here tomorrow — same table, fresh plate.