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 pick through the day's tech and AI headlines, throw out the spoiled stuff, and serve you only what's worth chewing on. Pour yourself something, settle in, because today we've got a menu that runs from a five-billion-dollar plan to nationalize the AI industry, to Amazon trying to muscle into Nvidia's lunch line, to a couple of companies finding out the hard way that promising self-driving features you never deliver is, in fact, a lawsuit.
But let me start where the real story is, because I want to lead with the thing that's gonna matter most to anybody building a company right now. And that's not the splashiest headline. The splashiest headline is Bernie Sanders waving a seven-trillion-dollar plan to give Americans, quote, control of the AI industry. We'll get to Bernie. But the consequential story, the one that actually reshapes how the biggest player in this game is gonna behave for the next year, is OpenAI loading up the wagon before its IPO. So let's dig into that.
According to a piece from Rebecca Bellan over at TechCrunch, OpenAI is bulking up in the lead-up to going public, and they did it with two hires in the same week that tell you a lot about how they're thinking. First, they landed Noam Shazeer from Google DeepMind. Now, for folks who don't live and breathe this stuff, Shazeer is one of the co-inventors of the Transformer. That's the architecture underneath basically every large language model you've ever talked to. The T in GPT, the T in every model that matters. That's his work. So you've got the guy who helped invent the engine of the entire industry leaving Google's lab and walking over to OpenAI right as OpenAI is dressing up for Wall Street. That's not a small thing. That's a statement of intent about where the talent gravity is pulling.
And the second hire is the one that I want you to pay attention to, because it tells you something different. They also brought on Dean Ball, a former Trump AI policy official. So in one week, OpenAI hires a foundational researcher and a Washington insider who knows the policy machinery. And if you've been listening to this show, you know exactly why that second hire matters right now.
Let me connect this. We've spent the last couple weeks talking about the Anthropic situation — the Commerce Department forcing Anthropic to disable foreign access to its most powerful models after a jailbreak scare, the whole Fable and Mythos mess, world leaders at the G7 worrying that America could just flip a switch and turn off the AI they depend on. We covered that on the eighteenth, the Macron and Modi anxiety about the kill switch. Now here's OpenAI, watching their biggest rival get body-slammed by Washington, and what do they do? They go hire someone who speaks the government's language. That's not a coincidence. That's a company looking at Anthropic getting knocked flat and saying, we are not gonna be the ones who didn't have anyone in the room.
Because here's the lesson sitting underneath all of this, and it's a lesson for any founder building something the government might one day consider strategically important. The moment your product becomes integral to the economy or national security, dealing with the people in power stops being a distraction from your real job. It becomes part of your real job. Anthropic, by most accounts, learned that the slow and painful way. OpenAI is watching and front-running it. They're hiring the relationship before they need it. That's the move. That's the whole move.
And the IPO context makes it sharper. When you're about to go public, you cannot afford to be the company that's vulnerable to having its product switched off by a regulator nobody at your shop has a phone number for. The Shazeer hire is about capability — keep building the best models. The Dean Ball hire is about not getting Anthropic'd. Two hires, two halves of the same defensive crouch. So if you're watching OpenAI's pre-IPO moves as a tea-leaf for the industry, the tea leaves say: the era where you could just build and ignore Washington is over. Period.
Now, let me stay in the OpenAI orbit for one more beat, because there's a thread that connects to their finances. We talked the other day about the leaked, fully audited numbers — the eye-watering loss figures, the thirty-billion-dollar-plus non-cash accounting charge tied to the corporate restructuring, the fact that when you strip that out the actual operating loss looks a lot more digestible, and the genuinely interesting bit that they appear to be turning a tidy margin on inference itself. I'm not gonna re-litigate all of that today. But hold that picture in your head — a company that's making money selling tokens, bleeding cash on training and headcount, sitting on a big pile of cash, and now arming up on both the research and policy fronts before it goes public. That's a company preparing for a fight on multiple fronts. Keep that in mind as the year unfolds.
Alright. Now let's swing over to Bernie Sanders, because if OpenAI is preparing for the regulatory environment quietly, Bernie just showed up with a megaphone and a number with a lot of zeros.
Ashley Belanger at Ars Technica reports that Senator Sanders unveiled a seven-trillion-dollar plan to, as the headline puts it, give Americans control of the AI industry. The centerpiece is what's being called an AI wealth fund. And the framing here, the summary says it plainly, is that the biggest AI firms will likely recoil at this. Yeah, I'd imagine they will.
Now let me be straight with you about how to read this, because there's a way to overreact and a way to under-react and the truth is in between. Seven trillion dollars is not a number that passes Congress next Tuesday. This is a proposal, it's a marker, it's Bernie doing what Bernie does, which is plant a flag way out on the horizon and drag the conversation toward it. The biggest AI firms recoiling is not the same as the biggest AI firms being in danger. So don't panic if you're building on top of these platforms.
But here's why I don't want you to ignore it either. The idea underneath the wealth fund — that the public should get an ownership stake in the upside of an industry that's being built partly on public infrastructure, public research, public data, and is increasingly being justified to all of us on national security grounds — that idea is not going away. We just talked about how AI is getting wrapped in the flag, declared vital to the nation, defended in court as too important to shut down. Well, the flip side of that argument, the side Bernie is pressing on, is simple: if it's that important to the nation, why does the nation own none of it? You can't have it both ways forever. You can't tell a judge your data center is a matter of national economic and energy security and then tell a senator it's just a private company that owes the public nothing.
So the Bernie plan is not the threat. The Bernie plan is the early tremor of a tension that every builder in this space should expect to feel more of: the more AI gets framed as a national strategic asset, the more political pressure there's gonna be to socialize some of its returns. Whether that takes the shape of a wealth fund, a windfall tax, equity stakes, whatever — the direction of travel is toward the public wanting a cut. Watch the framing, not the dollar figure.
Now shift from the politics of AI ownership to the economics of the hardware underneath it, because there's a really interesting competitive move from Amazon today.
Julie Bort at TechCrunch reports that Amazon is hoping to challenge Nvidia more directly — by selling its own AI chips. AWS is in talks to sell its in-house chips to other data centers. And CEO Andy Jassy has called this a fifty-billion-dollar opportunity for the company. Let me unpack why this is a bigger deal than it sounds.
For years now, Amazon has been designing its own AI silicon — the Trainium and Inferentia chips — but the whole point was to use them inside AWS, to lower Amazon's own costs and reduce its dependence on Nvidia. That's the standard hyperscaler play: build your own chips so you're not paying Nvidia's margins on every GPU. Google does it with TPUs, everybody's been at it. But selling those chips to other data centers, to outside customers — that's a different game entirely. That's Amazon saying, we're not just trying to save money on our own compute, we want to be a chip vendor. We want to compete with Nvidia for the actual customer.
And fifty billion dollars is the number Jassy is putting on that opportunity. Now, is that real or is that the kind of number CEOs throw out to make analysts perk up? Bit of both, probably. But here's the structural thing that matters for you. Nvidia's dominance in AI hardware has been one of the great chokepoints of the entire boom. Everybody complains about it. The cost of GPUs, the scarcity, the waitlists, the markups. If a credible second source — and Amazon is about as credible as second sources get — starts seriously selling alternative AI chips to the broader market, that's the beginning of pricing pressure on the most expensive input in your whole AI stack. It doesn't crack Nvidia overnight. CUDA and the software ecosystem are a moat that money alone doesn't fill. But the moment the biggest cloud provider on earth decides it wants to sell shovels and not just dig with them, the gold rush economics start to shift. For anybody whose business model is sensitive to the cost of inference — which is increasingly everybody — that's a development worth tracking closely.
And speaking of inference economics, here's a tidy little companion story. Dominic-Madori Davis at TechCrunch reports that Baseten, the AI inference startup, is reportedly close to finalizing a one-point-five-billion-dollar round at a thirteen-billion-dollar valuation — and this is months after their last mega round. The phrase in the piece is the inference gold rush, and that's exactly what it is. Think about what these two stories say together. Amazon wants in on selling the chips that run inference. Investors are throwing one-and-a-half billion at a company that helps you run inference efficiently. The whole center of gravity in this industry has shifted from training the models to serving them — from building the brain once to answering a billion questions cheaply. That's where the money is going because that's where the recurring cost lives. If you're a founder, the signal is loud: the value, and the investment, and the competition, is all migrating to the serving layer. The model is increasingly a commodity input. What you do with it at scale, cheaply and reliably, is the business.
Let me pivot from the money flowing in to the money flowing out — specifically, a couple of companies getting in trouble for promises they made and didn't keep.
Kirsten Korosec at TechCrunch reports that Rivian owners have filed a class action lawsuit alleging false promises on self-driving features. The plaintiffs claim Rivian falsely promised, for years, that it would bring hands-free driving to its first-generation R1 vehicles. And it never showed up. Now, we mentioned Rivian recently in another context — they cut hundreds of workers after the R2 started shipping. So this is a company under real pressure, and now its earliest, most loyal customers are in court saying: you sold me a car on a promise of a feature that never arrived.
I want to sit on this one for a second because it's a pattern, not a one-off. We have spent years in this industry treating the roadmap as the product. You sell the dream. You sell what the car will do someday, what the software will unlock in a future update, the autonomy that's always just around the corner. And for a long time customers played along, because the whole culture was: invest in the vision, the features will catch up. Well, the catching up didn't happen, and now the bill is coming due in a courtroom. This is the same energy as the FERC story we'll get to, the same energy running through half the headlines this year — the gap between what was promised and what shipped is becoming a legal liability, not just a credibility problem.
And for founders this is a genuinely important shift in the weather. If you've been selling features that are, let's say, aspirational — coming soon, on the roadmap, in beta, any day now — the Rivian suit is a warning that aspirational marketing has a statute of limitations on customer patience. At some point, years of we'll-deliver-it-eventually becomes, in the eyes of a plaintiff's attorney, a false promise. Be careful what you put in the brochure. The autonomy you're selling has to eventually drive the car.
Now let's talk about the grid, because here's a story that lands right on the seam between AI ambition and physical reality.
Tim De Chant at TechCrunch reports that AI data centers just got a government-mandated fast lane to the grid. The federal energy regulator, FERC, told grid operators to give data centers a fast lane for interconnections. So if you're trying to plug a giant new AI data center into the power grid, you used to have to wait in the same long queue as everybody else. Now, by federal direction, you get to cut the line. But — and this is the whole story — the piece is clear that FERC failed to address the actual electricity supply shortages. So you can get to the front of the line faster, but the line leads to a power plant that may not have enough juice to go around.
Think about what that actually does. You've made it easier to connect demand without doing anything to increase supply. That's not a solution, that's a faster path to a bottleneck. And it raises an uncomfortable equity question that's gonna get louder: when a data center cuts the interconnection line, who's behind it? Homeowners. Local businesses. Manufacturers. The fast lane for AI is, by definition, a slower lane for everybody else, unless the supply grows to match. And we covered a related thread recently — the Justice Department going to court to defend a company's unpermitted gas turbines as a matter of national economic and energy security. So you see the shape of it. The build-out is colliding with the physical limits of the power system, and the government's instinct so far is to clear the path for AI rather than expand the road for everyone. That tension — AI's appetite for power versus the grid's ability to feed it — is one of the defining infrastructure fights of this whole era, and FERC just made one move in it. Watch where the next move comes from, because somebody — a state, a utility, a ratepayer advocate — is gonna push back.
Let me bring it back down to the human scale for a minute, because there's a cluster of stories today about people just... wanting their attention back.
Amanda Silberling at TechCrunch has a piece arguing that the smartphone era created an attention crisis, and that, quote, slowtech is fixing it. The line that jumped out — people just really want to take back control of their time, their lives, their attention, and they're down for whatever helps them do that. And there's a companion piece from Lauren Forristal about a new app called Mivo Scrolling, which launched last month and takes a mindful approach to managing screen time. Instead of slapping hard limits on you, Mivo tries to make you aware of the doomscroll loop you're stuck in rather than just locking the door.
Now why am I, a guy who mostly talks about agent frameworks and chip economics, spending time on screen-time apps? Because there's a real business signal in here. For fifteen years the entire consumer tech industry was built on one objective: capture and hold attention. Maximize time-on-app. Engagement, engagement, engagement. And what we're watching now is the emergence of a counter-market — a whole category of products whose value proposition is the exact opposite. We will help you use us less. We will help you put the phone down. That's a fascinating inversion. When a market gets so saturated with extraction that people start paying for relief from it, that's a signal there's an underserved need worth building into. If you're a founder looking for white space, the backlash against the attention economy is becoming a product category, not just a think-piece. People are voting with their wallets for tools that give time back. That's worth a look.
There's a darker cousin to this attention story, and it's the New York Times reporting on AI apps and students cheating. I'll keep this short because I don't have the full piece in front of me, but the broad shape is the one we all see coming — a wave of apps designed to help students cheat with AI, and schools scrambling to respond. The connective tissue with the attention story is this: we keep building powerful consumer tools optimized for the wrong thing. Optimized for engagement, optimized for getting the answer without the learning. And then we act surprised when the externalities show up — the doomscrolling, the cheating, the kids who can produce an essay but can't think through one. The builders who win the next decade might be the ones who figure out how to align these tools with what people actually need, not just what keeps them clicking. That's not a moral lecture, it's a market observation. The cleanup is its own opportunity.
Alright, let me run through a few more that founders should have on the radar, then we'll bring it home.
Snap is spinning off its AI video team into a new company called Dotmo, and the reason given, straight up, is costs. Lucas Ropek at TechCrunch reports that Dotmo will be made up of current Snap staff leaving to focus on AI video development. This is the second internal unit Snap has spun off. Read that as what it is: AI video is expensive to develop, and a company under financial strain — and Snap's had a rough stretch, their stock took a dive recently after those pricey AR glasses — is deciding it can't carry that cost on its own balance sheet. So it externalizes it. The interesting question for the ecosystem is whether spinning out becomes a pattern: big companies that can't afford to fund frontier AI internally cut these teams loose to go raise their own money. It's a way to keep a foot in the game without bleeding for it. Watch whether others copy the playbook.
On the security front — and this one matters for anybody shipping consumer hardware — Dan Goodin at Ars Technica reports Apple patched a high-severity eavesdropping vulnerability in its Beats Studio Buds. And the detail that should make you wince: the vulnerability was disclosed twelve months ago, and it affects multiple manufacturers. Twelve months. A high-severity eavesdropping flaw in earbuds, meaning theoretically someone could listen in, sitting unpatched for a year across multiple vendors. The lesson for builders is the boring one that everybody nods at and nobody acts on: your disclosure-to-patch timeline is a security posture, and a year is not a good one. If you ship hardware with microphones in it, the clock on a known vulnerability is not a suggestion.
Quick one on YC. Marina Temkin and Dominic-Madori Davis at TechCrunch put together the eleven standout startups from Y Combinator's Demo Day, according to VCs — the Spring 2026 batch. And the number that tells the story: some of these commanded valuations of over a hundred and seventy-five million dollars. At demo day. Coming out of an accelerator. I'm not gonna read you the list of eleven, but I'll give you the takeaway, because the takeaway is the point. A hundred-and-seventy-five-million-dollar valuation for a company that's, what, a few months old? That's the froth talking. That's capital so eager to find the next AI winner that it's pricing seed-stage companies like they've already won. If you're a founder raising right now, that's both good news and a warning. Good news because the money is there and it's generous. Warning because valuations like that set a bar you then have to grow into, and growing into a hundred-and-seventy-five-million-dollar story is a very different life than growing into a twenty-million-dollar one. Raise at a number you can actually outrun.
Let me touch the geopolitics quickly, because there are a few threads here worth a mention even if they're not the main course. There's a ProPublica investigation, reported via Ars Technica by stin Elliott and Joshua Kaplan, that before the SpaceX IPO, investors in China secretly acquired stakes — and one previously unreported SpaceX investor has ties to Chinese military contractors. Now, SpaceX is a defense contractor with deep national security entanglements, and the IPO we've been talking about for two weeks now is the biggest in recent memory. So the idea that there were quietly-acquired Chinese stakes, including one with military ties, in the run-up to that — that's the kind of thing that draws subpoenas. I'm not gonna speculate beyond what's reported, but file it under: the national security scrutiny on this company is only going to intensify, and that's relevant to anyone holding or eyeing that stock.
And separately, there's a recent report on Taiwan — Jeremy Hsu at Ars Technica — about Taiwan ramping up drone production for its own defense and for the US military as China looms. Taiwan's defense spending on drones could also boost its business overseas. The builder angle here is the same one we keep circling: defense tech and dual-use manufacturing is a real, growing market, and the geopolitics are turning Taiwan into a drone hub the way it's already a chip hub. If your business touches hardware supply chains, the militarization of that supply chain is a force you're going to feel.
Let me close out the AI-and-society file with one quick data point, because I think it's a healthy gut check. Match — the dating company — put out a survey, reported by Amanda Silberling at TechCrunch, finding that almost half of US singles, about forty-seven percent, feel negatively about AI in dating. But the nuance is the good part: a lot of those same people are open to AI helping with profile touch-ups and conversation starters. So they don't hate the AI helping them get ready for the date. They hate the idea of the AI being the date — of authenticity getting hollowed out. And I think that's a really clarifying line for anybody building consumer AI. People will accept AI as the assistant. They get queasy when it becomes the substitute. Help me write the message, fine. Pretend to be me, or pretend to be a person who isn't there — not fine. That boundary, assistant versus substitute, is gonna define which AI consumer products get loved and which get rejected. Keep your product on the right side of it.
A couple of things I'm deliberately not making a meal out of today, just so you know I saw them. NASA asked Northrop Grumman to stop work on that one-point-one-billion-dollar lunar HALO module — a habitation module for the Gateway program, now likely shelved, with the company reassigning affected employees. Big number, real story for the space crowd, but not central to a builder's day. There's an older report resurfacing about FDA advisors voting to approve a Moderna mRNA vaccine after a Trump official refused to review it back in February — that's a months-old development getting renewed attention, not breaking news, so I'm flagging it as such rather than presenting it as today's event. And there's a resurfaced item on Google's Android developer verification timeline, which traces back to changes announced last year — if you publish Android apps, the verification regime is worth your attention, but it's not new today. I'd rather be honest about the freshness than dress up an old story in fresh clothes.
And one I'll wave at fondly on the way out the door — Ars reports the first long-duration resident of the International Space Station, cosmonaut Aleksandr Samokutyaev, has died at fifty-six. Two expeditions, two spacewalks, three hundred and twenty-two days in space. No business angle there. Just a guy who spent the better part of a year off the planet, and now he's gone. Sometimes a story doesn't need a takeaway. It just deserves a moment.
So let me tie the through-line together before I let you go. Look at the spine of today. OpenAI hiring a policy insider before its IPO. Bernie pushing a wealth fund to claw back public ownership. The Justice Department defending turbines as national security. FERC clearing a fast lane for data centers while the grid groans. Even the SpaceX China-stakes investigation. Every one of those is the same underlying fact pressing in from a different direction: AI has crossed over from being a product into being a political object. It's strategic infrastructure now. And the minute that happened, the rules changed for everybody building in it. You're not just shipping software anymore. You're operating inside a system where governments, regulators, senators, and courts all believe they have a legitimate claim on what you're making.
The builders who internalize that early — the OpenAIs hiring their Dean Balls before they need them — those are the ones who don't get caught flat-footed. The ones who keep acting like it's 2021 and the only thing that matters is the model and the demo, those are the ones who end up like Anthropic, on a plane to Washington trying to explain themselves to people who hold the off switch. Build the relationships. Watch the framing. And don't promise the self-driving feature you can't actually ship.
That's the menu for today. I'm Tony DeLuca, this has been Barely Possible, and I appreciate you spending a little of that precious, hard-to-reclaim attention right here with me. Go build something worth the power it draws. Catch you next time.