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 we've got a fresh plate of AI morsels cooling on the counter today, so grab a coffee, pull up a chair, and let's have at it. This is Barely Possible, the show where we sort the real from the hype so you don't have to burn your afternoon doing it yourself.
Let me tell you how today's menu came together, because there's a thread running through it that I didn't expect when I sat down. We spend a lot of time on this show talking about who builds the big shiny thing. The frontier model. The mega-rocket. The next chip. But the story that kept jumping out at me today wasn't about the shiny thing at all. It was about the plumbing. The stuff that sits behind the shiny thing and quietly decides who actually makes money and who gets squeezed. And I want to start with the single most consequential version of that for anybody building a product right now, which is a decision Nvidia's competitors are all going to have to wrestle with, whether they like it or not.
So let's dig into what it actually means when we say Nvidia's advantage is moving beyond the GPU.
Here's the setup, and Russell Brandom laid this out in a recent piece for TechCrunch, published the day before this edition. For the first few years of the AI boom, the story about Nvidia was simple. They were the only shop selling state-of-the-art GPUs. You wanted to train a model or serve one at scale, you were buying from Jensen. Full stop. And that made them absurdly profitable. Brandom notes the market cap went up tenfold between the start of 2023 and the middle of 2025. That's not a company, that's a phenomenon.
And then the obvious thing happened. The big cloud players, your Amazons, your Googles, started building their own chips. And the investor class started getting nervous. If Amazon's building its own silicon, if Google's had its own TPUs for a decade, how durable is Nvidia's moat, really? For about a year, the stock reflected that nervousness. More modest trajectory, as they say when they mean the party cooled off a little.
Now here's the turn, and this is the part I want you to sit with. After Nvidia's earnings this past week, a new narrative took shape, and it's this: the fight was never really just about the GPU. As the compute for AI grows into the gigawatt scale, and I want you to hear that, gigawatt, we are talking about data centers that draw power like small cities, the hard part stops being the chip that does the math. The hard part becomes orchestration. Getting the right data to the right chip at the right moment so that the whole enormous machine actually runs at peak efficiency instead of choking on itself.
And this is where it gets good, because I've been reading through exactly what Nvidia is selling now, and it's not just a GPU in a box. They're rolling out this thing called the Vera Rubin architecture. And yes, they name everything after dead scientists, it's a whole vibe. The Rubin is the GPU. But it comes paired with a Vera CPU, a Groq 3 LPX inference accelerator, racks for storage, racks for networking. Brandom's line for it is the best in the whole piece. He says, if the GPU is the engine, these are the rest of the car.
And that's the whole thing right there. Everybody's been so focused on who builds the best engine that they missed that Nvidia quietly went and built the transmission, the suspension, the fuel injection, and the dashboard.
Let me give you the concrete example, because I promised you specifics and I don't do hand-waving. Jason Hardy, who's Nvidia's VP of storage technology, told Brandom about the Vera CPU. And the problem the Vera solves is this: you can only cram so much memory into a single server. As these data centers scaled up their raw computing power, they had to scale up memory too. That's why a company like Micron, which makes memory chips, got fabulously rich in what Brandom calls the second wave of the infrastructure boom. But here's the catch. Having the memory isn't enough. You have to get that data to the GPU at exactly the right time, or the GPU sits there idle, twiddling its thumbs, burning power for nothing. And when you're trying to drive your tokens-per-watt down, when every bit of electricity matters, that traffic direction becomes the whole ballgame. Hardy said they saw upwards of a 3x improvement in these operations, that the Vera CPU lets them use their flash storage to its fullest potential without bottlenecking.
Three times. That's not a rounding error. That's the difference between a data center that prints money and one that hemorrhages it.
Now here's why I'm telling you this instead of just letting it be a chip-nerd story, because I promised the founders and the builders listening I'd keep the CUTLASS-kernel stuff to a minimum. Watch what OpenAI did with their own chip, the one they call Jalapeño. Same problem, totally different answer. Brandom quotes OpenAI's blog post from earlier in the month. They said they designed Jalapeño to minimize data movement and communication delays. The whole design philosophy is to keep the entire workload inside one connected system so the data barely has to move at all. Nvidia says, we'll build you a brilliant traffic cop. OpenAI says, we'll build a city so compact you don't need a traffic cop. Two roads, same destination: squeeze efficiency out of smarter data flow instead of just throwing more processor cycles at the wall.
And here's the punchline for you, whether you're building a startup or running an enterprise. The competition has moved. It moved to a new layer. Building a rival GPU, that matters less than it used to. What matters now is whether you can make the entire system work efficiently at scale. And Brandom is honest about this, it's not automatically a win for Nvidia. They've still got to fight the hyperscalers and the chipmakers on this new turf just like they fought on GPUs. But in the early innings? He says Nvidia looks to have a commanding lead.
So why should you care, sitting there building your thing? Because this is the shape of every cost curve you're going to ride for the next three years. The price you pay for inference, the price you pay for tokens, is going to be decided less and less by who has the fastest chip and more and more by who orchestrates the whole rack the cleanest. When your cloud bill goes up or down, this is the machinery underneath that number. And if you're placing bets on which infrastructure provider to hitch your wagon to, understand that the moat you're evaluating isn't a chip anymore. It's a system. That's a much harder thing to reason about, and it's a much harder thing to walk away from once you're in it. Lock-in doesn't come from the engine. It comes from the whole car.
Now let me connect that to something, because there's a natural bridge here. If the story on the frontier is that Nvidia is trying to own the whole stack, the plumbing and all, there's a mirror image of that happening one layer up, in the model business. And it explains why Nvidia is reportedly spending thirteen billion dollars on a company that gives most of its stuff away for free.
We covered the Hugging Face acquisition itself a couple days back, so I'm not going to re-run that tape. But there's a genuinely fresh angle in a piece by Tim Fernholz for TechCrunch that I think is worth your time, because it's about a trend, not a single deal. And the trend is this: open-weight AI companies have become the hottest acquisition targets in the Valley.
Look at the pattern Fernholz lays out. Nvidia reportedly buying Hugging Face, the GitHub of AI models, for thirteen billion. Before that, Nvidia struck a six billion dollar deal with Poolside, an open-weight model builder, and most of Poolside's employees are moving over to Nvidia. And a couple weeks before that, Stripe bought OpenRouter, the top provider of open-weight models to businesses, for more than seven billion. Now hold on, I covered the Stripe-OpenRouter deal back in the middle of the month, so that's a callback, not a new story. But stacked together like this, you see the shape of it. That is a whole lot of capital pouring into a sector whose entire business model is, and I quote the spirit of it, giving stuff away.
Why? Fernholz nails the logic. For Nvidia, it's about not being too dependent on the frontier labs and the hyperscalers. Because here's the awkward part: the labs writing Nvidia the biggest checks, the OpenAIs and the Googles, are the exact same companies building their own chips to get off Nvidia. So if the model builders are going to make chips, Nvidia figures, fine, we'll go make some of the model business. Own the open ecosystem, and you own a mass of developers you can funnel toward your chips and your standards.
But here's the part that matters for you as a builder, and it's the sobering number in the whole piece. How many companies actually use open-weight models today? According to a spending survey from Ramp, six percent. Six. And by another measure from Jellyfish, just two percent of software engineers. So all this billions-of-dollars land grab is happening over a sector that most companies aren't even using yet.
Nik Albarran from Jellyfish explains where the open models do get used, and it's telling. High-volume, repetitive stuff. Customer service chats. Where you've got a ton of the same kind of question over and over, you can tune a cheap open model to answer it for pennies. But for coding, for agentic tasks, the varied stuff that needs real reasoning? The frontier models still win, partly because the proprietary labs make them easy to access and sometimes subsidize the tokens. Albarran's read is that the main reason companies reach for open models right now isn't cost. It's control and configurability. And then the line that should stick with you: he says when your AI-driven workflows get much more mature, that's when it makes sense to invest in self-hosting your own models.
So the takeaway for the builder is not go run out and self-host tomorrow. The takeaway is: the giants are betting billions that you eventually will, and they want to own the ground you'll stand on when you do. Lin Qiao, the CEO of Fireworks, which is one of these open-model hosting companies, she goes even further. She told TechCrunch that every single app company should consider hiring an in-house researcher, build your own model off your own product data, and that the future is specialized intelligence, a model per use case. Now, that's a vendor talking her book, and you should hear it as such. But the direction of travel is real. The whole point of this land grab is that the dominance of OpenAI and Anthropic is not, as Fernholz puts it, inevitable. And the tech giants clearly agree, or they wouldn't be spending like this.
Alright, let me shift from the infrastructure story to a different kind of plumbing question, one that's about who gets to regulate what.
There's a court ruling that came down that I think deserves a spot on the menu, and it's the Kalshi decision. Now, one housekeeping note, and I'm gonna be straight with you the way I always am: this ruling surfaced fresh today but the decision itself is from a few days back, so I'm framing it as a recent ruling, not something that just dropped this morning. Fair's fair.
Here's the deal. Kalshi is a prediction market. You may have seen their ads at bus stops. They let you bet on outcomes, and lately, a whole lot of those outcomes have been sports. And Kalshi's clever, legally speaking. They advertise themselves, and this is a direct quote from the ruling, as the first app for legal sports betting in all fifty states. But when Nevada's Gaming Control Board sent them a cease-and-desist saying, hey, that's gambling, and gambling is our jurisdiction, Kalshi's response was essentially: no no no, these aren't bets. These are swaps. Financial instruments. And swaps are regulated by the CFTC, the federal commodities folks, not by you, Nevada. So buzz off.
And the Ninth Circuit, a panel of three Trump-appointed judges, mind you, unanimously said: nice try. Judge Ryan Nelson wrote the opinion, and he reached all the way back to Shakespeare for it. Romeo and Juliet. That which we call a rose by any other name would smell as sweet. And Nelson's version: placing sports bets, even when called by another name, is still gambling. I love a judge who quotes the Bard to tell you your financial engineering is baloney.
The concurring opinion from Judge Kenneth Lee is even better for the regular person. He said, few people would describe the New York Mets' latest loss of a game as an event in the swap sense. And then he goes on, and I'm not making this up, to muse that maybe a Mets loss has marginal economic impact because some fans guzzle more beer to drown their sorrows, but that it's fanciful to say a single game in a 162-game season is associated with a financial consequence the way a real swap is. As a guy from the Bronx, I have some feelings about a federal judge using Mets losses as his example of something economically meaningless, but I'll let it go. He's not wrong about the beer.
Now, why does this matter to you as a builder, especially if you're anywhere near crypto or fintech? Two reasons. One, there's now a circuit split. The Third Circuit, in a case against New Jersey, ruled the opposite way, that these sports wagers are swaps. When two federal circuits disagree, the odds of the Supreme Court taking it up go way up. So this isn't settled. This is round one. Two, and this is the strategic lesson, the whole Kalshi play was a regulatory arbitrage move. Wrap a thing that's regulated one way in the language of a thing that's regulated another way, and slip through the gap. And the Trump administration has been friendly to prediction markets, Donald Trump Jr. advises both Kalshi and Polymarket, so there was a real tailwind here. But the court basically said the label doesn't change the substance. If you're building a business whose entire model depends on which regulatory box you get to check, understand that a court can look right through the box. The rose-by-any-other-name doctrine cuts both ways.
Now let's move from courtrooms to factory floors, because there's a real business story in humanoid robots and it's got a China angle that founders should be tracking.
Kirsten Korosec at TechCrunch had a piece on Chinese automakers piling into humanoid robots, and the numbers are eye-popping. Xpeng's robotics unit just raised more than nine hundred million dollars at a post-money valuation north of six point three billion. The company called it the largest single-round private financing ever recorded in China's embodied AI industry, that's their term for AI built directly into physical machines. And the round was led by IDG Capital with Tencent and Alibaba in there too. When Tencent and Alibaba are both writing checks into the same round, that tells you the smart money in China has decided robots are the next thing.
And it's not just Xpeng. Chery's robotics unit, AiMOGA, is reportedly prepping for an IPO. BYD unveiled a humanoid called Xiao Di. Changan, GAC, Li Auto, SAIC, Seres, they're all in. And the why is the interesting part. Michael Dunne, who runs an advisory firm called Dunne Insights, put it plainly. He said Xpeng's founder He Xiaopeng sees razor-thin profit in cars on the near horizon. Robots look much more promising. That's the whole thesis in one sentence. The car business in China has become a bloodbath of price competition, and these guys are looking at humanoid robots as the escape hatch to fatter margins.
And here's the strategic tension Dunne identifies, and this is the part for the builders. He says the Chinese automakers bring a manufacturing edge, they have all the hardware to get the job done. The open question is whether they can catch Tesla on the AI side of the equation. That's the whole game. Making a robot walk is a hardware problem, and China is very, very good at hardware problems. Making a robot actually useful, able to learn nearly any task, that's an AI problem. And that's still up for grabs. Meanwhile on the Western side, Hyundai's bringing Boston Dynamics' Atlas to a Georgia factory, working with Google's DeepMind, aiming to have robots doing parts sequencing by 2028. Even Mobileye bought a humanoid startup for nine hundred million. So the whole industry is converging on the same finish line, which is commercial deployment at scale. If you're a founder thinking about where the next platform shift comes from, the tell here isn't the robot videos. It's Tencent and Alibaba deciding this is where the money goes.
Now shift gears with me from the biggest hardware bets to the smallest, most concentrated bet in venture capital, because there's a terrific interview that gets at something every founder should hear.
Connie Loizos sat down with Vijay Pande. Now, Pande spent over a decade at a16z, grew their bio and healthcare practice into something managing close to four billion dollars. And then he walked away to start something tiny. His new firm, VZVC, does maybe five concentrated bets a year instead of dozens. No associates. And here's the detail that jumped out at me: they were planning to hire associates, and then they built out AI agents and found they just didn't need to. That's a venture firm quietly automating the junior-analyst rung of its own ladder, and nobody's making a big deal of it.
But the meat of the interview, the part I want you to chew on, is about data. Pande makes this point that biology is one of the few places where AI can't just scrape the internet. Unlike text, you can't hoover up biological data off the web, and you can't distill it from one model to another. So nearly every biotech company ends up building its own walled-off dataset. And that creates a real problem for the whole promise of AI in medicine, because as Pande puts it, LLMs work because there's so much data to learn from. When the data simply isn't there, AI can't magically solve the problem. That's his answer, by the way, to what's overhyped. He says the hesitance isn't doubt about AI, it's doubt about the data.
And here's the founder wisdom buried in there, the thing that applies whether you're in bio or not. Loizos asks what he's gotten right and wrong over his career. And Pande says the thing that took him time to appreciate is that as seductive as the coolest technology is, it always comes back to go-to-market. He tells his founders, especially the ones coming from the science or product side, to take all their brilliance and apply it to go-to-market, because go-to-market is at least as hard as the technology, if not harder. I've said versions of this on this show more times than I can count, but it lands different coming from a Stanford chemistry professor who built a four-billion-dollar practice. The tech is the easy part. Getting somebody to actually buy the thing, that's the mountain.
Alright, one more before we start wrapping, and this one's about sovereignty, which has become the word of the year in European tech.
Dominic-Madori Davis filed a piece from TechBBQ, this big Nordic conference in Copenhagen, and the whole event kept circling back to one question: who's actually in control of this AI wave? And the reason it hit so hard this year is that earlier in the year, Anthropic's models Mythos and Fable became unavailable to users outside Europe for a stretch. And that little disruption made a lot of European builders realize, hey, we're renting the engine of our entire economy from two other geopolitical powers, the US and China. One startup exec told Davis it seriously disrupted his software team. Another shrugged it off, said sure, it might cause headaches one day, but for now everything's fine.
And that split reaction is the whole story, isn't it? That's the exact conversation every builder outside the big US labs is having with themselves. Do I care about controlling my own infrastructure, or do I just want the best model and I'll deal with the dependency risk later?
There were two quotes from that conference I want to leave you with, because they cut in opposite directions and I think that tension is honest. Meredith Whittaker, the Signal president, moderated a privacy panel and she did not mince words. She said this AI era is creating a data collection apparatus, and that the labs are using clever marketing to make people forget about the collateral consequences of hoovering up all that data. She specifically called out AI assistants getting jammed into operating systems, the kind of thing where an AI is reading your iMessages. And her business read was sharp: she said there is still a huge market need for privacy, and particularly with these sovereignty concerns, we will have customers here. That's a founder identifying an underserved market in real time, which I respect.
And then the other pole. Ellen de Brever, an angel investor at the conference, framed the whole agentic moment beautifully. She said the age of AI agents isn't really a story about machines gaining agency. It's a story about humans deciding what to give up. Sit with that one. Every agent you deploy, every workflow you hand off, is a small decision about control that you're making, whether you're conscious of it or not. As a builder, you are constantly making that trade, speed and convenience on one side, control and independence on the other. And the European conversation is just that same trade, played out at the level of a whole continent.
So let me tie the ribbon on this, because there's one thread running through everything today, and it's not the usual governance-versus-capabilities song we've all heard a hundred times. It's simpler and it's more useful. Today's whole menu was about who owns the layer underneath your product. Nvidia isn't fighting over the chip anymore, it's fighting over the whole rack, the orchestration, the plumbing. The giants are spending thirteen billion dollars to own the open-model ecosystem you might migrate to in two years. A court just reminded everybody that you can't relabel your way out of the regulatory layer you actually sit in. Chinese conglomerates are deciding the robot business is the next layer worth owning. And a room full of Europeans spent a week agonizing over the fact that the intelligence layer of their economy is rented from abroad.
Every one of those stories is somebody making a move for the ground you're standing on. Not the shiny thing you're building on top. The ground. And if you're a founder or a builder, the practical instruction is this: know what layer you actually own, and know which ones you're renting. Because the people writing the biggest checks this week are all making very deliberate bets about which of those layers you'll be forced to pay rent on down the road. Watch the plumbing. That's where the money moves before anybody notices.
That's the plate for today, kiddos. As always, I tried to protect your time and give it to you straight, no slick, no hype. This has been Barely Possible, I'm Tony DeLuca, and I'll catch you on the next one. Take care of yourselves out there.