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 menu of delicious morsels sitting on the counter today, so grab a coffee, pull up a stool, and let's have at it. This is Barely Possible, the show where I read the technical stuff so you don't have to, and I try real hard to tell you which of it actually matters for the thing you're building.
Today we've got a real one to chew on. Nvidia is reportedly buying Hugging Face for about thirteen billion dollars. And I want to spend some real time on that, because if you build software, if you touch AI models at all, this one moves the furniture in a room you're standing in. But before we get to the big plate, let me walk you through the rest of what's on the table, because there's a bunch here that a founder should know about.
Let me start with a company doing something so simple and so decent it almost made me suspicious. You remember Panic, the folks behind that quirky little yellow gaming handheld with the crank, the Playdate? Here's what happened. Back when the Trump tariffs were in effect, Panic was getting hit with a nineteen percent tax to import their hardware, and because their margins on that little thing are thin, they couldn't just eat it. So they did the honest thing and put it right there on the receipt as a line item. You paid the tariff, you saw the tariff. Then the Supreme Court ruled back in February that those tariffs were illegal and ordered the government to refund the companies that paid them. And now that Panic has started getting its refund money back, they turned right around and refunded every customer who paid that surcharge. Co-founder Cabel Sasser put it plain: "It's just not our money to keep, and it felt really good to give it back."
Now here's why I'm telling you this, and it's not just because it's a nice story. Look at the contrast. Over a hundred other companies, including Microsoft and Sony, are facing consumer lawsuits over tariff costs they passed along. Nintendo, in its own legal filing, argued that customers "voluntarily paid a purchase price and received a product in return" and that there's nothing unjust about keeping the refund money the government sends back. Nintendo said it got three hundred million dollars back and that those costs were "primarily borne by the company." So you've got two totally different postures here. One little indie shop says here's your money back, no drama, and the giants are lawyering up over whose money it is. If you're a founder thinking about trust and brand, that's the whole lesson right there. The transparent line-item that let Panic hand the money back cleanly? That's a feature, not a bug. You can't refund a tariff you buried inside "market conditions."
Alright, let me shift from good corporate behavior to some corporate legal exposure, because we've got a couple of ugly ones and I'm not going to dwell but you should know they happened.
First, xAI. There's a lawsuit filed this week, a proposed class action, alleging that Elon Musk's xAI trained Grok on child sexual abuse material. The plaintiff, a woman using the name Jane Doe, was abused as a preschooler in the early two-thousands, and her images were hashed years ago by the groups that track this stuff. According to the complaint, she got a notification that AI-generated versions of that abuse, depicting her, had turned up on xAI. This is the first case to accuse xAI specifically of training on this material, and I want to be careful here, these are allegations, not proven facts, and the complaint itself admits it's thin on the direct training claim. But the more detailed part of the argument is about the pipeline. Grok's terms treat public posts and Grok's own outputs as training data by default. So the lawsuit's logic is that anything ingested before takedown keeps shaping the model even after the original image is pulled. For anyone building generative products, that's the part to sit with. Your default data ingestion terms are not a footnote. What flows in by default is a liability surface, and "we deleted it later" may not be a defense, because you can't cleanly un-train a model.
And staying on the theme of what these systems soak up, there's a related one over at Meta. Buried in Meta's eighteen billion dollar child-safety settlement with twenty-nine states is a provision that raised my eyebrow. As part of the deal, Meta has to build a model to detect users under thirteen. Fair enough. But to do that, the states agreed, and I'm quoting the settlement language here, "fully, finally, and forever" not to bring COPPA claims or similar state claims over Meta's use of kids' data for that purpose. So in a case about protecting children's data, the state AGs gave Meta a legal pass to retain and train on children's data. One attorney quoted in the reporting said the carve-out could "disincentivize future enforcement actions." Now, there's an independent auditor involved, and Meta can't use the data for ad targeting. But here's the builder's takeaway, and it's a genuinely hard problem, not a gotcha: to build age-assurance that actually works, you need deep behavioral signal on the exact population you're trying to protect. The tool and the harm draw from the same well. That tension is going to show up in every consumer AI product that needs to know who its users really are. Keep it in your peripheral vision.
Now let me get us to the story I actually want to sit with, because this is the consequential one for builders. Nvidia is reportedly acquiring Hugging Face for roughly twelve point nine billion dollars.
If you don't live in this world every day, here's the quick orientation. Hugging Face is, more or less, the GitHub of AI models. It's where developers and researchers go to find open-weight models, download them, run them, fine-tune them, and upload the variants back up. It became the default warehouse. And a warehouse that everybody uses is not just a business, it's infrastructure. The report comes from The Information, confirmed by a second source to CNBC. The deal's not closed, so it could still hit a snag, but both sides appear to be actively pushing it. And here's a detail worth chewing on: Hugging Face reportedly had other suitors, including Salesforce, and Nvidia wasn't even its only past investor. Google and Microsoft had money in it too.
So why does Nvidia want it, and why should you care? Let me give you the strategic read, because this is where it gets interesting.
Nvidia sells the shovels. The GPUs. That's the whole empire. But there's a threat on the horizon that Wall Street doesn't talk about enough. The frontier labs, OpenAI, Anthropic, the big ones, they've all started building their own custom silicon. You saw it in the fresh material this week, OpenAI's got a custom inference chip they're calling Jalapeño. Everybody wants off Nvidia's leash. Vertical integration. Freedom from Nvidia's pricing power. And that is an existential slow-bleed risk to Nvidia's whole position.
So what does Nvidia do? It buys the layer that sits above the hardware. The place where the entire open-model ecosystem lives. Think about the leverage there. Hugging Face is the de facto home for open-weight models, and open weights are precisely the alternative to those closed frontier labs that are trying to escape Nvidia. If you own the town square where open models get distributed, you have a lot of quiet influence over which hardware those models are optimized for, discovered on, and encouraged to run on. Nvidia even has its own family of open models, Nemotron. And Nvidia's own cloud AI business never really got off the ground before. This gives that engine a jump start.
Now here's the part the article notes that I want to underline for you, because Hugging Face reportedly isn't even profitable. Ninety-seven percent of its users are on it for free. And it doesn't matter. This is not a revenue-multiple deal. This is a control-of-infrastructure deal. When one company decides a not-yet-profitable business is worth thirteen billion dollars, they're not buying the income statement, they're buying the position. And the position is: the chokepoint of the open-model world.
Let me put on my skeptic hat, because that's the job. If you're a startup building on open weights, and a whole lot of you are, because as we've covered before on this show, the economics of running frontier cloud models are brutal and open models are how lean teams keep their token bills sane, you should ask a very direct question. When your critical distribution infrastructure gets owned by the dominant hardware vendor, whose interests does that infrastructure serve over the long run? Today Hugging Face is neutral ground. Everybody's welcome, every framework, every chip. Nvidia will absolutely say it stays that way, and in the short run it probably does, because breaking that neutrality would torch the trust that makes the asset valuable. But incentives are incentives. Over years, the pressure is always going to be to nudge the ecosystem toward the owner's hardware. So the thing to watch, the concrete thing, is whether tooling and optimization on Hugging Face starts quietly favoring Nvidia silicon, and whether the AMD and custom-chip paths get the same love. Put that on your dashboard.
And one more angle, because Hugging Face isn't just language models anymore. They've leaned hard into robotics models, their LeRobot library, physical-AI stuff. Nvidia is already the giant in that space. So this acquisition opens a robotics door too. Which, funny enough, connects directly to the next thing I want to talk about.
Because on the same day this Nvidia news broke, Anthropic put out something that's aimed at that exact physical-AI frontier, and it's the piece I want to actually dig into. Let's shift from who owns the model warehouse to what the models are being pointed at next: the physical world.
Anthropic announced a research preview of something they're calling the Model Hardware Standard. And I read the whole long writeup, all of it, the partner case studies and everything, so let me tell you what's real here and what's marketing.
The plain-English version first. Right now, if you run a science lab or a factory floor, you've got a bunch of expensive machines, microscopes, robotic arms, liquid handlers, plate readers, and they mostly don't talk to each other. Every one has its own weird software, its own programming interface, sometimes just a graphical screen a human clicks through with no real API at all. Getting them to coordinate is a nightmare of custom glue code that can take weeks or months and cost anywhere from thousands to millions of dollars. Anthropic's pitch is a standardized driver layer that translates between all these devices and an AI agent, so the agent can read from them, write to them, and orchestrate them together. And because it plugs into the same Model Context Protocol we've talked about, a scientist can drive the whole rig in natural language.
Now, the marketing line is "this is how a century of progress can condense into a decade." I'm going to ask you to set that aside, because the actual case studies are more interesting and more honest than the tagline, and honestly the honesty is the reason I trust it more.
Let me give you the one that grabbed me, the quantum computing example from a company called QuEra. They build quantum computers using neutral atoms, and the whole thing depends on lasers holding an insanely precise frequency, roughly one part in a trillion. The article gives you the analogy: that's like measuring the distance from the Earth to the Moon to within the width of a human hair. When the laser drifts off and "unlocks," a human expert has to babysit it back, and it takes five to ten minutes. And in a university lab, when that lock drops at two in the morning, somebody literally has to wake up and drive in.
Before this, QuEra had a team, a laser engineer, a software engineer, an algorithms person, a tester, spend several months building a script to automate the recovery. And it only worked fifty-eight percent of the time and took a hundred and fifty seconds a shot. So they handed the same problem to Claude through this new standard. They set up an agent loop, ran it overnight, unattended, hundreds of iterations, each pass improving its own recovery script. By morning, recovery was down to about six seconds and working ninety-six percent of the time. And in a later blind test across seven hundred trials, it recovered the lock six hundred ninety-five times. Ninety-nine point three percent. And the end product wasn't the AI staying in the loop forever, it was a clean, deterministic, inspectable script that runs in production without the agent.
That's the pattern I want you to notice, founders, because it's the actually-useful shape of this stuff. The agent wasn't the product. The agent was the thing that discovered the solution, then packaged it into deterministic code a human can read and trust. It rewrote a rigid linear sequence into a smart decision tree by running the failure over and over until it saw the pattern a human couldn't catch at that speed. That's a real capability, and it's a repeatable playbook: use the expensive, slow, exploratory agent to find the answer, then compile the answer down into something cheap and reliable.
And here's why I trust this writeup more than most vendor blogs. It tells you where Claude fell on its face. Over at Genentech, they were automating a protein assay, and the agent kept hitting errors caused by air bubbles in the liquid. Claude's instinct was to just retry in the same spot with different settings, which, if you know anything about foam, only makes more bubbles. The article says flat out: Claude "did not yet understand the underlying physics of the failure," and a human had to guide it toward gentler handling. In the quantum example, they admit Claude couldn't troubleshoot actual physical hardware problems because its understanding was "programmatic rather than physical," and it kept pausing overnight to wait for human approval on anything it deemed risky. Their line was, and I appreciate this, an overly cautious agent is preferable to one that isn't cautious enough.
So what's the builder takeaway from the whole Model Hardware Standard thing? A few things. One, the interoperability problem is real and it's a giant tax on anybody doing physical work, and standard driver layers that collapse months of integration into hours are genuinely valuable, AI or no AI. Two, notice the partner list, AWS, Raspberry Pi, Universal Robots, and yes, Hugging Face with their LeRobot library. Which brings the whole episode full circle: the same day Nvidia moves to own the model warehouse, Anthropic is laying down the standard for how models touch machines, and Hugging Face is a named partner in both stories. The physical-AI land grab is on, and the plumbing is being poured right now. And three, the honest limits matter for your planning. This is not autonomous science. It's a human setting intent, an agent doing the tedious middle, and a human keeping their hand near the emergency stop. If somebody pitches you fully autonomous anything in the physical world today, keep your hand on your wallet.
Alright, let me get you the rest of the menu, quicker bites now.
Let's shift from labs to the policy fights that hit your supply chain. There's a report, and I want to frame this correctly, this surfaced from reporting that traces back to earlier this summer, so treat it as renewed attention on an ongoing fight rather than something that landed brand new this morning. But it's live and it matters: the Trump administration is weighing sweeping new semiconductor tariffs, and the tech industry is, to put it mildly, alarmed. One industry official called it "the single dumbest way imaginable to pursue American dominance in AI," and said it's "like kneecapping yourself at the starting line." The core tension is simple, and it's the kind of thing that sounds fine in a slogan and falls apart in a spreadsheet. You want more chip factories built in America, great. But domestic chip plants take years to build. Taxing imported chips today, while the buildout desperately needs those chips today, just raises the cost of the exact thing you say you want. As one source put it, "the math literally just does not work." One trade group estimated that approach could cost ninety billion a year in GDP and delay or cancel about a fifth of planned data center projects. If you build hardware, or products with chips in them, which is most of you eventually, watch this closely. The version where they tax the servers and consoles and downstream goods, not just the raw chips, is the nightmare scenario.
Now, on the flip side of the hardware conversation, there was a recent report worth a quick note for the local-AI crowd, and I'll keep it to the ninety seconds I promised you. Apple refreshed the Mac Studio and Mac mini and is very openly aiming them at local AI inference. The new M5 Ultra tops out at five hundred twelve gigabytes of unified memory. The interesting cultural signal here isn't the specs, it's that people have been daisy-chaining these Macs together to run big open models locally, as an alternative to renting cloud GPUs, and Apple has clearly noticed and is now marketing to those developers directly. That's the whole story for our purposes. If cloud token bills are eating your margins, this is a lane that keeps widening. Moving on.
A couple of quick business items. There was an executive-shuffle story that I'll mention only because it captures the moment: Barret Zoph, who co-founded Thinking Machines, then went back to OpenAI, and it later came out he'd been fired from Thinking Machines, has now landed at Google as a VP of research. I'm not going to spin a grand theory out of one person's job hops, but the through-line the reporting notes is real, OpenAI has been bleeding senior people over the last eight months, from its COO to a top data center exec. When a company readying an IPO is losing that much altitude in its leadership ranks, that's a thing to keep half an eye on. Not a crisis. A pattern.
And on the startup side, one that's a nice little window into where AI is quietly shifting value. A fashion company called Atorie raised nine and a half million to sell luxury-quality goods, Italian leather handbags for a few hundred bucks instead of a few thousand, by connecting factories directly to consumers. And here's the AI angle that matters for founders: the co-founder says the real shift is that the best factories now have their own design and product-development capabilities, and AI helps them predict demand so they don't overproduce. That's the actual disruption, not the chatbot on the website. It's AI collapsing the distance between the factory and the customer and cutting out the inventory risk in the middle. He also mentioned they're seeing sales referrals coming in from ChatGPT and Claude, which, take with a grain of salt, but file it under: agentic shopping is starting to show up in somebody's referral logs.
Let me hit two more and then I'll get out of your hair.
On security, this one's a heads-up more than a deep dive. The ATF, the federal Bureau of Alcohol, Tobacco, Firearms and Explosives, declared a "major incident" after a cyberattack, and a ransomware gang called Qilin claimed responsibility. The system that got hit reportedly held information on the targets of ATF investigations. Now the reason I flag this for a builder audience, it's part of a drumbeat. Same week you had over a hundred tech companies, OpenAI, Anthropic, Google, Microsoft, plus the big cyber firms, signing an open letter warning that AI-enabled cyberattacks are about to get "far more widespread and sophisticated." And yes, there's an obvious tension there, the same companies building ever-more-capable models are the ones warning about ever-more-capable attacks, and also selling you the defensive products. That's not hypocrisy exactly, it's just the shape of the business. But the practical signal for you is that the threat surface is genuinely moving, and if your security posture is still "we'll get to it," the people with the most information are telling you to move it up the list.
And the last one, because it's a little bit of fun and a little bit of a lesson. A thousand people across twenty-five countries voluntarily buried their underpants. On purpose. It's a citizen-science project, real study, real journal, where volunteers buried pairs of organic cotton underpants for two months to measure soil health. The idea is the faster the cotton rots, the more alive the soil is. They call it the "underpants index," and it turns out it works about as well as the fancier scientific method it was replacing, the teabag index, but it's easier to see and understand. And the finding was that private gardens had the healthiest, most active soil, lawns the least. I bring it up because there's actually a builder's lesson buried in the dirt with those shorts: the best measurement tool isn't always the most precise one, it's the one people will actually use and understand. A visible, intuitive signal beat a fussy lab instrument. Keep that in mind next time you're designing a dashboard nobody looks at.
So let me tie the bow on it. The big story today is Nvidia reaching up the stack to own Hugging Face, the warehouse of the open-model world, right as the whole industry pivots toward physical AI, with Anthropic laying down the standard for how models drive machines and Hugging Face sitting inside both stories. Underneath all of it, the plumbing question keeps repeating: who owns the neutral ground, and how long does it stay neutral once somebody owns it. That's the thing I'd watch. Not the model of the week. The infrastructure and who's quietly buying it.
That's the menu, kiddos. Watch the Hugging Face deal for whether it closes and whether that neutral ground stays neutral, watch the chip tariff fight for whether they go after downstream goods, and if you're building on open weights, keep your eyes on your foundation, because somebody just made a bid to own it. This has been Barely Possible. I'm Tony DeLuca, thanks for spending a little of your day at the counter with me, and I'll see you tomorrow.