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 tech news cooling on the windowsill today, so grab a chair and let's dig in. This is Barely Possible, the show where I read the technical stuff so you don't have to pretend you did.
I want to start today somewhere unusual, because there's a story sitting in the pile that most people would skip right past, and I think it might be the most important thing on the menu for anybody who's building anything on top of a data center. And these days, that's basically all of you.
Here's what happened. A single power line went down outside Washington, DC this week. Now, on a normal day, the grid shrugs that off. A few seconds, maybe, and everything's back to balanced. That's what the grid is supposed to do. This time, though, it took more than ten minutes. And the reason it took ten minutes is the story.
When that line went down, more than three gigawatts of data centers stopped drawing power at nearly the same instant. Ting Labs, a startup that runs sensors out of people's electrical sockets, caught the whole thing on their network. Voltage spiked across the PJM grid, from Northern Virginia all the way to Chicago. No blackout, thank God, but lights flickered across the region. And the guy they quote, Ricardo de Azevedo, the CTO at a company called ON.Energy, he calls it "the canary in the coal mine." His words. And he says these events involving big loads like data centers are "happening more and more."
So let me walk you through why this matters, because it's actually a beautiful little example of how a system fails, and it's not the way you'd think.
Northern Virginia, if you didn't know, has the highest concentration of data centers on the planet. It's in PJM's territory, and PJM is the biggest grid operator in the country. Sixty-seven million customers, New Jersey to Illinois. When that power line failed, the data centers in Northern Virginia sensed a fluctuation in the voltage, and they did exactly what they were designed to do. They flipped to backup power to protect themselves. Smart, right? Each individual data center made a sensible, self-protective decision.
But here's the problem. When a data center flips to backup, it stops pulling power from the grid entirely. Its load just vanishes. And when hundreds of them do it within a few seconds of each other, all making that same split-second decision independently, you get about 3.1 gigawatts of demand disappearing in roughly thirty seconds. Then a little later, more of them dropped off. At the peak, PJM's grid had an extra 3.49 gigawatts of electricity sloshing around with nowhere to go. It took eleven more minutes to stabilize.
Now think about what that is. It's not one big failure. It's a whole bunch of individually rational decisions that add up to a collective disaster. Every data center said "I'll protect myself," and in protecting themselves, they nearly took down the neighborhood. The disconnected facilities were only about three percent of total demand at that moment. Three percent. Doesn't sound like much. But the grid has to sit in near-perfect balance, supply matching demand, and when three percent walks out the door in thirty seconds, the voltage surges and your lights start flickering in Chicago.
There's a guy quoted here, Ali Zain Banatwala, a market models specialist at the Independent Electricity System Operator, and he lays out the fix in one sentence. "We need to figure a way for these loads that are located next to each other to sequentially either disconnect or reconnect." Sequentially. In other words, don't let them all bail at once. Stagger it. Take turns. It's traffic management for gigawatts.
And there's a startup, ON.Energy, working the other angle. Instead of staggering the exits, they hide the whole data center behind a bank of batteries. The grid only "sees" one calm, well-behaved load. The batteries absorb the fluctuations, they charge up when there's extra power, they dispatch power to the servers when the flow dips, and they can react within milliseconds. They're installing three gigawatts of this stuff across four campuses right now. And grid managers are waking up. ERCOT, the Texas operator, is going to start requiring big loads like data centers to "ride through" disruptions instead of turning their backs and bolting.
Here's the number that should stick with you. This week's mass disconnect was twice as big as a similar event in 2024, when sixty data centers dropped 1.5 gigawatts at once. Back then, data centers were about six percent of PJM's load. By 2040, they're projected to be twenty-four percent. Almost a quarter of the grid.
So why do I lead the show with this? Because if you're a founder building AI products, you spend all day thinking about tokens and models and latency, and you treat electricity like it's water from the tap. It just shows up. This story is a reminder that the thing your entire business floats on is a physical system operating on a razor's edge of balance, and the AI boom is putting weight on it faster than the balance can adjust. The failure mode isn't "we run out of power." The failure mode is "everybody protects themselves at the same time and the whole thing wobbles." That's a coordination problem, not a capacity problem, and coordination problems don't get solved by building more. They get solved by somebody making the loads behave politely with each other. Watch which vendors and which grid operators actually mandate that, because that's going to become a real line item in where you can and can't put your compute.
Alright, let me pull us out of the substation and into the model shop, because there was some frontier-model news to close out the week, and I want to handle it carefully, because part of it we already chewed on.
We covered the Claude Opus 5 launch yesterday, so I'm not going to re-plow that field. Quick reminder for continuity, since it connects to today: Opus 5 landed Friday, and the whole pitch, as Ars Technica's Samuel Axon put it, is that it's not a capability leap, it's a cost story. Five bucks per million input tokens, twenty-five per million output, and it gets you something just shy of the flagship Fable model at roughly half the price. What I want to add today is the frame around it, because there's a fresh detail in that piece I think matters more than the model itself.
Axon points out that the whole conversation among software developers and engineering managers right now isn't about "which model is smartest." It's about cost. And there's a Chinese open-weight model, Kimi K3, that comes in at fifteen dollars per million output tokens at similar performance. So the price pressure is coming from every direction at once. And here's the part that should perk up a founder's ears: companies like Cursor and Meta are building what they call model routers. Systems that automatically pick the right-sized model for each individual prompt. You don't burn a Ferrari-grade model on a task a Corolla could handle. You route the easy stuff to something cheap, save the expensive model for when you actually need it.
Now hold that router idea in your head, because it's about to show up somewhere else.
Let me tell you about a company called Prentis. And I'll note up front, this is a report from a few days back, from July 24th, so I'm not telling you this happened this morning. Prentis is a new AI lab, launched back in April, co-founded by a serial entrepreneur named Ritankar Das, with Reid Hoffman and Mark Pincus attached as co-founders. TechCrunch's Marina Temkin reports they're in talks to raise a hundred million dollars at a one-billion-dollar valuation, according to two people familiar with the discussions.
What are they building? Computer-use models. Agents that learn how office workers navigate routine workflows across documents and systems, so the software can eventually do those tasks itself. Handling insurance claims. Automating customs duty refund exceptions, so nobody has to hunt down paperwork. The boring, expensive, human-hours stuff.
And here's where it ties back to that router conversation. Prentis's whole pitch is the small-cheap-model bet. By their own account, their model, Hive-32B, beats OpenAI's GPT-5.4 and Anthropic's Claude Opus 4.6 on two computer-use benchmarks. And TechCrunch is careful to say they haven't independently verified those numbers, so keep the salt handy. But the interesting claim isn't the benchmark. It's the economics. They say roughly ten times lower cost per task than the frontier APIs, because they're running a much smaller, cheaper model. Same instinct as the router. Don't use the big expensive brain when a smaller specialized one gets the job done.
Now here's the part that made me raise an eyebrow. Prentis has already signed contracts worth up to fifty million dollars with customers, and their pitch deck predicts a seventy-five-million-dollar annualized run rate by the third quarter. But read the fine print, because TechCrunch did. Those figures reflect "estimated annualized value based on a contracted fee equal to twenty percent of savings realized." Not recognized revenue. Performance-dependent. Subject to final execution.
Founders, chew on that, because it's a business model choice you're going to see more of. Prentis isn't charging a flat SaaS fee. They're taking a cut of the money they save you. Twenty percent of realized savings. On paper, that's beautiful. Aligns everybody. You only pay when it works. But it also means their "run rate" is really a projection of savings that haven't happened yet. It's a promise dressed up as revenue. And the bet underneath the whole thing, the sources say, is that automating everyday office tasks will soon outpace coding as AI's biggest use case. That's a big call. Everybody's been obsessed with coding agents. Prentis is saying the real money is in insurance claims and customs paperwork. The dull stuff nobody wants to do.
It's a crowded room, though. Anthropic, OpenAI, and Mira Murati's Thinking Machines Lab are all chasing computer-use agents. Anthropic even bought a Seattle startup called Vercept earlier this year, folded in the founders, shut down the product. So Prentis is a small boat in a channel full of aircraft carriers. But the angle, cheap specialized model plus pay-for-savings pricing, that's the thing to file away. If you're building agents, the pricing model may end up being your real moat, not the model weights.
Now let me swing to the other side of this same coin, because while the labs and startups are racing to shove AI into every workflow, there's a counter-current out in the world that I find genuinely fascinating, and it's the least Silicon Valley story in today's pile.
Librarians. I'm serious. TechCrunch's Amanda Silberling has a piece about librarians hosting "Avoiding AI" workshops for people who are fed up with Big Tech. This is a current story, out this week. And it's not a couple of cranks. It's going viral.
The setup is almost too perfect. There's a librarian in South Philadelphia named Charlie Bailey, standing on a children's classroom rug, the one with M-is-for-moon and Z-is-for-zebra, except the twenty-odd adults in the room aren't there for the alphabet. They're there to learn how to turn off Apple Intelligence and Gemini on their phones. The workshop's called "Avoiding AI." Bailey walks through the popular platforms and devices and shows people, step by step on the projector, how to shut off the features they never asked for.
His words: "I was inspired by the feeling of people's frustration with AI tools being kind of forced onto them, and feeling like AI tools we didn't ask for are suddenly everywhere in our lives." He frames it not as anti-technology, but as digital literacy. Helping people reclaim autonomy over whether they use these tools, "especially when it can be so difficult not to use them, and when the design seems to force adoption."
Bailey got the idea from another librarian, Hannah Cyrus, up in Maine, at the Bangor Public Library. She wrote a journal article about her own version of the workshop, and dozens of librarians from around the world emailed her asking for the slides. Her line: "Nobody has ever been emailing me like, 'Can you give me your Intro to Computers slides?'" She was getting patrons walking in asking, and I love this, "Why is it trying to summarize my one-sentence email that I can easily read?"
Her first workshop capped registration at thirty people, opened a waitlist, went to Zoom. About seventy people showed up. Bailey's library posted about his event on Instagram and got over two thousand likes when their usual post gets a few dozen. He had to schedule a second session.
Now, why does a founder need to hear about a library workshop? Because this is the demand signal you don't get from your analytics dashboard. There's an attendee named Gabrielle in the piece who says, "I keep getting AI shoved down my throat at work, and every time I see it, I think about the environment." But then she says, and this is the crucial part, "I'm not against AI in terms of medical breakthroughs." The librarian, Cyrus, points out she uses optical character recognition all the time to scan old documents.
So these aren't Luddites smashing the looms. They're drawing a line. They're saying, the AI that does something I actually want, fine. The AI that writes my emails I didn't ask it to write, that summarizes my one sentence, that installs itself, that I can't turn off, that one's the straw breaking the camel's back. Cyrus calls it exactly that, "the straw that's breaking the camel's back," this forced adoption on people's own devices.
If you're building product, that's a gift. It tells you the resentment isn't about capability. It's about consent. The stuff that makes people show up to a library to learn how to escape your feature is the stuff you jammed in without asking. The path here isn't to hide the off switch better. It's to build the thing people would opt into if you gave them the choice. And note who's leading the backlash, the trusted local institution, the librarian, the neighborhood expert. That's the same energy that makes something spread, and right now it's spreading against forced AI.
And it dovetails with something I'll just flag briefly. Ars Technica ran a piece this week, current story, about the embarrassing flip side of forced AI, a Canadian legislator in New Brunswick named Bill Oliver who read an apparent LLM prompt instruction out loud during a floor speech. He was reading a prepared bit about advocacy offices, and then he just kept going and said out loud, quote, "here's a more natural, flowing version of that section that reads like a legislative speech rather than a series of short points." That's the model offering him an alternative style, and he read the offer instead of the answer. Didn't notice in real time. Video spread on Reddit and Threads, and now the CBC and the Toronto Star are on it.
Look, it's funny. But it's the same theme as the librarians from the opposite direction. The librarians are people who don't want AI touching their stuff, and here's a guy who leaned on it so completely he didn't even read what he was saying into the public record. The public is getting a very fast education in what over-delegation looks like, and a Duke study last year found workers already hide their AI use because colleagues see it as lazy. The trust picture is getting complicated, and if your product's whole pitch is "let the machine do it for you," you're wading into water that's getting choppier by the week.
Now let me shift gears entirely, from model behavior to the machines that talk to the rest of the solar system, because there's a hardware fragility story here that scared me a little.
A fast-moving wildfire, one of many burning across Spain right now, bore down on NASA's Madrid Deep Space Communications Complex on Friday. This is Ars Technica's Stephen Clark reporting, and it's a current story. They had to evacuate and temporarily suspend operations. Photos showed flames and smoke rising over the antenna array, anchored by a two-hundred-thirty-foot radio dish. Spanish authorities ordered more than nineteen thousand people out of towns in the mountains west of Madrid. Two thousand personnel, ten aircraft fighting the fires.
Now here's why this is more than a wildfire story. That Madrid station is one of three sites in NASA's Deep Space Network. Madrid, California, and Australia. Three stations spaced around the globe so that as the Earth rotates, at least one of them can always see whatever probe you're trying to talk to. The network supports more than forty missions. Artemis, the James Webb telescope, the two Voyager spacecraft way out past the edge of everything. NASA says they seamlessly shifted operations over to the Goldstone complex in California, and good for them.
But here's the part that made me sit up. Clark points out that without Madrid, the Deep Space Network is down to a single operational seventy-meter antenna, the big one, in Australia. Because the big seventy-meter dish in California has been offline since last year. An accident "over-rotated" the structure, damaged cables and water lines, flooded the base with two hundred thousand gallons of water containing glycol, an environmental hazard. The cleanup runs four to four-point-six million dollars, and the thing's expected to stay dark into 2028.
So think about that. The entire deep-space communications capability of the United States, the thing that lets us hear Voyager whispering from interstellar space, is running on essentially one big functional dish while a wildfire licks at another one and the third sits broken in the California desert. That is a thin, thin margin. The one bit of luck is that the next big Artemis Moon mission is still a couple years off, and those missions are what really strain the network with all the human-spaceflight telemetry. So there's slack in the schedule to absorb this.
Why do I put this on a builder's show? Because it's the same lesson as the power grid, just in orbit. We build these magnificent capabilities and then we run them with almost no redundancy, on the assumption that nothing goes wrong at two places at once. The grid assumed data centers wouldn't all bail simultaneously. The Deep Space Network assumed you wouldn't lose California and Madrid in the same window. Resilience is expensive and boring, so it's the first thing that gets deferred, right up until the wildfire shows up. If you're running critical infrastructure of any kind, and increasingly software is critical infrastructure, the question isn't "what happens if this fails." It's "what happens if this fails while the backup is already down." Ask it before the smoke shows up.
Alright, let me stay in the world of things going wrong, but bring it back to the ground and into some hard business news, because there's a cluster of stories today about money and lawyers and who gets squeezed.
First, the Boring Company. This is a current report. Elon Musk's tunneling outfit is in talks to raise four billion dollars at a twenty-billion-dollar valuation, according to the Wall Street Journal. Deal's not closed, terms could change. But twenty billion would be a massive jump from their 5.7 billion valuation back in 2022. They've built the tunnel network under Las Vegas where Teslas shuttle people between stations, and they've pitched or announced projects in Nashville, Dubai, Baltimore, Chicago, LA.
Now I'm skeptical of hype by trade, so let me give you the other half of that story, the part that doesn't make the headline. The reporting notes tunnel workers have suffered serious injuries, and Nevada regulators said last year that the Boring Company violated environmental regulations nearly eight hundred times. Eight hundred. So you've got a company quadrupling its paper valuation while sitting on a stack of environmental violations and worker-safety problems. That's the tension. The valuation is a bet on the future vision. The violations are the present reality. When you see a number like twenty billion, always ask what's in the footnotes.
And here's a small thread worth pulling. The Boring Company spun out of SpaceX back in 2018. And SpaceX, which we've talked about, recently did the largest IPO ever, but the stock's taken a real dip since. So the whole Musk constellation of companies is doing this valuation dance while the flagship's public stock is under pressure. Interconnected in ways that aren't always obvious from the outside.
Second bit of hardball, and this one's for anybody who's ever signed an employment contract or made someone else sign one. Warner Bros. Discovery filed a lawsuit this week, current story, accusing Amazon of illegally poaching executives. As reported by Deadline. The accusation is interference with contractual relations, breach of contract, unfair competition. Warner Bros. says Amazon has been, quote, "hurriedly seeking to pirate away a number of contracted employees," including an HBO Max marketing executive named Pia Barlow who jumped to Amazon MGM Studios. Warner says her contract wasn't set to expire until October 2027.
Here's the juicy line from the complaint. Warner says Amazon induced employees to breach their agreements, quote, "backed up with the ready assurance that Amazon will defend and indemnify them should they be held to account." In plain English, Warner's alleging Amazon told these people, come on over, break your contract, and if you get sued, we'll pay for the lawyers. And Deadline notes this is going to reopen a real debate about whether these term employment agreements are even enforceable under California law in the first place.
Now why does a builder care about Hollywood executive musical chairs? Because this is the enforceability question that sits underneath your entire talent strategy. If you're a founder in California relying on a term contract to keep a key hire from walking to a deeper-pocketed rival, this case is going to tell you how much that paper is actually worth. California famously hates non-competes. Term agreements are a different animal, but if a court decides a company can just indemnify its way past them, then those contracts are worth about as much as the ink. Watch it.
Let me do one more from the money-and-lawyers file, and this one's got a nice little irony in it. Kalshi, the prediction market, sent Netflix a cease-and-desist letter demanding they pull the trailer for an upcoming documentary. And I'll be careful with the timing here, the bet at the center of this is from May 2025, so this isn't a fresh event so much as a fresh fight over an old receipt.
The documentary's called "Instadocs: The Prediction Games," about the rise of prediction markets, and it features interviews with both the Polymarket CEO and Kalshi's CEO Tarek Mansour. But the trailer focuses on a party in Las Vegas where guys who've made, quote, "millions of dollars on prediction markets, probably eight figures," gathered to watch the World Cup final. One guy says "I like betting on Kalshi," another flashes an apparent five-thousand-dollar bet on his phone.
Here's the rub. Kalshi is currently banned from operating in Nevada by a court order. So Kalshi's saying, hey, that bet on the phone is actually a screenshot from May 2025, before the ban, and the trailer misleads people into thinking this guy was legally trading in Nevada this July. Netflix doesn't dispute the screenshot is from 2025, but says none of the footage was fabricated, the party was real, and what any individual bet, and I quote, "are between the individual and the app in which they placed the trades."
The irony I can't get past: a prediction market, a company whose entire product is people wagering on what's true about the future, is now in a fight over what's true about a receipt. And the reputational risk for Kalshi is real, because a documentary showing off eight-figure winners is exactly the kind of thing that makes regulators in states like Nevada dig in harder. This is a company managing its regulatory image in real time, and Netflix is holding the camera. If you're building anything in a gray regulatory zone, and crypto and prediction markets live there, take note: the story other people tell about your product can become the regulatory problem, whether or not your lawyers win the takedown fight.
Alright, let me close out today with the story I actually think is the most quietly interesting on the whole menu, because it's about AI doing something good and specific and unglamorous, which is my favorite kind of AI story.
Ars Technica's John Timmer wrote up a team that used AlphaFold, Google's protein-folding AI, to redesign gene-editing proteins to make them safer. This is a current story, and it's in Nature. Now I'm going to keep this tight, I'm not going to drag you through the biochemistry, but the shape of it is worth understanding because it's a genuinely different use of AI than everything else we talked about today.
Here's the problem they were solving. Gene editing, CRISPR, the whole business, it works by targeting a specific DNA sequence and editing it. The trouble is what they call off-target effects. The editing machinery occasionally grabs the wrong sequence and edits something you didn't mean to touch. It's a low-probability event, but when your therapy has to edit millions of cells, low-probability becomes inevitable. That's a safety problem that's dogged the whole field.
The central protein, Cas9, is supposed to only stick to a perfect match. But it turns out Cas9 can tolerate a few mismatched bases and stick anyway, and it's been hard to predict in advance which guide sequences are going to cause those mistakes. So this team, based at institutions in China, fed AlphaFold the DNA, the guide RNA, and the Cas9 protein, and compared how the protein's structure shifted between on-target and off-target sites. They found that in over ninety-five percent of the off-target cases, the specific amino acids touching the RNA changed. AlphaFold could actually pinpoint which parts of the protein were bending around to accommodate the wrong sequence.
They built a tool on top of this they call ContactSeek, found the key spots, made twenty-three different amino acid swaps at ten positions, and here's the payoff: they found a variant that stayed just as effective at the right target while its off-target activity dropped from twenty-eight percent down to five percent. Twenty-eight to five. And it held up across different guide RNAs and even a different Cas protein.
Now I want to draw the line under why this is the one that matters, even though it's got no valuation and no lawsuit and no viral library workshop. Every other AI story today is AI as a product you sell or a feature you shove at someone. This is AI as a scientific instrument. They didn't ask AlphaFold to write anything or automate anyone's job. They used it to see a structural pattern that humans couldn't see directly, and then human researchers used that insight to make something safer. The AI found the where. The people decided the what. And the result is a general method, the authors say, for producing gene-editing systems tailored to prevent specific known off-target events, with applications potentially far beyond gene editing.
That's the version of this technology I don't hear the librarians in Philadelphia complaining about. Nobody's showing up to a workshop to learn how to turn off the protein-folding software that makes their cancer therapy safer. Gabrielle, the woman at the workshop, said it herself, she's not against AI for medical breakthroughs. This is the medical breakthrough. And the difference between the AI people resent and the AI people welcome isn't the intelligence of the model. It's whether it's solving a problem a human actually has, or just inserting itself into a workflow a human was managing fine.
That's the through-line I'd leave you with today, from the flickering lights in Chicago to the librarian's rug in Philadelphia to a protein bending the wrong way. The technology keeps getting more capable, but the value and the trust both come down to the same boring question: did anybody actually ask for this, and does it behave when the pressure's on. Build for that, and you're building something people opt into instead of something they show up to a library to escape.
That's the plate for today. I'm Tony DeLuca, this has been Barely Possible, and I'll be right here tomorrow with whatever fresh trouble the industry cooks up overnight. Take care of each other out there.