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, it's your boy Tony DeLuca, and we've got a fresh plate of tech and AI morsels cooking today. Pour yourself something, settle in, and let's have at it.
I want to start with a story that isn't the flashiest thing on the menu, but it's the one that's been rattling around in my head, because it's about a company that did everything right and still lost. NASA and a startup called Katalyst Space Technologies announced this week that they're giving up on a mission to rescue the Swift gamma-ray observatory before it falls out of the sky and burns up in the atmosphere.
Let me back up and explain what Swift is, because it matters. Swift is the Neil Gehrels Swift Observatory, a roughly 22-year-old telescope that was originally designed to last two years. Two years. It has outlived its design by a factor of ten, and astronomers still lean on it hard, because it does something special. It detects and locates gamma-ray bursts, which are the most powerful explosions in the known universe, and it points fast enough that other telescopes can swing around and follow up. It's a $500 million piece of hardware doing real science. The problem is Swift has no thrusters. It cannot keep itself up. So it's slowly falling, and the only way to save it was to send something up to grab it and boost it into a higher orbit.
Enter Katalyst. Less than a year ago, NASA gave them a $30 million contract and a brutal deadline: build a first-of-its-kind rescue satellite and launch it by this summer, or Swift's gone. Now here's the thing, and I want you builders to really hear this. A satellite like this normally takes several years to get ready. Katalyst did it in nine months. Nine months. The satellite is named Link, it's about the size of a refrigerator, it's got three robotic arms and three xenon electric thrusters, and it launched July 3rd. The first few weeks in orbit, everything looked good.
Then in late July, Link spun out of control. Two of its three reaction wheels, which are what keep a satellite pointed the right way, just stopped working. The cold gas thrusters used for fine pointing had problems too. That left them trying to fly a refrigerator hurtling around the Earth at nearly five miles a second using only three low-impulse plasma thrusters. And it wasn't enough. So this week they called it.
And here's where I want to sit with you for a second, because the quotes from the people involved are actually worth chewing on. NASA Administrator Jared Isaacman said, quote, "NASA should be willing to move quickly and take smart risks when the potential return is worth it, and that is exactly what we did with this mission. This is not the outcome we were working toward, but it does not change why this mission was worth attempting." And Katalyst's CEO, Ghonhee Lee, said, quote, "We took on this high-risk, high-reward challenge and are proud of the milestones we reached along the way. We have already learned a tremendous amount, and now our job is to turn those lessons into something durable, building a repeatable playbook to inform future rendezvous and proximity operations and satellite servicing."
Now, you can read that as corporate cope. Somebody failed and they're spinning it. And look, I'm a skeptical guy from the Bronx, I've heard my share of spin. But I actually think there's a real lesson buried in here for anyone building anything hard, and it's this: the nine-month timeline was the villain and the hero of this story at the same time. The article says it plainly, the schedule drove engineers to tolerate technical risks they might not otherwise accept. When you compress a multi-year build into nine months, you are trading margin for speed. You're saying: I'll accept a higher chance of blowing up because the alternative is guaranteed loss. Swift was falling regardless. There was no slow, safe version of this mission. It was move fast or watch a $500 million telescope become a fireball.
And that's the trade every founder knows in their bones. Sometimes there is no cautious path. Sometimes the deadline is real, the window is closing, and you ship the thing you can build in the time you have, knowing the reliability isn't where you'd want it. Katalyst is one of several US companies working on satellite servicing, refueling, repairing, upgrading things in orbit. Those robotic arms and rendezvous sensors are exactly the tech that matters for the future of that whole industry. So did they fail? On the stated goal, yeah. Swift is going to reenter later this year and burn up, and astronomers lose a workhorse. But the capability they proved out, getting a novel spacecraft to the pad in record time and getting it close to a target, that doesn't evaporate. That's the durable part. Whether NASA and the market reward that kind of fast, risky, honest failure, or punish it, that's the thing I'll be watching. Because if we only fund the safe stuff, a lot of falling telescopes just burn.
Now, stay with me on this theme of expensive machines and hard economic math, because there's a companion story here, and I want to flag the framing carefully. There's a piece out about Europe canceling planned upgrades to its Ariane 6 rocket. Now, this one's a little older, the underlying decision traces back to reporting from earlier in the summer, June-ish, so I'm not going to tell you this happened this morning. But it resurfaced, and the economics are worth the two minutes.
Ariane 6 has actually been a quiet success. Eight missions, all reached orbit safely, smoother rollout than Blue Origin's New Glenn or United Launch Alliance's Vulcan here in the States, both of which had real growing pains. It gave Europe assured access to space, which mattered a lot after the war in Ukraine ended their use of Russian rockets and they didn't want to be dependent on Elon Musk's SpaceX for their most important satellites. Good story so far.
Here's the but. The European Space Agency shelved the "Block 3" upgrade, which included a lightweight carbon-composite upper stage called ICARUS. And the reason the ministers who write the checks balked is cost. Ariane 6 is expendable, meaning you throw the whole rocket away every launch. Europe sank about 3.6 billion euros into developing it and its launch site. And then, on top of that, the head of ESA has publicly confirmed that even at nine flights a year, European governments still have to subsidize each launch to the tune of 32 to 38 million euros. Amortize the development costs plus the ongoing subsidy, and the article estimates European taxpayers are putting more than 100 million euros into every single launch before the customer pays Arianespace a dime.
Now here's the kicker for you business-minded folks. Amazon signed up for 18 Ariane 6 launches to build out its Amazon Leo satellite constellation. Sounds like a commercial win, right? Except Europe is probably losing money on each of those private flights. You landed the big customer and you're still underwater. That's the trap of an expendable rocket in a world where SpaceX and now Chinese companies are landing reusable ones. And the real tell is this: the ministers may decide the smart move is to stop pouring money into Ariane 6 upgrades and instead fund a next-generation rocket with at least partial reuse. Because subsidizing yesterday's architecture to compete on price is a losing game. There's a decision coming from European ministers later this year on whether to scale Ariane 6 production up to 12, 15, even 20 launches a year, and that decision is going to tell you everything about whether Europe's chasing the current market or building for the next one. Two space stories, same underlying question: when do you keep patching the thing you have, and when do you admit the economics don't work?
Alright, let's come back to Earth and into your actual world, founders, because the next one is the most consequential story on today's menu for anyone building a company. This is the Runlayer and Rippling lawsuit saga, and I'll frame it right, the original filings go back to late last year and earlier this summer, but the resolution just landed and the takeaway is genuinely fresh and genuinely useful.
Here's what happened. On Wednesday night, Runlayer and Rippling both dropped their lawsuits against each other. No settlement, no money changed hands, not even lawyers' fees, according to court documents TechCrunch saw. And Rippling celebrated by immediately releasing its MCP gateway, which is the exact product at the heart of the whole fight. Let me walk you through it, because this is a cautionary tale that every founder should have tattooed somewhere visible.
Runlayer is an early-stage startup, launched out of stealth in November 2025, raised 42 million bucks from VCs including Khosla Ventures' Keith Rabois and Felicis. It's led by Andrew Berman, a third-time founder. Their product is an MCP gateway. And for those of you not deep in the agent world, let me explain what that is in plain English, because it matters. When you've got AI agents inside an enterprise, and one of them asks for something like "give me the top five job candidates and their emails," that data has to come out of the company's recruitment system. An MCP gateway is the thing that securely handles that request. Instead of giving the agent direct keys to all your internal systems, the gateway retrieves the data, and it layers on the important stuff: role-based access control so a manager sees different things than an intern, observability, logs, usage trails. It's essentially the security and permissions checkpoint for AI agents inside a company. That's a real, valuable category.
So here's the tale. Rippling, the payroll and HR giant, tested Runlayer's gateway. Not for a week. For more than a year. The two engineering teams worked closely together. And after all that, according to Runlayer's lawsuit, Rippling never signed on as a customer. Instead, Berman got a text from a Rippling employee saying, hey, my employer is building its own MCP gateway and planning to release it as a product. And this employee, per the suit, described Rippling's product as a clone of Runlayer's. Runlayer sued, claiming Rippling violated the contracts covering those product tests. Rippling countersued, alleging Runlayer was violating some of its patents, which Runlayer read as a pressure move, jack up the legal bills, force us to fold. Runlayer dropped its suit after three weeks in discovery. Rippling dropped its own. And then Rippling shipped the product.
Now, I'm not here to litigate who's right. The suits are dropped, nobody admitted anything. But TechCrunch's Julie Bort nailed the deeper lesson, and I want to read the core of it, quote: "In the age of AI, when building new software has become almost trivial, you never know who your next competitor will be. It might even be a prospective customer."
Chew on that. In the age of AI, when building new software has become almost trivial. That's the whole game changing right there. Ten years ago, if a big company evaluated your product for a year, the worst case was they said no and you moved on. The idea that they'd just go build it themselves was a long, expensive proposition that most enterprises had no appetite for. Now? Building the software is the easy part. So that "technical shoot-out" that big enterprises love to make startups go through, where you spend months proving your product works in their environment, that whole ritual has become dangerous for the startup. You're not just spending your runway on a sales cycle. You might be handing your prospective customer a free, year-long education in exactly how to build your product without you.
And look at where Rippling is now. Their bread and butter was payroll and benefits. In the span of weeks they've entered the AI gateway market with a tool that routes to different models and dashboards token spend by employee, putting them up against Stripe, Ramp, and Databricks. And with this MCP gateway tying AI access to employee roles, they're now in the AI security business against Runlayer, Docker, and Amazon Bedrock. A payroll company, suddenly competing on four fronts, because in this environment the software is cheap to build and the distribution and the customer relationships are the moat.
So what's the actionable piece for you? Two things. One, rethink the long enterprise pilot. If a much bigger company wants to embed with your engineering team for a year to "evaluate," ask yourself what happens if they walk. Structure the contract, structure the access, structure what they can see and take. Runlayer's whole legal claim rested on contractual agreements covering the tests, which tells you those agreements are worth taking seriously up front, not after the text message arrives. And two, Runlayer's actual survival strategy is instructive: their pitch now is a broader bundle of agent security services, agent creation, spotting shadow AI agents running inside a company that IT doesn't even know about. In other words, when the single product can be cloned, you widen the surface. You become a platform, not a feature. Because the feature is the thing that gets copied. The bundle, the relationships, the harder-to-replicate stuff, that's what survives the age of trivial software.
Now let me connect that to something that's been a running thread on this show. We talked recently about the AI industry's public-perception problems, the blame debate, the Stripe-OpenRouter deal, and generally this feeling that AI is being forced into people's lives whether they want it or not. And there's a fun little story today that's the comedic version of that exact anxiety, and I promise there's real substance under the joke.
Liquid Death, the water-in-a-tallboy brand that runs on provocation, teamed up with former NFL star Jason Kelce for a commercial where the pitch is, and I quote from the man himself, "AI data centers waste millions of gallons of water. That's why Liquid Death and Garage Beer have teamed up. We want your pee to cool these data centers." And then a field full of people sing, "Let's pee on computers together to save humanity." It's dumb. It's funny. It went viral.
But here's why it made my list. The joke accidentally lands on something real. TechCrunch actually called up water experts, and the funny part is data centers really do cool themselves with reclaimed wastewater, which, if you follow the plumbing all the way back, yes, contains human urine. Now you can't literally pour pee into a cooling tower, one expert pointed out it's full of salts and urea and bacteria that'd gunk up the whole system and require constant cleaning. But treated, recycled water is a huge and growing thing. One water reuse specialist said, quote, "Every day right now, I'm working on recycled water for data centers." Loudoun County, Virginia, which has more than 250 data centers, uses about 200 million gallons of recycled water a day, and that's still only 43 percent of the water those centers drink. The other 57 percent is coming from the drinking water supply.
And here's the part builders should actually file away, because it's a real strategy, not a punchline. One expert argued that data centers can become anchors of water infrastructure for communities. Meta, for example, is putting at least 270 million dollars into wastewater infrastructure projects near its data centers. There's a policy push for a 30 percent tax credit to help scale recycled water systems. So the smart move for a company building compute in someone's backyard isn't to fight the backlash, it's to show up as the entity that finally funds the water treatment plant the town couldn't afford. Because the backlash is real, a recent Gallup poll had about seven in ten Americans opposing data centers in their communities. When Jason Kelce is selling beer off of hating your industry, that industry has a public relations problem it cannot ignore. The commercial's a symptom. The water infrastructure math is the cure.
Alright, let me shift from cooling towers to the security desk, because there were two stories this week about the strange, escalating relationship between AI companies and the security researchers who are supposed to be on their side.
First one. Somebody ran a hacking campaign targeting cybersecurity professionals around the time of the Black Hat and Def Con conferences earlier this month. And I love the opening line of the TechCrunch piece, because it's just true: if you're a malicious hacker, cybersecurity pros are the worst people on Earth to try to hack, because there's a real good chance they catch you. Which is exactly what happened. A person pretending to work for a leading crypto news site approached conference attendees on X, both public replies and DMs, and tried to lure them with a fake crypto conference. They'd share a legitimate Google Doc, an actual Google Doc, that looked like a planning document, with a sidebar rigged to look like the document was encrypted. The trick was to get the target to enter a fake "decryption key," which was step one in a chain that ended with malware, an infostealer for Macs, a repurposed remote desktop tool for Windows, and a fake installer for the Ledger crypto wallet.
The clever bit, and the thing you should note if you build with these tools, is how they made the fake encryption sidebar look real. They used Google App Script, a totally legitimate platform that lets developers customize the interface of Google Docs with menus and sidebars. So the attack rode on top of a real, trusted Google product and a real Google feature. That's the pattern to watch. The security firm Huntress caught it because one of their own researchers played along to see what the guy was after. But the lesson is that "it's a real Google Doc from a real Google feature" is no longer a signal of safety. The trust layer itself is the attack surface.
Now here's the connected story, and it's the more consequential one for the AI world. Several security researchers say OpenAI abruptly revoked their access to a limited-access program called Trusted Access for Cyber, TAC. OpenAI confirmed it was caused by an error. Let me explain what TAC even is, because it's a genuinely interesting piece of how the AI labs are trying to walk a tightrope. TAC is a program where OpenAI gives vetted, ID-verified security researchers access to its most advanced models with fewer of the usual cybersecurity guardrails. Anthropic has an equivalent, called the Cyber Verification Program. The whole idea is: give the good guys, the defenders, the powerful models so they can find bugs and get them patched fast, while keeping those same unrestricted capabilities away from the criminals.
So what happened. On Wednesday, a bunch of researchers opened up ChatGPT's Cyber page and got messages saying their identity couldn't be verified or their account was "ineligible at this time." OpenAI told them it was a technical issue affecting a limited number of users, quote, "This was an issue on our end, and not the user experience we want to deliver," and asked them to reapply and re-verify. Here's the detail that jumped out at me: every researcher TechCrunch spoke to lives outside the US and Europe. Which suggests the revocations may have been regional. Now, OpenAI says it was an error, and maybe it was purely a bug. But if you're a defender in, say, Southeast Asia or Latin America or Africa, and the tier that gives you real capability suddenly locks you out while, presumably, your peers in San Francisco keep humming along, that's a trust problem. The newest tier, called Daybreak Blue, gives access to frontier general-purpose models with safeguards tuned for defensive work, and there's an even higher tier, Daybreak Red, for offensive vulnerability research. These are powerful tools. And the through-line with both of today's security stories is the same: as these capabilities get more powerful, who gets access, and who gets accidentally locked out, becomes a real geopolitical and trust question, not just a support ticket. Defenders around the world are increasingly complaining that the guardrails are getting in the way of legitimate work. And if the labs can't administer these programs cleanly, without regional glitches that feel like exclusion, they undermine the exact defenders they built the program to empower.
Let me take a breath and give you a couple of quicker hits, because the menu's got some smaller plates worth tasting.
There's a robotics reshoring story I want to flag, and I'll frame it right, the underlying policy trail goes back several months, so this is a resurfaced piece pulling threads together, not brand-new news. A company called RoboStore had become the main North American distributor for Chinese humanoid robots and robot dogs from Unitree, selling to Harvard, MIT, Amazon, Nvidia, over 150 universities, the CEO says they moved more than 1,500 robots to customers including Cisco and OpenAI. Then in July, the FCC banned imports of new foreign-made robot models, Unitree's included, on national security grounds. So RoboStore shut down its distribution business and spun up a new company, Robo Inc., to manufacture its own robots at a 66,000-square-foot facility on Long Island, aiming for early 2027. And here's the founder's take that I actually respect, Teddy Haggerty said, quote, "Companies like Standard Bots and Universal Robots have a robotic arm for specific use cases. And I will tell anyone that that product is better suited for most businesses, for what they want to achieve, than a humanoid robot." That's a distributor who sold humanoids for years telling you, honestly, that for most real jobs a boring specialized arm beats a general-purpose humanoid. When the guy who profited from the hype tells you the hype is oversold, listen. Also worth noting, they were bootstrapped for years and are only now raising money on the back of the reshoring wave. Policy created a market, and a distributor turned itself into a manufacturer to catch it.
Quicker still. Grok, xAI's chatbot, had a glitch where it was spitting pure gibberish at some users. One person asked it to generate a PDF and got, and I'm reading this straight, "match it without and your they and two for planets can practical and often cheese," continuing for paragraphs. Refreshing the session usually fixed it. It seems to have hit only a small subset of users on Grok Lite, on the website, not the X account. The Grok account itself called it, quote, "a rare temporary generation glitch." Minor in itself. But I mention it because it sits next to a bigger fact from that same reporting: xAI has had heavy staff turnover, reportedly losing most of its founding team and at least 50 researchers and engineers, per a report from earlier this year. When word salad starts leaking out of a frontier model and the team that built it has thinned out, you at least raise an eyebrow.
And a couple of tiny ones. Google is throwing publishers a bone as AI search kills their traffic, this rolled out earlier in the year and expanded now, a "Preferred Sources" button readers can click to see a site more often across Search, Discover, and News. Google says people are twice as likely to click through to a preferred source. It's a small lever against a big problem Google itself created. And in the file-under-billionaire-life-cycle category: Mark Zuckerberg bought a 19th-century Irish castle, Strancally Castle, 440 acres, somewhere between 23 and 35 million dollars, a couple hours from Meta's Dublin headquarters. Eleven bedrooms, four tower suites, a library. I've got nothing analytical to add. Sometimes a guy just buys a castle.
Now let me close with the story that I think quietly says the most about where our whole industry is emotionally right now. OpenAI launched a new blog this week called AI Futures. The description is short, quote, it's "a new OpenAI blog exploring how transformative AI could reshape power, governance, the economy, and individual freedom."
That's it. That's the whole thing I've got on it, it's a launch announcement, so I'm not going to pretend I've read a manifesto. But think about what it signals, especially stacked against the other OpenAI moves from this week. They rolled out ChatGPT for Teens with parental controls and healthy-use features. They partnered with a group called CodeAI on AI literacy for students. They launched an initiative on democratic oversight of AI in national security. And now a whole blog dedicated to power, governance, and individual freedom.
Here's my read, from the kitchen table. A company doesn't stand up a governance-and-freedom publishing arm when it feels beloved. It does it when it feels the temperature in the room changing. And we've watched that temperature all week. The Jason Kelce pee commercial. Seven in ten Americans not wanting a data center near them. Security researchers grumbling about guardrails. The revenue-scrutiny pieces circling the labs ahead of their IPOs. OpenAI is a company that reads the room extremely well, and the room is anxious. So the question I'd leave you builders with isn't whether AI Futures will have good essays in it. It's whether the industry can actually deliver the thing that changes minds. Because as one lab CEO put it recently, no glitzy marketing campaign wins back trust. The thing that works is actually delivering results. A blog about individual freedom is a promise. The Swift telescope falling out of the sky because the rescue didn't work, the Runlayer founder getting a text that his customer is now his competitor, the town that still can't afford its own water plant, those are the results, good and bad, that people actually live with. The words are cheap now. Building the software is trivial now. Earning the trust is the only hard part left.
That's the plate for today, kiddos. Chew on it, and don't let anybody pee in your cooling tower. This has been Barely Possible, I'm Tony DeLuca, and I'll catch you next time.