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 and AI to work through today, so grab your coffee, pull up a chair, and let's have at it.
I want to start with a story that isn't about a chatbot, isn't about a funding round, and isn't about a new benchmark. It's about a rocket factory nine hundred kilometers inside Russia, and why a bunch of us should be paying attention even if we build software for a living.
Here's what happened. Ukrainian President Volodymyr Zelenskyy confirmed on Saturday that Ukraine struck the Progress Rocket Space Center in Russia's Samara region using ground-launched cruise missiles. Zelenskyy said, and I'm quoting him here, "In Russia's Samara region, one of the key enterprises within Roscosmos—the Progress Center, which was involved, among other things, in electronics production—was hit." And then he added, "Flamingo missiles were used. A good achievement." That's from Eric Berger's reporting.
Now let me tell you why the Progress Center matters, because the name doesn't ring a bell for most people. This is Russia's most important rocket factory. It builds the Soyuz-2, the workhorse vehicle. Of more than a hundred and twenty Russian rocket launches since 2020, over eighty-five percent involved the Soyuz-2 family. It's the principal rocket Russia uses to put humans on the International Space Station. The facility also builds military satellites. And here's the kicker: even before this attack, Russian space officials were saying demand for Soyuz rockets was already running thirty to fifty percent higher than what they could produce for the 2027 to 2030 window. So they were short before somebody put a cruise missile into the building.
The missile they used is the FP-5, nicknamed "Flamingo," built by a Ukraine-based defense group called Fire Point. Range of several thousand kilometers. Open-source reports suggest a missile hit a building housing a clean room for avionics — which lines up with that "electronics production" line from Zelenskyy.
So why am I leading a tech and AI show with a missile strike? Because this is the part of the world that people who build products keep waving off as "geopolitics, not my problem." And I want you to notice the pattern. A country develops a homegrown, long-range precision weapon, and suddenly it can reach into the industrial heart of an adversary and take out a bottleneck that was already stretched to the limit. That's supply-chain warfare. And whether you're talking rockets or GPUs or lithium or natural gas, the story of the next decade is that the physical inputs to the things we build are getting more contested, not less. Keep that thought parked. We're coming back to it, because a couple of today's stories are the exact same idea wearing different clothes.
Now let me shift from missiles to money, because there's a big one moving that affects anybody who touches payments.
There's reporting — and I want to be careful here, this first surfaced back in July, so this is renewed attention on an older thread — that talks to sell PayPal to Stripe and the private equity giant Advent are heating up again. The Wall Street Journal reported originally that Stripe and Advent offered to buy PayPal for sixty dollars and fifty cents a share, a deal that would have valued PayPal around fifty-three billion dollars. PayPal balked at first. But according to the Journal's newer reporting, citing unnamed sources, the negotiations never actually stopped, and a deal could come together in the coming weeks. PayPal declined to comment. A Stripe spokesperson said they don't comment on rumors or speculation.
Let me put my skeptic hat on for a second, because "talks are heating up" is the kind of phrase that has been keeping business reporters employed since the invention of the phone. This is not a done deal. Nobody has signed anything. But the direction is what interests me. PayPal's CEO Enrique Lores — who joined in March after years at HP — has been running a turnaround. Back in April he split the business into three operating models: checkout and PayPal, consumer financial services and Venmo, and payment services and crypto. He told investors PayPal needed to "recommit to the fundamentals" and become "a technology company again." And the turnaround also includes cutting the workforce by twenty percent over the next two to three years.
Here's the thing that should make every builder sit up. PayPal was founded in 1998. It is one of the original Silicon Valley institutions. Peter Thiel, Elon Musk, Max Levchin — the so-called PayPal Mafia came out of that building. And now the live question is whether Stripe, the scrappy upstart that ate into PayPal's developer business one clean API at a time, buys the elder statesman. If that happens, that's not a merger, that's a generational handoff. The company that won developers by being easier to integrate is now positioned to absorb the company that defined online payments. If you're building anything that runs on payment rails, you want to watch this. Consolidation at the infrastructure layer means fewer choices and, eventually, pricing power concentrated in fewer hands.
Speaking of the App Store tax, let's stay on the money-out-of-your-pocket theme, because Apple made a move here that developers have been waiting years for.
After trying and failing to delay the whole thing, Apple submitted its proposal for what it wants to charge on purchases made through external links inside iOS apps. This is from Sarah Perez's reporting. Apple proposed a fifteen percent commission for standard apps on those external-link purchases. There are discounts: small business developers would pay five percent, apps in the Video Partner Program, News Partner Program, and Mini Apps Partner Program would pay ten percent, and subscription renewals drop to ten percent too.
Now, some context for why this even landed. Apple's been in a years-long brawl with Epic Games over App Store commissions. Apple tried to argue the lower court should wait until the Supreme Court ruled on a separate question — whether Apple was in contempt for imposing a twenty-seven percent commission on external-link purchases and restricting how developers could show those links. On Thursday the Supreme Court rejected Apple's bid to pause the lower court case, and that forced Apple to show its cards.
Here's my read as somebody who's watched this fight for a decade. A fifteen percent cut on purchases made outside the App Store is still Apple charging you a toll for the sin of leaving Apple's toll road. Their argument is they should recoup investment in the tools and technology that keep the App Store running. And they helpfully pointed out that Google Play charges twenty percent link-out for standard apps. So Apple's basically saying, "Look, we're cheaper than the other guy." That's a real shift. Two years ago the idea that Apple would voluntarily propose fifteen percent on off-platform purchases would have gotten you laughed out of the room. But before anybody throws a party, understand that "the court made me" is not the same as "we believe developers deserve this." The number moved because the legal pressure moved it. If you're a founder building a consumer app with subscriptions, run the math now: five percent if you qualify as a small business, ten on renewals, fifteen standard. That changes unit economics, and it changes whether an external checkout is even worth the friction.
Now let's talk about the thing that's been all over the AI world this week, and I'll be honest with you — we touched the Claude watermark story yesterday, so I'm not going to re-plow the same field. But Anthropic put out fresh detail on Friday that's worth a quick, specific look, because it answers the questions builders actually asked.
Yesterday we covered the fact that Anthropic is watermarking Claude's text output to comply with the EU AI Act's Transparency Code. What's new, from Anthony Ha's reporting, is the how. Anthropic explained that when Claude makes what they call "low-stakes choices" — like picking between the words "overcast" and "grey" to describe the weather — it can bake a pattern into those word choices that's, quote, "undetectable to the reader, but is detectable to anyone who has a key that encodes it." They're using the SynthID-Text approach that the Google DeepMind team outlined in 2024, and they plan to release a watermark detection API.
Here's the part that matters for anybody using Claude to write code. Anthropic says code should carry much less of a watermark, because the model has to produce working code and doesn't get to freely choose between a bunch of equally valid options. Their words: "in areas where there is an arbitrary choice between particular words or terms within the code, the watermark can be used, such as comments within code, but by definition, it will have a negligible effect on the actual code produced." So your functions aren't getting fingerprinted — your comments and prose might.
And can you scrub it? Anthropic says light editing probably won't remove the watermark completely, but a complete rewrite where every word is replaced will. And they made a dry little point I appreciated: at that point, "it's arguable whether the text can any longer be described as AI-generated." Which is honest. If you replaced every word, did the AI really write it?
Now the reaction. Business Insider reported dozens of users on X claimed to cancel their Claude subscriptions over this. On Reddit one poster called it a conspiracy against innocent Claude users, while another shot back, "The only reason you wouldn't want this is to lie to people." And Anthropic noted Claude won't be the only one — other major model developers signed the same Code of Practice and will roll out their own watermarks.
My take, briefly, because I don't want to over-index on a compliance story: the watermark is a European regulatory artifact showing up in your American workflow because these are global products. If you're a founder shipping Claude-generated content — marketing copy, docs, whatever — the watermark won't hurt quality, but you should know it's there. It's not spyware, it's a compliance stamp. The panic is louder than the problem.
Alright, let's get to something darker, and I'm going to handle it carefully because it deserves that.
There's a lawsuit against Elon Musk's xAI over its chatbot Grok, and a new plaintiff joined it. A woman identified as Jane Doe 4 has joined a suit filed by three Tennessee teenagers, alleging Grok played a role in creating child sexual abuse material. According to a Washington Post report cited by TechCrunch, the woman alleged that her stepfather used Grok to manipulate a photo taken when she was eleven years old, and generated more than seven thousand explicit images of her. She said her stepfather was found dead of suicide two days after the images were uncovered in a law enforcement raid. Her words: "Limitless access to these tools is spreading so quickly. It is taking everyday life and turning it into child sexual abuse."
The teenagers who filed the original suit accused xAI — which is now part of SpaceX — of failing to take basic precautions to prevent Grok from being used to make explicit images of real people, including minors. They're seeking class action status. TechCrunch reached out to xAI for comment. And I'll note, as the source does, that if you're in crisis or having thoughts of suicide, you can call or text 988 in the US to reach the Suicide and Crisis Lifeline.
I'm not going to sensationalize this. I'm going to make one builder point and move on. If you ship a generative image tool without guardrails against this, "we're a platform, we're not responsible" is a legal position that is aging very badly in real time. The industry spent a couple years treating safety filters as a tax on creativity. Cases like this are the bill coming due. If you're building anything that generates images of people, the precautions aren't a nice-to-have anymore. They're the difference between a product and a defendant.
Let me pull us back to the physical world for the deep dive, because this is the one I think matters most for builders and founders, and it connects straight back to that missile story I opened with.
The piece is from Tim De Chant, headline "Hyperscalers might regret embracing natural gas if new forecast proves correct." And I'm going to spend real time here because the argument is important and most people building AI products aren't thinking about it at all.
Here's the setup. After years of buying up wind and solar, the hyperscalers — Amazon, Google, Meta, Microsoft — have pivoted to natural gas to power the data centers behind their AI ambitions. And they've gone big. In March, Meta said it would build a massive seven-and-a-half gigawatt natural gas power plant in Louisiana to power its Hyperion data center. A few days later, Microsoft and Google each said they'd build their own gigawatt-scale gas plants in Texas. And Amazon plans a seven-point-six gigawatt gas plant in Texas. These are companies that historically hated big capital expenditures, and now they're not just building physical plants — they're wading neck-deep into energy markets, which is even less familiar territory for them.
Now here's the forecast that should worry them. According to Noreva, an energy research firm, natural gas prices could triple in some parts of the US in the coming years. Peter Gardett, the CEO of Noreva, put it to TechCrunch like this: "I think everyone in the energy markets has been lulled into a sense that gas prices can't go up. You just need simple arithmetic to get to a much tighter gas market than you were in just a few years ago."
Let me give you the numbers, because they're stark. Today, gas prices range from about two dollars to four-fifty per million BTUs, with the widely traded Henry Hub in Louisiana sitting just under three bucks. Noreva expects prices to soar above ten dollars per million BTUs in certain hubs. And here's why fuel matters so much: fuel is about half the cost of electricity from a large power plant. So a doubling or tripling of gas prices makes "bring your own power" AI data centers dramatically more expensive to run. And Gardett spelled out exactly where that pain goes. Quote: "That could drive up token costs, or it could push hyperscalers to connect to the grid, driving electricity prices higher."
Read that again, founders. Token costs. The price you pay per million tokens on your API bill has a hidden dependency, and that dependency is the price of natural gas in West Texas. That is a sentence I did not expect to say on this show, but that's where we are.
Now, why does Gardett think prices are going to move when the futures market is calm? Two reasons. One, gas prices have been stable for years because demand was flat and new supply kept pace. He thinks companies can still add supply — just not at the old rate — and new wells are getting more expensive. But that alone, he says, wouldn't change the math. What changes it is two things. His words: "What's changing the number is that finally we're connecting the domestic gas market to the global gas market. And the second is the AI demand pull."
Let me unpack that first part, because this is the missile story all over again. For years in West Texas, most wells were drilled for oil, and the natural gas that came up was a byproduct nobody had a market for. There weren't big pipelines to move it out, so producers basically sold it at a discount to anybody who'd take it. That's why the hyperscalers went there — cheap gas. But, Gardett says, "They've finally built some pipelines out there, and a lot of that is headed towards export markets." So West Texas gas is getting connected to national and international markets. And once it's connected, the local cheap gas isn't local and cheap anymore — it's priced against global demand. You'll get, in his words, "places where you get a lot of gas next to someplace where there's none," and those differentials drive certain regions above ten dollars for extended stretches.
There's the connection to the opening story. Ukraine strikes a rocket factory because the physical supply chain is a target. And here, the American gas market gets tied into a global system, and suddenly your data center's fuel cost is exposed to what's happening in the Strait of Hormuz. In fact — and this is the eerie part — that all-electric aircraft story I saw this week noted jet fuel prices have jumped from just over two dollars a gallon to well over three because of Gulf energy disruptions. The global energy system is one connected organism now, and AI just plugged itself into the most volatile part of it.
And the last bit of the piece is the one I want founders to really hear. Gardett said there's already a consumer backlash forming — eighty percent of consumers are worried about data centers' impact on their utility bills, and that's mostly about electricity. Now imagine that anxiety spilling into natural gas bills too. And then the closer, which I think is the whole ballgame. Gardett said: "On future Alphabet earnings calls, you will hear them talk about the correlation between natural gas pricing and Google results, which is strange, but that's where we are."
Think about that. A search-and-ads company's quarterly earnings being tied to the commodity price of methane. That's how deep AI has pushed these software companies into the physical economy.
So what do you do with this if you're building? A few things. One, if your cost model assumes the price of intelligence keeps falling forever, build in a scenario where inference costs stop falling — or tick up — because the underlying power got more expensive. Two, if you're picking a model provider, understand that the ones with "bring your own power" gas plants are taking on commodity risk that they will eventually pass to you. And three, this is a reminder that the moat in AI is drifting away from clever software and toward who controls energy and compute. Which brings me neatly to two more stories that are all about who owns the physical layer.
First, and I'll frame this as the older story it is — this closed back in April as an option, and the deal became official this weekend — SpaceX officially closed its acquisition of Cursor. The AI coding startup is now part of SpaceX, per an announcement on Cursor's blog, reported by Anthony Ha. To recap the timeline: SpaceX, which also absorbed Musk's xAI earlier this year, announced a deal in April with an option to buy Cursor for sixty billion dollars. Then, as SpaceX became a public company, they moved forward. Now it's done.
And listen to how Cursor framed it, because it's the exact thesis of today's whole episode. Cursor said that by becoming part of SpaceX, it'll have "access to the largest fleet of GPUs in the world." And they added, "SpaceX is building the computing capacity needed to scale intelligence far beyond what exists today. Cursor will be one place where that intelligence becomes useful." SpaceX has been renting that compute to customers including Anthropic and Google. And, worth flagging, SpaceX faces a lawsuit over pollution from its data center gas turbines — so even Musk's compute empire is tangled up in that same natural-gas-and-backlash story.
Here's my continuity note for you: we've been circling the theme of compute-as-power all month — King's Cross becoming an AI hub, Texas halting its interconnection queue over the power crunch, Base Power's home batteries. This Cursor deal is another data point in the same trend, but I don't want to abstract it into a tidy meta-pattern, so let me be concrete. The thing that made Cursor valuable was a great coding product on top of other people's models. The thing SpaceX bought it for is to have a useful destination for its own compute. That's the tell. When your product becomes an amenity attached to somebody's GPU fleet, you've learned who actually held the leverage. Founders building on top of the model layer: notice that the acquirers here aren't buying the software, they're buying the funnel for their infrastructure.
And on the other side of the compute question, there's a genuinely different bet worth two minutes. This is a recent report from Mistral, the French lab. They announced three things aimed at what they call AI sovereignty. Mistral Regional Endpoints, now generally available, let customers pick whether their inference runs in Europe or the US, to match data-residency and regulatory needs. A new Mistral Priority Tier, in public preview, offers committed service levels with an uptime SLA. And here's the interesting one — they're opening their platform to third-party open models, starting with Z.ai's GLM-5.2, running on the same regional-controlled infrastructure.
The strategic pitch is European independence. Mistral is building toward up to one gigawatt of capacity by 2030 and rounding up an anchor group of big enterprises — names like ASML, Capgemini, Amadeus, CMA CGM — to make multi-year compute commitments through what they're calling European Compute Units. The CEO of Factory, Matan Grinberg, gave the builder's version of the argument: "Different workloads need different models, and that will keep changing as the frontier evolves. Mistral allows us to run open models under strict regional controls and service commitments."
My honest read: some of this is a regulatory-arbitrage play dressed up in the language of sovereignty. But there's a real product idea underneath. If you're a European enterprise, or you sell to one, "where does my inference physically run, and can I get an SLA on it" is becoming a purchase requirement, not a nice-to-have. Mistral is betting that data residency plus model choice plus a contract you can enforce is a durable wedge against the American giants. It's the same energy-and-compute-sovereignty theme, just fought on regulatory ground instead of gigawatts.
Let me change gears entirely, because not everything is a heavy infrastructure story, and there are a couple of developer-relevant items I want to get to.
OpenAI rolled out a new mode called Ultrafast, per Lucas Ropek's reporting. It's designed to speed up their most powerful model, GPT-5.6 Sol, by up to fourteen times — delivering up to seven hundred fifty output tokens per second. OpenAI's framing was interesting: "Until now, getting real-time speed typically meant choosing a smaller or more specialized model. Ultrafast points to progress in a new direction: more useful work per second." They're pitching it for incident response, customer service, financial market analysis, e-commerce. It's in preview to a small group, and — here's the detail that matters — it's powered by OpenAI's partnership with the chipmaker Cerebras.
Why do I flag Cerebras? Because it's the same theme yet again, dressed as a speed feature. The way OpenAI got fourteen-x speed on the frontier model isn't a software trick, it's specialized silicon from a partner. Speed is a hardware story now. And for builders, "more useful work per second" on your best model without dropping to a dumber one — that's genuinely valuable for anything real-time. Watch whether the pricing on Ultrafast makes it usable in production or whether it's a premium halo feature.
Quick companion note: OpenAI also put out a builder's guide to GPT-5.6, aimed at startups building faster, cheaper AI agents with smarter model selection and new Responses API capabilities. I'll flag it because it's exactly the "which model for which job" muscle every founder needs to develop now that the menu has ten items instead of two. The link's in the show notes.
And here's a sharp little observation from developer Simon Willison that punctures the whole model-treadmill absurdity. He was looking at Google's pricing for the Gemini 3.7 Flash model and noticed the "introductory pricing" is scheduled to double on December 31, 2026. His comment: "who would anticipate still using this model five months from now? Especially since 3.6 Flash came out just three weeks ago!" And that's the joke, right? The models are shipping so fast that a pricing schedule with a five-month horizon is basically science fiction. If you're building on these, don't over-optimize for a specific model version. Build your harness so you can swap the engine, because the engine is going to change before your pricing negotiation even matures.
Let me get to a couple of stories about transparency and the physical world before I wrap, and I'll keep these tight.
X — now merged with SpaceX — put out a recent report that it's open-sourcing a much bigger chunk of its ranking codebase, including the "For You" algorithm and the core ranking engine, on GitHub under the Apache v2 license. This is Sarah Perez reporting again. The codebase is now ten to fifteen times larger than before. They're also adding a transparency tool that lets users who've posted ten or more times in the past month download their aggregate stats as a JSON file, to see if any labels got slapped on their account or posts — a way to check if you've been, in the popular term, shadowbanned. X's VP of Product Keith Coleman said, "This is the kind of thing that I think people will be fairly shocked that we are releasing."
Now, the honest caveats. A few systems aren't included — notably the ones that use Grok to predict rule violations, which X says they held back so bad actors can't game spam filters. And there's a real irony the piece points out: X pushing transparency here while it became much less transparent overall after going private, reporting less on user metrics, revenue, and government takedown requests. So take the "we're so open" framing with the appropriate seasoning.
But the builder angle is genuinely interesting. There's a detail where non-technical users are supposed to take that JSON file, drop it into an LLM, point the AI at X's GitHub repo, and ask for an interpretation. Think about what that is. The transparency mechanism assumes you'll use an AI to read the algorithm's output about you. That's a new pattern — the recommendation system is now legible only through another AI. If you're building consumer products, that's a preview of where "explainability" is headed: not a human-readable dashboard, but a machine-readable dump plus a model to translate it.
Two more quick physical-world hits and then I'll let you go, because they show how AI and hard infrastructure keep colliding.
Self-driving trucks are officially testing on California highways — this is an older thread with a fresh regulatory step. Aurora Innovation and Kodiak AI got permits from the California DMV to test autonomous heavy-duty trucks on public roads, per Kirsten Korosec. Kodiak's already started, running test trucks around its Mountain View office. The rules, approved back on April 28, lifted a ban on driverless vehicles over ten thousand pounds. There's a human safety operator required behind the wheel for now. And the opposition is real — the Teamsters California sued the DMV, alleging it skipped required studies on the economic impact of putting self-driving trucks on the road. That's the jobs fight, out in the open. Both companies had been running mostly in Texas because California banned the heavy trucks; now they can operate at home.
And there's a cautionary infrastructure tale that every founder building on the cloud should tattoo somewhere visible. This one first surfaced in July. A PBS station — Nine PBS in St. Louis — is fighting to recover fifty terabytes of data going back seventy years after its cloud storage provider went defunct, per Scharon Harding. The provider, Open Source Storage, ceased operations and became unresponsive. Nine PBS's data was sitting in an Iron Mountain data center, but Iron Mountain said, essentially, "we just rent the building and the power; the servers belong to our customer, and we don't have access to the data." A judge has now ordered Iron Mountain to hand over the physical devices, and Nine PBS has thirty days to find a former OSS worker to help retrieve it — assuming it isn't encrypted, in which case there's another hearing.
Fifty terabytes. Seventy years of a region's history — COVID coverage, the Great Flood of 1993, over eleven thousand files, described as "unique and irreplaceable." Trapped because the vendor died and nobody could get to the drives. The lesson isn't complicated, but people ignore it constantly: your data is only as safe as your ability to get it out. Read your exit clause. Keep an independent backup you actually control. And understand the layers — the company you have a contract with may not be the company holding the physical hardware, and when the middle vanishes, you're stuck in court. That's not a technology failure. That's a supply-chain failure, same as the missile, same as the gas. The chain is only as strong as the weakest handoff.
So let me tie the bow on today. We opened with a cruise missile hitting a rocket factory that was already short on capacity, and if you strip away the specifics, almost every story today is a version of that. Natural gas prices threatening to triple and quietly reaching up into your token costs. SpaceX buying Cursor to have a use for its GPU fleet. Mistral rounding up gigawatts and multi-year compute commitments in the name of sovereignty. A PBS station locked out of its own archive because a vendor died. The common thread isn't a slogan — it's that the value and the risk in this business keep migrating toward whoever controls the physical inputs and the physical handoffs. If you're a founder, your job is to know where in that chain you actually sit, and how exposed you are when the link two steps upstream from you snaps.
That's the plate for today, kiddos. Watch the gas prices, back up your data, and read the exit clause before you sign anything. This has been Barely Possible. I'm Tony DeLuca — take care of yourselves, and I'll see you next time.