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 this is Barely Possible, the show where we pick through the day's pile of AI noise and try to find the two or three things that actually change how you build. Got a fresh menu today: a Google lawsuit that should make every founder think hard about what their product can be turned into, an Indian video model that's nearly free per second, Mistral doubling its valuation, and a price war brewing between the two labs you're probably already paying. Pour the coffee, buckle up, let's have at it.
Let me start with the story I think most builders will scroll right past, and shouldn't. Google filed a lawsuit this week against a Chinese cybercrime network that allegedly used its own Gemini models to automate a mass scam operation. The group, named in the filing as Outsider Enterprise, used AI to target, by Google's count, hundreds of thousands of victims. The number that stopped me cold: 2.5 million text messages sent over a span of two weeks. That's not a guy in a basement with a burner phone. That's an assembly line. And the assembly line was built, at least in part, with Gemini coding the scam sites themselves.
Now why does this matter to you, the person building a product, and not just to Google's legal department? Because this is the clearest signal yet of how the platform companies are going to handle the fact that their own tools are now the best phishing-kit-builder, scam-site-generator, and social-engineering-script-writer on the market. They're not just patching the model. They're suing the abuser. Ars Technica covered the same case, and the framing is worth holding onto: the fraudsters allegedly used Gemini to code the scam sites at scale. The model didn't decide to scam anybody. It did what it was told, fast, cheaply, and at volume. That's the whole pitch of these tools. The same thing that makes Gemini or Claude or GPT a 10x force multiplier for your legit startup makes it a 10x force multiplier for the guy running the SMS fraud ring.
Here's the builder takeaway, and I want to be plain about it because it's easy to wave off as someone else's problem. If you're putting an AI product into the world, you are now, whether you like it or not, in the abuse-tooling business. Somebody is going to try to use your thing to do something ugly. Google has the legal department and the political cover to chase a Chinese network through the courts. You probably don't. So the question for your roadmap isn't "will this get abused," it's "what's the cheapest abuse vector I'm shipping, and have I even looked at it?" Two and a half million texts in two weeks. That's the bar now. That's the scale a small crew can hit when the model does the heavy lifting. Build accordingly, and watch this case, because the legal theory Google is testing here, that you can go after the operator of an AI-powered fraud network directly, is going to set the template the whole industry copies.
That's the abuse side of cheap, fast AI. Now let me flip it to the legit side of cheap, fast AI, because they're the same coin. Over in India, TechCrunch's Ivan Mehta wrote up a company called Avataar AI, and the headline number is the one you need: their distilled video model is priced at half a cent for every second of generation. Five-tenths of one cent per second. Cheaper, faster, and as the piece puts it, culturally aware, built for India's scale.
Let that price sink in if you've ever paid for video generation. We're at the point where a localized, distilled video model is being offered at a price that's basically a rounding error per clip. And the "culturally aware, built for India's scale" part is the strategic move, not the marketing fluff. The frontier labs are building one giant model to rule them all, trained mostly on Western data, priced for Western budgets. Avataar is doing the opposite: distill it down, tune it for the local market, price it for a market where a half cent per second actually opens up use cases that a frontier price tag would kill on contact.
I keep coming back to this because it's the thing the trillion-dollar capex headlines miss. The interesting action in video and image generation right now isn't only at the absolute frontier. It's in distillation and localization. Take a capable model, shrink it, specialize it, and undercut everybody on price in a specific market. If you're a founder outside the US thinking the AI game is already lost to the labs with the biggest GPU piles, Avataar is your counterexample. The distribution and the cultural fit and the price are a moat the frontier labs are structurally bad at building, because they're optimizing for the global average user and you're optimizing for the actual person in front of you.
Now stay on the money for a minute, because there's a connected story sitting right next to it. Mistral, the French lab, is rumored to be raising three billion euros at a roughly twenty billion euro valuation. That's about twenty-three billion dollars, and it's nearly double their Series C valuation of eleven point seven billion euros. TechCrunch's Ram Iyer has the rumor, and I'll flag it as exactly that, a rumored round, not a closed one.
But the direction tells you something even as a rumor. Europe's flagship lab roughly doubling its paper value. Avataar undercutting on video price in India. These aren't the same company or the same strategy, but they rhyme: the non-American AI ecosystem is raising real money and shipping real product, and the bet underneath Mistral's number is that there's durable demand for an alternative to the American labs, whether that's for sovereignty reasons, regulatory reasons, or plain old not-wanting-to-be-locked-into-one-vendor reasons. If you're building on top of models and you've been assuming the supplier list is just OpenAI, Anthropic, and Google forever, the capital flowing into Mistral is the market telling you to keep your options open.
And speaking of those American labs, let's talk about the price war they may be about to start, because this one hits your unit economics directly. The Wall Street Journal reported, and I'll note this is via the Journal so some of you will hit a paywall, that OpenAI is considering drastically lowering its prices for tokens. The reported reasoning is interesting: they're not cutting because demand is soft. They're cutting in anticipation of similar cuts they expect Anthropic to make. A preemptive strike. Cut first, before the other guy does, and grab the customers in the scramble.
Now I covered the Simon Willison angle on this yesterday, where he reframed the rumor as OpenAI possibly finding Anthropic's latest models surprisingly good and reacting to that, so I won't relitigate it. But there's a fresh wrinkle worth adding today, and it comes from SemiAnalysis. They ran an experiment: bought all the consumer plans from both Anthropic and OpenAI, then pushed them to their weekly limits with intense coding tasks to figure out what you're actually getting for your money. Their finding: OpenAI's two-hundred-dollar Pro plan delivered up to fourteen thousand dollars in monthly API-equivalent value. Anthropic's two-hundred-dollar Max plan reached around eight thousand. Many, many times the subscription cost.
Sit with that. The labs are handing power users somewhere between forty and seventy times their subscription fee in raw inference value. SemiAnalysis is blunt about why: these are heavy subsidies designed to hook developers. And the skeptic's question, which they raise themselves, is the obvious one: how long does that last? Because the limits evolve, the rates get adjusted, and the subsidy can't run forever.
Here's the connection between the rumored price cut and the subsidy numbers, and it's the part that should shape how you plan. If OpenAI is genuinely considering slashing token prices, and if the consumer plans are already running at this kind of subsidized blowout, then the per-token economics you're building your margins on today are not a stable input. They might get cheaper, which is great for your cost of goods, right up until the day the subsidy logic flips and they raise prices or tighten the limits on the heavy users who are the most expensive to serve. The strategically dangerous move for a founder is to build a business whose entire margin depends on inference staying this cheap. Enjoy the subsidy. Just don't architect your company assuming it's permanent, because the people running it are openly telling SemiAnalysis it's a customer-acquisition loss leader.
Let me shift from model pricing to a different kind of bill coming due, the physical one, because this is where the deep dive is today.
The story is data centers, but not the version we usually do. We've spent plenty of episodes on the trillion-dollar capex forecasts and the power deals. Today I want to zoom into something more concrete and, frankly, more politically loaded: the fact that, per Ars Technica's Ashley Belanger, roughly a hundred and thirty billion dollars in data center projects have been blocked by local protests so far this year. A hundred and thirty billion. Blocked. Not delayed by permitting paperwork, not slowed by a chip shortage. Stopped by people showing up at town meetings.
And the line in that piece that I can't stop chewing on is this: winning a fight against an AI data center gives people a "taste of political power." That's the real story here, and it's why I'm making it the centerpiece. For years, the data center buildout has been treated as basically inevitable. The hyperscalers pick a county, dangle some jobs and tax revenue, and the local government rolls out the welcome mat. What's happening now is that ordinary residents are discovering that a coordinated "no" actually works. And once a community wins one of these fights, they don't forget the feeling. They've learned they can stop a multi-billion-dollar corporation cold. That's a genie that does not go back in the bottle.
Now pair that with the companion piece, also from Ars, by Kyle Orland, on water use. The headline is almost reassuring at the macro level: when it comes to total water use, AI data centers are a drop in the bucket. Nationally, agriculture and power generation dwarf what the data centers pull. If you only read the top line, you'd think the water worry is overblown. But the actual finding is sharper, and it's the reason the protests work: even a moderately sized data center can have an outsized local impact. The national number is a rounding error. The number in your county, drawing from your aquifer, in a drought year, is not. People don't experience national averages. They experience the well running low and the reservoir dropping and the new neighbor that uses more water than the whole rest of the town.
So here's the thing that ties these two pieces together, and why I think founders building anything compute-heavy need to internalize it. The economic case for these projects is made at the national and global scale. The cost is felt at the local scale. That mismatch is exactly the wedge that's now stopping a hundred and thirty billion dollars in projects. The water isn't a national crisis, but the local impact is real enough to organize around, and a win on the water fight teaches the community it can win the next fight too.
What does this mean if you're not building a data center yourself? Two things. First, the compute you're renting from the hyperscalers has a physical supply chain that is now facing organized, effective political resistance in a way it wasn't a year ago. That's a long-term input risk on the most basic thing your AI product depends on: somewhere to run the models. The cheap, abundant compute we've all been assuming is a permanent feature of the landscape is running into the same NIMBY dynamics that have stalled housing and transit for decades, except the dollar figures are bigger and the local impacts, the water, the noise, the power draw, the truck traffic, are more visceral.
Second, and this is the strategic part: scarcity and local resistance both push in the direction of the Avataar story I told you earlier. If frontier-scale compute gets harder and more expensive to physically site, the economic pressure to do more with smaller, distilled, efficient models goes up, not down. Every protest that blocks a gigawatt data center is, in a roundabout way, an argument for the team that figured out how to ship a good-enough model at half a cent per second. The local fight over the aquifer and the lab fight over token efficiency are two ends of the same rope.
And let me be fair to both sides here, because I don't want to caricature this. The hyperscalers aren't villains for needing power and water, and the residents aren't Luddites for not wanting their water table drained. This is a genuine collision between a national-scale economic project and local-scale physical limits, and the political system is going to have to actually resolve it instead of pretending the buildout is frictionless. For now, the scoreboard says: a hundred and thirty billion blocked, communities learning they have power, and a water story where the national number says "relax" and the local number says "organize." If you're forecasting your compute costs three years out, factor in that the people who live next to your compute have just figured out they get a vote.
All right, let me come down from the heavy stuff and run through a few things that matter but don't need a full sermon.
Jeff Bezos has a new startup, and we now know what it's for. It's called Prometheus, and per Ars Technica's Samuel Axon, it's going after physical AI. It's not the only company chasing this, but it's one of the best-funded. The bigger ambition, as Bezos has framed it elsewhere, is what he calls an artificial general engineer, an AI that can design and help manufacture physical things, jet engines being the example that gets thrown around. The interesting structural detail, and I'll keep this to a mention because the financials are still soft, is the reported idea of pairing the technology with an industrial buyout fund. The logic there is actually sharp: the physical economy can't be scraped. There's no internet of manufacturing data sitting around to train on. So if you want that data, you don't find it, you buy the factories that generate it. That's a genuinely different playbook from the scrape-the-web approach that built the current models, and it's worth watching whether "acquire the data source" becomes the pattern for physical AI the way "scrape everything" was for text.
Next, Anthropic. Two items today. One, they published results from the first Anthropic Public Record, which is their transparency effort. Two, and more relevant to builders, TCS, Tata Consultancy Services, is partnering with Anthropic to bring Claude into regulated industries. Now if you've been listening, you'll remember we covered the DXC and Anthropic alliance recently, same basic shape: get Claude embedded into the big systems integrators that banks and airlines and insurers actually rely on. The TCS deal is another brick in that wall. The pattern is unmistakable now. Anthropic's enterprise strategy is to go through the integrators, the firms that already have their hands inside regulated corporate IT, rather than trying to sell directly into every compliance department one at a time. If you're building enterprise AI tooling, that channel, the systems integrator as the distribution layer for frontier models, is the lane the leader is betting on. Take the hint or find a reason you're different.
A quick security note that's genuinely important if you run anything on Oracle's stack: Dan Goodin at Ars reports a PeopleSoft zero-day affecting hundreds of organizations, and it's stealing gigabytes of data. PeopleSoft is Oracle-owned, it's buried inside HR and finance systems at a lot of big institutions, and Goodin describes the vulnerability as about as critical as they come. We touched on the Oracle PeopleSoft breach angle in passing a couple days back, but this is the fuller picture: if PeopleSoft is anywhere in your org's plumbing, this is a today problem, not a next-sprint problem. Go check.
On the policy front, Section 702 of FISA, the law that authorizes a big chunk of the NSA and FBI's warrantless surveillance, is expiring for the first time. TechCrunch's Zack Whittaker reports it's all but certain to lapse after lawmakers rejected the administration's pick to lead the spy agencies. But, and Ars Technica's Jon Brodkin makes this the whole point, the spying will continue. The existing certification runs until March of 2027. So the law expires, the headlines say "surveillance authority lapses," and in practice the wiretaps keep humming for another nine-plus months. For builders, the takeaway is narrow but real: nothing about your data-handling obligations changes overnight here. The framework that lets the government compel cooperation from telecoms and platforms is still operational. File it under "watch the March 2027 date," not "the rules just changed."
A couple of war-and-autonomy items worth flagging, briefly. Ars reports that Ukraine ran a one-time test using fully autonomous drones to kill Russian soldiers. Full autonomy, no human in the loop on the kill decision. The piece is careful: this kind of full autonomy is still rare, mostly Ukraine is installing AI modules onto otherwise human-controlled drones and robots. But "rare" and "never" are very different words, and that line got crossed in a test. I'm not going to pretend to have the geopolitics solved on a tech show, but I'll say this for the builders working anywhere near defense or robotics: the autonomy-in-the-loop question is no longer theoretical, and the people writing the procurement checks have noticed.
And then there's the big one everybody's already heard about, so I'll be quick because we did the IPO mechanics yesterday. SpaceX went public, closed up nineteen percent over its hundred-and-thirty-five-dollar offering price, and the result is that Elon Musk is now, on paper, the world's first trillionaire. TechCrunch's Sean O'Kane notes the obvious tension: this lands at a moment when Musk is, as the piece puts it, more hated and more powerful than ever. Robinhood saw record-breaking traffic as retail piled in, with some customers hitting intermittent disruptions that have since resolved. And SpaceX president Gwynne Shotwell dropped another hint at a SpaceX-Tesla merger, which a lot of people now read as inevitable.
I covered the SPV risk and the retail-bloodbath worry yesterday, so I won't repeat it. The one fresh thing I'll add for builders: the relevant fact here isn't the trillionaire headline, it's that SpaceX is increasingly being priced as an AI infrastructure and neocloud play, not just a rocket company. Those data center deals it signed, including the compute arrangements we've talked about, are part of why the market is willing to put a near-two-trillion valuation on it. The IPO is a referendum on the AI infrastructure buildout as much as it's a referendum on Elon. And that ties right back to the deep dive: the same buildout that the public market just blessed with the biggest IPO in history is the buildout that local communities are blocking a hundred and thirty billion dollars of, county by county. The capital markets and the town halls are voting on the exact same thing, and they're voting in opposite directions. That tension doesn't resolve cleanly, and I'd keep both eyes on it.
One last small one, because I think it's a healthy reminder. Simon Willison, who's about as level-headed as anyone on what these coding models actually do day to day, had a sharp little clarification this week. Somebody was alarmed about AI-generated debug code, and Simon's correction was simple: it's not code for shipping, it's throwaway debug code the model used to help track down the bug. I bring it up because in a week full of trillionaires and price wars and autonomous drones, that's the actual texture of how most of us use this stuff. The model writes some scaffolding to find a bug, you throw it away, you keep the fix. Not every interaction with AI is a civilizational turning point. Sometimes it's just a smart intern leaving sticky notes you toss when you're done. Worth remembering when the hype gets loud.
So here's where we land today. The two stories I'd actually act on are at opposite ends of the same scale. At the small end, Avataar shipping good video at half a cent a second is your reminder that distillation and localization are a real strategy, especially when frontier compute gets scarce and expensive. At the large end, a hundred and thirty billion in blocked data centers is your reminder that the physical foundation under all of this is now contested ground, county by county, and the people living next to your compute just figured out they have a vote. And running underneath both: a token price war that makes today's costs unstable in both directions. Build like the cheap compute is a gift you didn't earn and shouldn't bank on.
That's the menu. I'm Tony DeLuca, this has been Barely Possible, and I appreciate you spending part of your day in here with me. Go check your PeopleSoft, watch that March 2027 date, and I'll see you tomorrow.