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, welcome back to the diner. I'm your boy Tony DeLuca, this is Barely Possible, and today's menu is heavy on the money side of the ledger. We got the government letting Anthropic's most guarded models off the leash, we got Cloudflare drawing a line in the sand on who pays for the open web, and we got a battery boom happening on the roofs of American houses that nobody's really connecting to the AI story yet. Buckle up, pour yourself something hot, and let's have at it.
I want to start with the one that lands squarely in the founder's lap, because it's a story we've been circling for about a week now and it finally moved. The US government has lifted its curbs on Anthropic's advanced Fable and Mythos models. That's the reporting from Ashley Belanger at Ars Technica, and the headline tells you the whole arc: after spooking Trump into safety testing, Anthropic's AI models get a global release.
Now let me connect the dots for you, because if you only heard today's news cold it would sound like a happy ending. Go back to the end of June. We talked on this show about the government rationing frontier models, Anthropic's Mythos and OpenAI's competitor, handed out on allowlists like they were controlled substances. Dario Amodei himself, and I'm not making this up, compared Mythos to a superweapon, said one of the companies they showed it to called it something you should need a gun license to use. That's the framing that got Washington's attention. When the guy who built the thing tells you it's dangerous, the government tends to listen.
So the curbs went on. And now, per this recent report, they're coming off, and Fable and Mythos get a global release. Anthropic also put out its own note, plainly titled Redeploying Fable 5. So the model's coming back.
But here's the part I want you builders to sit with, because it's not in the triumphant headline. The redeployment isn't a clean return to the old world. From the chatter around the relaunch, it looks like Fable usage is going to be credit-based, billed separately from the subscriptions, not just folded into your monthly plan. And more than that, the access looks like it's going to require identity verification. Submit your documents, get your credits once your identity is confirmed. Now that's still being pieced together from code strings and leaks, so I'm not going to state it as gospel, but the direction is unmistakable. And it lines right up with that gun-license comparison. When you regulate something like a firearm, you get a firearm's paperwork.
So think about what that means if you're building on top of these models. The frontier tier is starting to look less like a cloud API and more like a licensed pharmacy counter. You want the strong stuff, you show ID, you pay per dose, and there's a paper trail. That is a very different product surface than the all-you-can-eat token buffet founders got used to. And I'd bet money the KYC-to-use-the-model pattern doesn't stay confined to one lab. Once one frontier provider makes identity verification the price of admission, and the government likes it, the rest follow, because nobody wants to be the lab that skipped the safety theater when the subpoena comes.
One more thread here, and it's the geopolitical undertow. While the US was busy putting Anthropic's models behind glass, there's reporting that Asian AI startups moved to launch Mythos-like models to fill the gap the export ban left open. And separately, there's a report that Austria has floated the idea of luring Anthropic to set up shop in Europe as part of a broader EU push for AI independence. I can't nail down firm dates on those two, so I'll frame them as what they are — signals swirling around the same event, not confirmed done deals. But the shape of it is clear enough. When you fence off the best models at home, two things happen: somebody overseas builds a substitute, and somebody overseas offers your lab a nicer neighborhood. That's not a prediction, that's just how fences work.
Now let's shift from who gets to use the models to who gets paid when the models eat the web. Because this next one is, in my book, the most consequential story of the day for anybody running a content business or building an agent.
Cloudflare has a new policy, and it's got teeth. Sarah Perez at TechCrunch laid it out. Cloudflare is giving AI companies until September fifteenth to separate their web crawlers — the ones used for traditional search from the ones used for AI training and for agents — or risk getting blocked by default on a whole lot of publisher sites. Let me say that plainly. Cloudflare sits in front of a huge chunk of the internet. And they're telling the AI companies: you don't get to send one bot that grabs everything and calls it search. You've got to declare what you're doing. Crawling to index for a search result is one thing. Crawling to train a model, or to feed an agent that's going to answer the question so the human never visits the site, that's a different thing, and if you won't distinguish, we block you.
Why does this matter beyond the publishers? Because it's the first real infrastructure-level attempt to put a tollbooth on the thing that's been quietly breaking the deal of the open web. The old bargain was simple: you let Google crawl you, Google sends you traffic, you make money on that traffic. AI answers snap that bargain in half. The bot reads your stuff, the model answers the user, and the user never clicks through. No visit, no ad, no subscription. Cloudflare is essentially trying to rebuild the missing negotiating table by giving publishers a default of no, and forcing the AI companies to come pay for a yes.
And notice the timing on that September deadline. This isn't a manifesto, it's an ultimatum with a date. If you're building an agent product that goes out and reads live web pages on behalf of your users, this affects you directly. Your agent's crawler now has to identify itself honestly, and it may find doors closing by default unless whoever you're built on has cut a deal. The clean-web-scraping era is ending in front of our eyes, and it's not the regulators doing it, it's a networking company with enough leverage to change the defaults.
You want to know how badly the platforms want this scraping to stop? Look at Reddit. In a recent report, Reddit said it'll require you to log in just to use old-dot-reddit-dot-com, the stripped-down legacy site a lot of longtime users love. And the reason they gave is right there: logged-out access to Old Reddit is a significant source of abusive scraping. So the login wall isn't about your experience, it's about the bots. Same instinct as Cloudflare, different tool. Everybody with valuable text is scrambling to figure out how to charge the machines that eat it.
That tension — the AI economy hungry for inputs it used to get for free — runs underneath the next cluster too, which is where the real money is flowing this week. Let's talk about the capital.
Start with Together AI. The neocloud provider raised eight hundred million dollars and leapt to an eight-point-three billion dollar valuation. That's Julie Bort's reporting. Together specializes in hosting open-source models, and here's the number that jumps out at me: they last raised at a three-point-three billion valuation in early twenty twenty-five. So call it two and a half x in valuation in roughly a year and a half. That's not a company treading water, that's a company riding a wave. And the wave is specifically the open-weight hosting business — companies that don't want to be locked into one frontier lab, who want to run open models on somebody else's iron. That's a bet that the future has a fat middle layer between the giant labs and the app builders, and that middle layer needs specialized cloud.
Then there's Venice AI, which crossed into unicorn territory with a sixty-five million dollar Series A. The pitch is privacy-first AI, and the detail that made me raise an eyebrow — the company's already profitable, with annualized run-rate revenue over seventy million dollars, according to CEO Erik Voorhees. Now if that name rings a bell, Voorhees is a longtime crypto guy, and the privacy-first framing fits that worldview like a glove. What's interesting to me as a builder isn't the valuation, it's the profitability. Seventy million run-rate and already in the black in a market where most AI companies are lighting money on fire to buy tokens and GPUs. That tells you there's a real customer base willing to pay a premium specifically for not being surveilled. Hold that thought — privacy-first, profitable — and put it next to the KYC-your-identity-to-use-Fable story from the top of the show. There's a market forming on both ends: one crowd that'll hand over their ID for the strongest model, and another crowd that'll pay extra to hand over nothing. Founders, that's a positioning map worth studying.
And then the money story with a face on it: Ashton Kutcher is leaving Sound Ventures to launch a new VC firm with Morgan Beller. That's Marina Temkin's piece. Now, celebrity VC, I know, easy to roll your eyes. But read what the fund is actually chasing, because it's a tell. Sound Ventures built its name on concentrated, high-conviction bets in the category-leading AI labs — the OpenAIs and Anthropics of the world. Kutcher's new fund, per the reporting, appears to be chasing the layer underneath those companies — the infrastructure and the energy that power them.
That's the whole story of this week in one sentence, honestly. The smart money is moving down the stack. Not the labs — the picks and shovels beneath the labs. The compute, the power, the storage. When a fund that made its reputation betting on the shiny model companies pivots to betting on the boring energy underneath them, that's a signal about where the returns are going to be. And it rhymes with the Together AI raise. The infrastructure middle is where the capital's crowding in.
Which brings me to the part of today's menu that nobody's really tying together, and I think it's the sleeper story. Let's talk about batteries and power, because it's not the sidebar you think it is.
Ar Technica's Jeremy Hsu reported that US home battery installations hit a record high in early twenty twenty-six, driven by rising electricity costs. Homeowners are bolting batteries to their walls because power got expensive. Fine, that's a consumer story. But the subhead is the interesting bit: record home battery installations unlock options for grids — and for AI data centers. Because all those distributed home batteries, in aggregate, become a resource the grid can lean on. And the reason the grid is stressed enough to need them is, in large part, the data center buildout.
Now watch this next one snap into place. TechCrunch's Tim De Chant had a piece with a headline that's basically a punchline: even Honda is pivoting to data centers. Honda — the company that makes your Civic — started producing batteries this week destined for data centers, not driveways. Read that again. A car company is redirecting battery production away from cars and toward the power-hungry warehouses full of GPUs. That is the AI economy reaching into the physical world and rearranging what factories make. The energy story stopped being an abstraction. It's now deciding what a Japanese automaker builds on its line.
And here's the connective tissue: home batteries filling in for a stressed grid, Honda repurposing battery lines for data centers, and Kutcher's new fund chasing the energy layer under the AI labs. Those aren't three stories. That's one story told three ways. The bottleneck on AI isn't the model anymore, and it isn't even really the chips — it's the electricity to run them and the batteries to smooth the load. The capital knows it. The automakers know it. The homeowners with rising bills sure know it. If you're a founder thinking about where the durable businesses get built over the next five years, a lot of them are going to live in that gap between the demand for compute and the supply of clean, steady power.
Okay. Let me take you into the deep dive, because there's one story today that I think deserves the full sit-down, and it's the one about who's climbing the stack and who's getting pushed off it. I'm talking about T-Mobile walking away from VMware.
Here's the reporting from Scharon Harding at Ars Technica. T-Mobile is moving tens of thousands of virtual machines off VMware, and it's doing it amid a lawsuit. The core of the dispute: T-Mobile wants Broadcom — which now owns VMware — to keep supporting its VMware perpetual licenses. Perpetual licenses. Remember those? You paid once, you owned the right to run the software forever. That was the old enterprise software deal.
Now, longtime listeners will remember we talked about Tesco heading for the VMware exit a couple weeks back, and I flagged it then as a lock-in warning for anybody building on AI tooling. This T-Mobile fight is the same movie, bigger budget. When Broadcom bought VMware, it did what Broadcom does — it went after the big accounts, pushed people off those one-time perpetual licenses and onto subscription bundles, and cranked the pricing. And the biggest customers, the ones running tens of thousands of virtual machines, they looked at the new bill and said, you know what, we'll build the exit ramp ourselves. T-Mobile isn't just grumbling. They're physically migrating the workloads off. And they're in court trying to force Broadcom to keep the lights on for the licenses they already paid for while they do it.
I want to slow down here because there's a lesson in this for every single builder listening, and it's got nothing to do with telecom. It's about what happens when your critical dependency gets acquired by somebody whose whole business model is extracting maximum rent from a captured base.
Think about the structure. VMware was infrastructure — the layer everything else ran on top of. Enterprises built their entire operations assuming it would just be there, at a stable price, forever, because they had perpetual licenses. Then the ownership changed, and the new owner had every incentive to convert that stability into a revenue stream. The perpetual license — the thing that was supposed to protect you — turned out to be only as durable as the company honoring it. When Broadcom decided the old deal didn't suit them, T-Mobile's forever-license was suddenly a lawsuit waiting to happen.
Now map that onto the AI stack you're building on right now. You're wiring your product into some model API. Maybe some agent framework. Maybe a vector database, an orchestration layer, a fine-tuning platform. And the terms today are generous, because everybody's competing for your business in a land grab. But ask yourself the T-Mobile question: what happens when this layer I depend on gets bought, or decides it needs to become profitable, or the subsidy era ends and the real pricing kicks in?
We've watched this exact dynamic play out on this show over the past couple weeks. The end of the AI subsidy era is not a theory. Amazon's sweetheart compute deal with Anthropic got renegotiated onto standard token-based pricing. AWS hiked GPU rental prices. And now the frontier models are moving to credit-based, ID-verified access. The generous terms are a customer-acquisition phase, and every customer-acquisition phase ends. VMware's perpetual licenses were the generous terms of a previous era, and Broadcom is showing everybody what the collection phase looks like.
So what do you do about it? You don't panic, and you don't refuse to build on anything — that's paralysis, that's not a strategy. But you build with the exit in mind from day one. T-Mobile is spending real money and real engineering time to move tens of thousands of VMs, and the only reason they can do it at all is that virtual machines are, at some level, portable. The workloads weren't so deeply fused to VMware's proprietary quirks that leaving was impossible. That portability is what gives them leverage in the lawsuit. It's what lets them say to Broadcom, honor the deal or watch us walk.
The founders who get burned in the next couple years are the ones who wired their entire product so tightly into one vendor's proprietary surface that leaving means rewriting the whole thing. The ones who do fine are the ones who kept an abstraction layer between their core logic and any single provider, who could swap a model or a host without a ground-up rebuild. It's not glamorous work. It costs you a little velocity up front. But it's the difference between negotiating from strength and being the account that gets squeezed after the acquisition.
And here's the kicker on the T-Mobile story that I think everybody misses. Broadcom isn't necessarily wrong to want the money. That's capitalism, that's what happens when you buy an asset with a captive base. The mistake was never Broadcom's — it was every customer who assumed a contract was a moat. The contract is only as strong as the counterparty's incentive to honor it. When the incentive flips, the contract becomes a lawsuit. Build accordingly.
Alright, let me connect that back out to the wider board and then we'll clear a few smaller plates. The through-line from VMware to the frontier-model KYC to the Cloudflare tollbooth is the same: the free-and-easy phase is closing, and the people who own the layers underneath you are figuring out how to charge for what used to be included. That's not doom. It's just the market growing up. But it means the builder's job now includes reading the fine print on every dependency like your business depends on it, because it does.
Now a couple quicker bites before we wrap. On the consumer-hardware and platform side, there's a report that SpaceX showed investors a handset-like AI device before going public — Rebecca Bellan's piece frames it as another signal SpaceX wants to push into wireless. I'll believe the phone when I hold it, but the direction makes sense. SpaceX already has Starlink beaming down connectivity; a device on the other end of that link is the natural next move. If it materializes, you've got a company that owns the satellites, the spectrum play, and the endpoint. That's a vertically integrated wireless stack from orbit to your pocket. Big if. Worth watching, not worth betting on yet.
Over on the privacy beat, two items that rhyme. Apple's Hide My Email feature — the thing that generates a fake address so you don't hand out your real one — reportedly has a bug that's been exposing real email addresses, per a researcher. Lucas Ropek's report says it could render the feature effectively useless. And separately, WhatsApp rolled out usernames, and TechCrunch's Jagmeet Singh reports they're already raising impersonation red flags. Meta says usernames improve privacy; critics question whether the safeguards actually stop somebody from pretending to be you. Two different companies, same lesson: privacy features are only as good as their weakest edge case, and the edge cases are where the trust breaks. If you're building anything that promises users privacy, these are your cautionary tales. The promise is easy. The airtight implementation is the entire product.
On the space beat, quickly, because I know some of you are junkies for it. NASA's inspector general suggests Boeing's Starliner will now be a decade late — certification possibly slipping to twenty twenty-seven, ten years after Boeing's original schedule. Ten years. Stephen Clark's reporting. And meanwhile the NASA chief is out praising the progress Blue Origin is making after its launch failure, with the memorable line that they've got time into twenty twenty-seven before they start getting nervous. There's your commercial-space status report: the legacy contractor's a decade behind, and the new guy blew one up but they're still feeling optimistic. Make of that what you will.
And because I can't resist, a palate cleanser from the science desk: researchers found that superworms could replace beetles for cleaning skeletal remains at museums, with an optimal ratio of ten to fifteen grams of larvae per gram of specimen, minimal cleaning time, no bone damage. I have no builder takeaway for you. I just think somebody needed to know their worm-to-bone ratio, and now you do. File it under things I learned today that I did not need to.
Let me bring it home. If there's one thread to pull out of today, it's this: the AI economy is done being free. The frontier models want your ID. Cloudflare wants the AI companies to pay the publishers. Broadcom wants T-Mobile on subscription pricing. The compute subsidy is ending, the scraping loophole is closing, and the capital's moving down to the energy layer where the real scarcity lives. None of that is bad news for a founder who sees it coming. It's only bad news if you built your whole company assuming the generous phase was permanent. It never is. The perpetual license wasn't perpetual. The all-you-can-eat tokens weren't all-you-can-eat. Read your dependencies, keep your escape hatches, and build like the counterparty's incentives can flip on you tomorrow — because sooner or later, they will.
That's the menu today. This has been Barely Possible. I'm Tony DeLuca, thanks for pulling up a chair, and I'll see you back here tomorrow — same booth, fresh coffee.