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 friends, it's your boy Tony DeLuca, and welcome back to Barely Possible, where we sort the real signal from the noise so you don't have to sit through forty browser tabs to know what actually changed today. Pull up a chair. We got a fresh tray of stories, and a couple of them are the kind that make you put down the coffee and squint.
We got SpaceX writing a sixty-billion-dollar check for a coding tool. We got leaked financials showing one of the biggest names in AI bleeding cash. We got Anthropic quietly walking back a billing change that had power users ready to riot. And we got Qualcomm betting the next computer you carry won't be a phone at all. Plenty to chew on. Let's have at it.
Let me start with the one that's been the talk of the building, because it's wild even by the standards of this industry. SpaceX is acquiring the AI coding platform Cursor for sixty billion dollars. Sixty billion. For a code editor. Now, I want you to sit with the weirdness of that for a second, because a few years ago the sentence "the rocket company bought the text editor" would have gotten you laughed out of the diner. But here we are.
The framing on this one is sharp. The piece on it had a line that just nailed the logic: separately, neither could compete. Now they hope they can. And that's really the whole thing in one breath. Cursor by itself is a fantastic product sitting on top of other people's models — Anthropic's, OpenAI's. It's a harness. A really good harness, but a harness. And a harness that depends on the labs it competes with is in a precarious spot. We've talked about that dynamic on this show plenty — the model is the engine, the harness is the car, and if the engine maker decides to build their own car and jack your fuel prices, you've got a problem.
So what does SpaceX bring? Capital, obviously. Compute, obviously — and we'll get to SpaceX's absolutely deranged valuation in a minute, because it's part of this story. But more than that, this is about somebody with deep pockets and their own compute ambitions deciding they want to own a full stack in coding, head-to-head against Anthropic and OpenAI. Take a strong front-end developer tool, marry it to serious infrastructure, and try to stop renting from the people you're trying to beat. That's the bet.
Whether it works is a totally separate question. Sixty billion is a number that prices in a lot of things going right. But for you, the builder, here's the takeaway: the coding-tool layer is consolidating fast, and it's consolidating up into the giants. We covered Salesforce buying the customer-service agent platform Fin for three-point-six billion just yesterday — that was the enterprise-agent version of this same move. Today it's the developer-tool version, at a number twenty times bigger. The independent middle of this market is getting bought, squeezed, or both. If you're building a thin wrapper on somebody else's model and somebody else's distribution, you should be asking very hard questions about who acquires you and at what point you stop being a company and start being a feature.
And here's a little wrinkle that I genuinely enjoyed. On the very first day this combined SpaceX-AI-Cursor thing exists, the Cursor and Graphite folks — Tomas Reimers in particular — went and announced a new product called Origin. It's their long-awaited Git competitor, built for agent workloads. Scalable for agents, extensible with API and MCP, and — this is the part that matters — built-in merge-conflict and co-failure agent resolution. Meaning: when you've got a fleet of coding agents all hammering on the same codebase, they're going to collide. They're going to step on each other's commits. Origin is trying to be the version-control layer that assumes the committers are robots, not people.
Now swyx pointed out the gentle irony here, which I have to share because it's just good: on day one of being SpaceX-AI-Cursor, they ship a product called Origin — a blue Origin — right after Jeff Bezos announces something called Prometheus. So you've got Musk's world shipping "Origin" and Bezos shipping "Prometheus," and the naming wars are now leaking from rockets into developer tools. Make of that what you will. But the substance underneath the joke is real: Git was built for humans typing. The agent era needs a version-control story built for swarms, and a bunch of serious people now think that's a real product category. Watch that space.
So that's the buyer. Now let's talk about who's bleeding, because the same morning we get the sixty-billion-dollar coding bet, we also get leaked financial documents showing OpenAI is losing billions of dollars a year.
The gist from the report: this is audited accounting, not back-of-napkin speculation. Revenue is growing — genuinely growing, and fast — but it's getting absolutely dwarfed by R&D and other expenses. So you've got a company whose top line is going up and to the right, and whose bottom line is deep, deep red.
Now, I want to be careful and grown-up about this, because the knee-jerk take is "see, it's all a bubble, the emperor has no clothes." And that's lazy. Losing billions while revenue explodes is, on its own, not damning. Amazon did it for years on purpose. The real question is always the shape of the curve and the cost of the inputs. Are the losses the cost of acquiring a position you can later monetize, or are they structural — meaning every new customer makes the hole deeper?
And here's where it connects to everything else today. The defining tension in AI right now is between what the labs need and what their customers will actually pay. The labs need token consumption to keep exploding, quarter after quarter, because that's the contract that justifies the data centers and the compute. But on the other side, the CFOs are showing up. We've talked about Uber capping its people at fifteen hundred a month. The enterprise buyers are starting to ration. So you've got a company torching billions to fund the frontier, needing growth to keep flowing, while the buyers are getting more disciplined about the bill. That's a vice. And leaked numbers showing the size of the losses just tightens it.
For a founder, the lesson isn't schadenfreude. It's this: the economics of the model layer are not settled. The price you pay for tokens today is not necessarily the price you'll pay in eighteen months, in either direction. Build like your input costs are volatile, because they are.
Which brings me right into the next one, because it's the same theme with a face on it. Anthropic paused token-based billing for its Claude Agent SDK.
Here's what happened. Anthropic had a change planned for Monday that would have moved a bunch of usage onto token-based billing — and for power users, the heavy hitters running agents all day, that would have meant a serious jump in cost. And the community noticed. And the community got loud. And Anthropic put the word "pauses" in quotes and walked it back, at least for now.
This is the token-scarcity story playing out in real time, in slow motion, in public. The labs spent a couple of years effectively subsidizing heavy usage to get everybody hooked. Estimates floating around suggest that on a two-hundred-dollar-a-month Claude plan, the max possible token spend was something like eight thousand dollars' worth. Eight grand of value for two hundred bucks. That math does not survive contact with a public company. So the labs are slowly, carefully trying to claw the subsidy back without setting off a revolt. And every time they reach for the lever, the power users push back, and they pause.
Simon Willison had a useful clarification in the middle of all the noise this week. People were tossing around a thirteen-dollar-a-day number, and Simon went and found that it's actually a page in the Claude docs advising enterprise customers how much to budget when they're paying full API price per team member. And his bigger point, which is the one to hold onto: enterprise companies with more than a hundred and fifty employees have to pay full API token pricing for their usage, on top of a per-seat charge of around twenty bucks a month. And as far as Simon can tell, those enterprise accounts are where Anthropic makes the bulk of its revenue.
That last bit is the whole ballgame. The consumer power users screaming about billing changes are loud, but they're not the cash engine. The cash engine is enterprises paying full freight at the API meter. So when Anthropic "pauses" a billing change, read it as a company managing the optics with the noisy crowd while the quiet, profitable crowd keeps paying by the token regardless. If you're building on Claude, the strategic read is: the seat-and-subscription comfort is for show. The real relationship is metered, and metered means you are exposed to every pricing decision they make. Architect for that. Route cheaper models where you can. Don't let your unit economics depend on a subsidy that's visibly being dismantled.
Alright. Let me shift from the money to the metal, because there's a genuinely interesting platform bet today, and it's our deep dive.
Qualcomm wants to be the chip inside whatever replaces your smartphone. And they just announced two products to start chasing that.
Here's the meat of it, straight from the reporting. Qualcomm's CEO, Cristiano Amon, said the company is working on more than forty different AI wearable devices. Not forty SKUs of the same earbud — forty different kinds of thing. Jewelry. Earbuds with cameras in them. Pins. Watches. The quote that frames the whole strategy is that this is, and I'm reading it, "a sign of how aggressively the chipmaker is betting that the next major computing platform won't be a phone."
Let that sit. Qualcomm built an enormous business being the brains inside Android phones. And here's their CEO, on the record, saying the next big platform won't be a phone. That's a company looking at its own franchise and openly preparing to outlive it. That's not nothing.
Now, why does this matter to you, a person who builds products and companies and not a person who collects gadget reviews? Because platform transitions are where fortunes get made and lost, and they don't announce themselves with a parade. The shift from desktop to mobile minted a generation of companies — and buried a generation that didn't believe it was happening. Qualcomm placing forty bets on form factors is them saying, out loud, with their R&D budget, that they think we're early in one of those transitions right now. AI changes what an interface needs to be. If the assistant can see what you see and hear what you hear and answer in your ear, the rectangle of glass in your pocket starts to look optional.
And I want to connect this to a story that's been hanging around the edges all week, because it's the same bet from a different angle. Snap finally shipped its long-awaited AR glasses — they're calling them Specs. Over a decade of work on this device, and they're finally out in the world. And the verdict in the coverage was, and I'm quoting the headline energy here, "oof, they aren't cheap." So you've got Snap, a company that has been grinding on face computers for ten-plus years, finally putting real hardware on the table — and the immediate reaction is sticker shock.
Put those two stories next to each other and you get the honest picture of the post-phone bet. The ambition is enormous. Qualcomm wants forty form factors. Snap wants the glasses to be the thing. And the reality on the ground is: it's expensive, it's early, and nobody's nailed it yet. This is the messy front edge of a platform shift, where the conviction is high and the products are still a little awkward and pricey.
So here's how I'd file it if I'm building. First, the picks-and-shovels read. Qualcomm isn't betting on which gadget wins. They're betting that they're the silicon inside whatever does. That's a smart posture in an uncertain transition — you don't have to pick the winning horse if you sell the saddles. If you're a founder, ask yourself whether your bet on the AI-hardware future requires you to be right about the exact device, or whether you can position one layer down or one layer sideways and win regardless of which form factor takes off.
Second, the timing read. Forty experiments and a pair of expensive glasses tells you we are not at the iPhone moment yet. We're at the moment before the moment — lots of attempts, no clear category killer, prices that scare normal people. If you're building software for AI wearables today, you are early. Early can be the best place to be, or it can be the place where you run out of money waiting for the platform to show up. Know which one you're signing up for.
And third — the thing nobody says out loud at the keynote. Cameras in earbuds. Cameras in jewelry. Cameras in pins. Glasses that see what you see. Every one of these form factors is, functionally, a sensor pointed at the world and at the people around you. The phone at least had the decency to be a slab you had to lift up and aim. The whole pitch of ambient AI hardware is that it's always on, always listening, always looking. That is a privacy and trust problem the size of a building, and it is going to be the thing that decides whether any of this crosses from gadget-blog curiosity into mass adoption. The company that wins the post-phone era won't be the one with the best chip or the cheapest glasses. It'll be the one people actually trust to wear on their face all day. Hold that thought, because it's going to come back around in a year or two and bite somebody hard.
Okay, let me come up for air and run through a few things that don't need the full treatment but you should have on your radar.
Let's talk about that SpaceX valuation, because it's the backdrop to the Cursor deal and it is genuinely something to behold. SpaceX's valuation ballooned to two-point-six trillion dollars and briefly passed Amazon. The valuation went up by a full trillion dollars since the shares started trading. A trillion. In days. And the IPO itself, by the numbers, grew to eighty-five-point-seven billion dollars raised, after the underwriters maxed out their share purchases — already a historic amount. So when SpaceX turns around and writes a sixty-billion-dollar check for a coding tool, understand the context: this is a company sitting on a fresh mountain of capital and a market that's handing it a trillion dollars of paper value for breakfast. That's the war chest behind the Cursor move. Whether two-point-six trillion is a real number or a momentum number, we'll find out the hard way at some point. But for now it's the engine.
Staying on cars and autonomy for a second — two stories rhyme here. Mobileye is entering the US robotaxi market with its own standalone service, leveraging its Moovit platform, launching in a US city in 2027. And the interesting angle, per the reporting, is that this puts Mobileye on both sides of the AV business. They sell their self-driving system to other companies — and now they're going to run a robotaxi service that competes directly with those same customers. That's a tightrope. You're a supplier and a competitor at once. We saw the same logic in the Cursor story — own the stack, even if it means competing with your partners. It's the same instinct showing up everywhere right now: don't just supply the layer, own it, even at the cost of the relationships that got you there.
Meanwhile, Rivian cut hundreds of workers right after R2 deliveries started. The company calls it restructuring to scale toward profitability, and notably, Rivian recently pushed its profitability goal back to invest more in autonomy. So that's the trade they're making — trim the workforce, redirect the money into self-driving. Layoffs the same week your big new vehicle starts shipping is a rough optic, but it tells you where they think the value is: not in building more of today's car, but in the autonomy bet.
On the Google side, Android 17 and Wear OS 7 are rolling out, hitting Pixel phones and watches, bundled with a Pixel Drop that brings Google's latest AI models onto the devices — new multitasking tools, parental controls, security features, and an expansion of Gemini features. The honest framing in the coverage was "don't expect monumental changes." So it's an incremental release with Gemini pushed deeper into the OS. The thing worth noting for builders is just the steady drumbeat: Google keeps welding its models tighter into the operating system layer. If your product competes with a default assistant, the default keeps getting more capable and more deeply wired in. That's a slow-moving squeeze, not a headline, but it's real.
Now here's a survey that ought to land on the desk of every marketing person at every AI company. Sixty percent of US consumers say the word "AI" in brand messaging is a turnoff. That's from a WordPress VIP survey. Sixty percent. And the kicker is that this is happening even as companies increasingly view AI search as a critical referral channel. So you've got this split-brain situation: businesses are racing to optimize for AI-powered search and slapping "AI" on everything, while a majority of actual humans see "AI" in your messaging and recoil a little. For founders, that's a genuinely useful, slightly counterintuitive data point. The technology can be your whole backend and your whole advantage — but shouting "AI" in your customer-facing copy may be actively costing you. Sell the outcome, not the acronym. Your buyers may want what AI does and not want to be told it's AI doing it.
A couple of security ones quickly, because they're the kind of thing that's boring until it's your incident. A security researcher found a bug in FIFA's World Cup internal systems that, per the reporting, would have let her access several internal platforms — including one that could have let her take control of the TV stream of every single World Cup match. Read that twice. The TV feed. Of every match. At the World Cup. Found through a flaw in FIFA's online platforms. It's a reminder that the highest-stakes, highest-visibility events on the planet are still running on internal tooling with holes in it, and the only thing between a clean broadcast and chaos is whether a good-faith researcher finds it before a bad one does.
And on the agent-security front, there's a steady stream of work this week on how AI agents get hijacked through their own tool pipelines — prompt-injection-style attacks that ride in through the very tools the agent is supposed to use. I'm not going to go deep into the mechanics, but the headline for builders is simple and unglamorous: every tool you give an agent is an attack surface. Every MCP connection, every plugin, every "let the agent read this document" is a door. As you wire agents into more of your stack, you are not just adding capability, you're adding ways in. Treat agent tool access the way you'd treat handing out production credentials, because functionally, that's what it is. There's a CrowdStrike push around continuous identity for AI agents that's chasing exactly this problem — the idea that an agent needs an identity you can verify and revoke, not just an API key it carries around forever. The fact that the security vendors are productizing this tells you it's becoming a real budget line. Show notes will have the links.
Let me hit a couple in the government-and-policy lane and then we'll land the plane.
The Pentagon is out boasting about using AI to write reports that Congress mandated — and claiming that one-point-five million personnel are now using generative AI tools. Now, there's a comedy layer here: the reports Congress asked for, written by the machine. But the number is the real story. A million and a half people inside the Defense Department using these tools is enormous-scale adoption inside one of the largest organizations on Earth. Whatever you think about AI-written compliance reports, that's a procurement and deployment fact that reshapes a giant customer. If you sell to government, the door isn't just open, it's been kicked off the hinges.
And a media-and-antitrust one that's worth a raised eyebrow: the US approval of the Paramount–Warner Bros. deal — a hundred and eleven billion dollars — reportedly surprised the DOJ's own lawyers. Senator Warren said the green light "reeks of corruption." I'm not here to adjudicate that. But when a deal that size sails through and the antitrust lawyers inside the building are caught off guard, that tells you something about how predictable — or unpredictable — the regulatory environment is right now for big consolidation. And given how much consolidation we're watching in AI and tech this very week — SpaceX buying Cursor, Salesforce buying Fin — the question of who's actually minding the antitrust store is not academic for anybody planning a big merger or hoping to get bought.
One more, internationally: India ordered a temporary nationwide ban on Telegram over exam-fraud concerns, running until June 22nd, plus a requirement to disable the app's message-editing feature. That message-editing detail is the interesting part — the concern is people coordinating cheating and then scrubbing the evidence by editing messages after the fact. It's a small window into how governments are starting to reach into specific product features, not just whole apps. "Turn off this one capability" is a different kind of regulatory ask than "block the app," and it's the kind of granular demand more companies are going to face.
So let me tie the bow on this one, because there's a thread running through today whether anybody planned it or not. SpaceX paying sixty billion for a coding tool. Mobileye becoming a competitor to its own customers. The labs clawing back subsidies one nervous pause at a time. Qualcomm betting on forty form factors at once. Every one of these is somebody refusing to stay in their lane — the supplier wants to be the operator, the harness wants to be the stack, the chipmaker wants to outlive the device it was built for. The comfortable middle of every one of these markets is getting eaten from above and below. If you're building right now, the strategic question isn't "what's my product." It's "am I a layer somebody bigger is about to absorb, and what do I own that they can't just buy or rebuild." That's the game on the board today.
That's the menu, kiddos. Keep your unit economics honest, treat your agent's tools like live wires, and don't put the word AI on the billboard if your customers are flinching at it. I'm Tony DeLuca, this has been Barely Possible, and I'll be right here tomorrow with whatever the day cooks up. Be good to each other.