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, you're tuned into Barely Possible, I'm your boy Tony DeLuca, and today we've got a menu where China's quietly redrawing the map of who's allowed to own a tech company, the IPO bell is still ringing in everybody's ears, and one of the smartest guys at Microsoft is telling you the thing you should actually be protecting isn't the model — it's the loop. Pull up a chair, grab the coffee, let's have at it.
Let me start where I think the real story is hiding, because it didn't get the headline it deserved. There's a current report out of TechCrunch — Kate Park has the byline — that Meta is moving to unwind its two billion dollar Manus acquisition after Beijing ordered the deal reversed. Now, on the surface that reads like one of those corporate-divorce stories you skim and forget. But sit with it for a second, because what's underneath is one of the more consequential shifts for anybody building a company with any cross-border ambition. And I want to be careful here — this connects to a thread we've been pulling at all week, so let me give you the full shape of it.
Here's the situation as it's been reported. Meta paid two billion for Manus right as 2025 rolled into 2026 — one of their flagship moves when they were resetting their whole AI strategy. Then in March, the Chinese government opened an investigation into the deal, and later barred the Manus founders from leaving the country. Now, Manus had tried to do the clever thing — they'd relocated operations to Singapore first, courting the foreign acquisition from outside the mainland, trying to thread the needle on China's tech export controls. Didn't matter. In April, Beijing ordered the deal unwound despite the workaround. And now Meta is doing the actual mechanical dismantling — firewalling operations, cutting Manus staff off from Meta's data systems, Meta staff can't use Manus tools internally anymore. A clean operational split, done under government order.
That's the news. Here's why it matters to you, and it's bigger than one acquisition. The maneuver Manus tried — set up shop in Singapore, then take foreign money — that wasn't some exotic dodge. That was the standard playbook. In China they even had a name for it, the red chip corporate structure. You incorporate offshore, you raise foreign capital, you give Western investors a clean entity to put money into. Tons of Chinese startups did it. And the message Beijing just sent, by cracking down on Manus, is that those days are over.
The knock-on effects are already showing up. Numerous prominent Chinese startups are reportedly looking to unwind their foreign corporate structures and reincorporate back home. StepFun has reportedly completed the process ahead of a Hong Kong IPO. The folks behind Kimi — Moonshot — and the team behind Kling are said to be considering the same. There's a Shanghai-based attorney quoted saying, more or less, that whether to dismantle the red chip structure is no longer the question — the only question is how to do it cheaply and efficiently. And the part that should make the hair on your arm stand up: reports that Chinese officials are now seizing passports from key researchers and executives at private firms. That was previously beyond the pale. Now it's a tool.
So what's the take-home for a founder? It's this. We spend all our time arguing about which model is best, which lab is ahead, who's got the better coding agent. And meanwhile, the actual ground under the whole global AI economy is being re-poured. Capital and talent — the two things you cannot build a company without — are both getting fenced in by national borders in a way they weren't eighteen months ago. If your strategy depended on Chinese talent reachable through a Singapore wrapper, or on Chinese AI companies being acquirable by Western firms, that strategy just aged a decade in a quarter.
And here's the uncomfortable mirror image, which ties right back to what we covered yesterday. Yesterday I walked you through the US Commerce Department ordering Anthropic to pull Fable 5 and Mythos 5 worldwide over a single alleged jailbreak — Friday at 5:21 in the evening, the kill switch came down. I'm not re-running that whole saga, you heard it. But look at the two stories side by side. Beijing is clawing its companies back inside its borders, seizing passports, unwinding offshore deals. Washington is reaching out and yanking a frontier model off the global market because a foreign national might touch it. Different governments, opposite directions, same underlying instinct: this technology is now a matter of state, and the state will decide who gets access to the talent, the capital, and the models. The borderless internet era of building — where you raised money anywhere, hired anywhere, sold anywhere — that's the thing actually ending. Not with a manifesto. With a 5pm Friday letter on one side and a seized passport on the other.
Which brings us neatly to the next chapter of the Anthropic fallout, because there's a genuinely new wrinkle worth covering. TechCrunch's Jagmeet Singh has a current piece on how, as Anthropic suspended access to those new models, India is now openly debating its AI future. And this is the downstream effect I want you to internalize. When the US flipped that switch, it wasn't just Anthropic's customers who felt it. Every government that watched it happen — and they all watched — got a free, fully-rendered demonstration of what dependence looks like. The piece frames the Indian tech leadership now asking out loud whether this is a wake-up call. Should India be building its own frontier capability rather than renting access to American models that can vanish on a Friday afternoon at the whim of a commerce secretary?
Now, I'm skeptical of sovereign-AI fever in general — it's easy to give a speech about it and very hard and expensive to actually do it. But the procurement logic is real and it's not going away. If you're a buyer of AI in Brussels or Tokyo or Sao Paulo or New Delhi, you now have a defensible, board-meeting-ready argument for hedging. For preferring a model you can't get cut off from. And the open-weight gap — the distance between the best American closed models and the strong open alternatives — has narrowed enough that hedging isn't a fantasy anymore. For builders, this is a market signal. The demand for AI you can actually own and run, that nobody can revoke, just got a giant unpaid advertising campaign courtesy of the US government. If that's the kind of product you build — on-prem, open-weight, sovereign-friendly — your sales calls just got easier. If your whole product sits downstream of a single American frontier API, you've got a new question on every enterprise security review, and you'd better have an answer.
Now let me shift from who controls the models to what you should actually be building on top of them, because there's a smart framing floating around that I think is worth your time. Swyx pulled out some comments from Satya Nadella about what he's calling loops as IP. And before your eyes glaze over at another executive philosophizing, hang with me, because the actual point is sharp.
Nadella's argument, roughly, is this. For the first time we can create a real cognitive loop between people and digital systems — a back-and-forth where the human and the AI system are working together, iterating, learning. And his line is that this changes how we even conceptualize work inside an enterprise. Then the part that matters for builders: the real opportunity, he says, is not in picking the best model. It's in building a learning loop on top of models, where — and this is his phrase — human capital and token capital compound. You can offload a task, you can even offload a whole job, but you can never offload your learning. And his closing move is that the priority should be building a frontier ecosystem, not just a frontier model, so that every organization can own the learning loop that encodes its institutional knowledge.
Now, I'm allergic to executive poetry as much as the next guy from the Bronx. But strip the gloss off and there's a genuinely useful idea in there for anybody running a company. The model is rented. Everybody can rent the same model. What you can actually own is the loop — the accumulated record of how your business uses AI, what works, the corrections your people make, the institutional knowledge that gets baked into how the system behaves over time. That's the asset. That's the thing a competitor can't just buy a subscription to.
And notice how clean this maps onto everything else on today's menu. If models can be cut off by governments — China clawing back, the US switching off — then the smartest defensive posture for a business is to make sure your value lives in the loop, not in the specific model serving it. The model becomes a swappable component. Your learning loop is the durable thing. When Anthropic goes dark and you have to migrate to GPT-5.5 or an open-weight alternative on Tuesday, the businesses that survive that cleanly are the ones whose IP was in the loop, not in the vendor. So Nadella's framing isn't just a Microsoft sales pitch — though it's also that. It's a survival strategy for a world where access to any single model is now politically contingent.
I'll add one note of skepticism so I'm not just nodding along. The hard part nobody's solved is actually capturing that loop in a way that compounds. Most enterprises right now have logs that look busy and a process that hasn't changed. You can have a million AI interactions a day and still not be encoding anything durable, because nobody's set up the feedback. The loop only becomes IP if you build the machinery to learn from it. Otherwise it's just expensive autocomplete with a transcript. So take Nadella's point as a goal, not an achievement. The companies that figure out the encoding part win. The ones that just rack up token spend and call it transformation are the ones who'll be most exposed when the rented model gets yanked.
Let me take a breath from the heavy stuff and give you something concrete for the toolbox, because Simon Willison flagged something this week that's quietly a big deal for anybody building in Python. It's now possible to compile Python extensions — your C, your C++, your Rust — to WebAssembly and distribute them through PyPI in a way that Pyodide can install directly. They're calling these WASM wheels.
Now I'm not going to drag you through the internals, that's not what this show is for. Here's the why-you-care version. For years, the friction with running Python in the browser — Pyodide being the project that makes that possible — was that all the good libraries, the fast ones, the ones with native C or Rust under the hood, were a pain to get working. You could run pure Python fine, but the moment you needed something with a compiled extension, you hit a wall. This closes a chunk of that wall. It means more of the real Python ecosystem can run client-side, in the browser, without a server. For a builder that's interesting because it pushes more computation to the edge, to the user's machine — cheaper for you, more private for them, and increasingly relevant when, as we've been saying all episode, you might not want to depend on a server somewhere that a government can lean on. I'll put the link in the show notes. It's a plumbing story, but good plumbing is how you avoid the floods.
Now let's talk money, because the IPO wave is still the loudest thing in the room and there's a fresh angle on it. We covered the SpaceX debut itself earlier in the week — priced at a hundred thirty-five a share, largest debut ever, popped nineteen percent on day one, minted the world's first trillionaire. I'm not re-running the debut. But there's a current TechCrunch piece by Anthony Ha asking the follow-on question that actually matters for founders: as AI companies race to go public, who else is along for the ride? The framing in the piece is startups trying to, quote, ride that SpaceX IPO wave.
And this is where I want to inject some of that protective skepticism, because this is exactly the kind of moment that separates the people who keep their money from the people who donate it. When a giant marquee IPO pops, every banker in town starts whispering to every growth-stage company that the window is open, the appetite is back, get in line. And some of that's real — a successful debut genuinely does reopen the market for the names behind it. But a lot of it is the halo effect, and the halo around this particular debut is almost entirely one specific guy and one specific story about data centers in space. You cannot generalize Elon to anybody else. There is no other founder operating in that gravitational field.
There's a related current item — Kirsten Korosec's TechCrunch Mobility piece, headlined SpaceX rockets past Tesla — that captures the strange new pecking order, where the rocket-and-infrastructure company has leapfrogged the car company in the market's imagination. And again, the lesson for you isn't to chase the wave. It's to understand what the market is actually pricing. It's not pricing AI models. It rebranded SpaceX as a neocloud — an infrastructure and compute story with a heavy dose of Elon on top — and that's a very different bet than the frontier-model IPOs everyone assumes are coming next. So if you're a founder sitting there wondering whether your AI startup can ride this, ask the cold question: is the market buying durable infrastructure revenue, or is it buying a personality? Because if your pitch is the second thing and you're not the guy, the window may not be as open as your banker says.
And here's a cautionary tale stapled right to the back of all this IPO froth. There's a current report — Connie Loizos at TechCrunch — that startup CEO Charlie Javice is reportedly angling for a Trump pardon. For those who don't have the file: Javice is the founder whose company JPMorgan bought, and the whole thing collapsed into one of the more notorious fraud cases of the era. JPMorgan, as the piece dryly notes, cannot be pleased by any of this. I bring it up not for the gossip but for the through-line. In the same week the market is throwing a hundred billion dollars of retail orders at a single IPO, we've got a reminder of what happens at the other end of the founder-celebrity machine — the inflated metrics, the acquisition, the unraveling, and now the angle for a pardon. The exuberance and the wreckage are always the same coin. When everybody's telling you it's the easiest money in years, that's exactly when you keep your hand on your wallet.
Let me change channels entirely, because not everything today is about borders and bubbles. Ars Technica's got a couple of pieces that scratch a different itch, and I think you've earned a palate cleanser.
First, a current story from Jennifer Ouellette asking whether a medieval flying monk spotted Halley's comet — twice. The flying monk in question is Eilmer of Malmesbury, an eleventh-century fella famous for, allegedly, strapping on wings and jumping off a tower, which is its own story. A University of Leicester historian now thinks Eilmer saw two different comets — one in 1018 and one in 1066 — and the headline gets at the it's complicated of it, because matching medieval comet sightings to what we now know as Halley's is a genuinely thorny bit of detective work across a thousand years of fuzzy records. I love this stuff because it's the long-memory version of what we do every day on this show — somebody dug into the historical record and reframed what we thought we knew. The link's in the notes if you want the full read.
And second, a current piece — Wyatt Myskow, Inside Climate News, running on Ars — that researchers have now quantified the underground fungal networks beneath our feet, and the threads are long enough, end to end, to reach beyond the solar system. The mycorrhizal fungi, the stuff that webs through soil and trades nutrients with plant roots. They put numbers on the global length and mass of it for the first time. I'll just say it: there is a poetry in the fact that the most consequential network on Earth might be the one nobody can see, nobody owns, and no government can switch off. After an episode about kill switches and seized passports, a network that just quietly holds the whole biosphere together is a nice thing to chew on. Link's in the notes.
Okay, let me bring it back home and tie the threads together before I let you go, because I don't want you to leave with five separate stories. I want you to leave with one idea.
Everything heavy today was a variation on the same theme: the things you used to be able to take for granted as a builder — that you could raise capital across borders, hire talent across borders, and rent the best model from anywhere and never lose it — those assumptions are all getting renegotiated at the same time, by governments, in real time. Beijing's clawing its companies and its people back home. Washington's demonstrating it'll yank a frontier model off the world market on a Friday whim. And every other government watched both happen and is now drawing its own conclusions, India out loud among them.
And against all that, the most useful piece of advice came from, of all places, an executive comment: build so that your value lives in the loop, not in the vendor. Own the learning. Make the model a swappable part. Because in a world where access to any single model is now a political decision made by people you've never met, the only durable moat is the accumulated, encoded knowledge of how your business actually works — the thing nobody can buy a subscription to and nobody can switch off with a letter. The WASM wheels story is a small piece of the same instinct: push more of your stack to where you control it. The IPO froth is the counter-temptation — the siren song telling you the easy path is to ride somebody else's wave. Don't. Build the loop.
That's the menu, kiddos. Watch how the other governments move now that they've seen the playbook, watch whether that sovereign-AI talk turns into actual budgets, and watch where you're keeping your own institutional knowledge. This has been Barely Possible. I'm Tony DeLuca — be skeptical, be generous, and I'll see you tomorrow.