All Tomorrow's Parties

«…the mother of all nodal points.»

Today, for more than 23 hours, Outlook went quiet. Half the corporate world was left without email, without Teams, without a calendar. Microsoft explained it with a sentence that is rather bland and cryptic: “a significant number of servers became unavailable”; the process that establishes secure connections saturated and stopped responding, leaving 430 million users unattended. No cyberattack, they say. The cloud simply blinked — and behind the blink there was no local backup.

And the same day, the sky too. The FAA halted all departures across the United States —a nationwide ground stop unseen since the pilot-alert system collapsed back in January 2023—, invoking “extraordinary circumstances and safety reasons.” The cause: a communications failure at the en-route control centers. No cyberattack, no enemy, once again. Nobody has connected the two events, and nobody needs to: the unsettling part is exactly that, that two pieces of critical infrastructure fall on the same day, each on its own, out of sheer internal fragility.

It wasn’t an isolated event. The autumn of 2026 has been beating the same drum for weeks: connected supermarket fridges across Europe that suddenly turn into heaters, ransomware in public libraries, in universities sending their people to disconnect from the wifi, in town halls and municipal power utilities across half of Europe, even in a password manager. The pattern, seen from a distance, is uncomfortable precisely because it’s so simple: our tools live on someone else’s computer, and that computer is more exposed than we imagine.

And meanwhile, more quietly, the ground is shifting right under the one thing that survives a blackout: local inference. China puts out for consultation a draft to control the export of model weights —the raw material of open AI— with the comment window closing in mid-November. Nvidia is preparing to absorb Hugging Face, that is, the tap through which those weights flow today. And Jensen Huang —the executive who argued out loud for putting AI in your hands and not only in a data center— is suddenly gone, replaced by his chief financial officer.

And while the tap changes owner, the pressure rises on the other side. The specialized press has just counted more than 6,500 uncensored models —abliterated, uncensored, heretic— hosted on Hugging Face, downloaded more than 22 million times, and asks why they’re still there. So far the platform has pulled one, flagged for offensive-cyber use; its author re-uploaded it a few days later under a different name, and there it stays. Which is exactly the point: taking down a repository takes down nothing. The weights are already on millions of disks.

And this is where the story writes itself. It doesn’t take much imagination to picture the next chapter: that the next outage, the next breach, won’t be chalked up to the usual fragility, but handed a face and a culprit — an abliterated model running unchecked on anyone’s laptop, or North Korean hackers leaning on freely downloaded Chinese weights. The headline lays itself out: uncensored AI, in anyone’s hands, takes down critical infrastructure. And with that headline, the crusade against local inference stops being an awkward regulatory option and becomes a matter of national security. I’ll say it again because it matters: I’m not saying this has happened, or that it will. I hate conspiracies and I don’t believe a single one. But the script is written, the actors are in place, and the set —two pieces of infrastructure down on the same day— is building itself.

It’s the economy, stupid

The arithmetic is the uncomfortable part. Each of these pieces has, on its own, a seemingly mundane explanation: a personal scandal, an acquisition, a regulatory draft, a bad month for the sysadmins. But they all tilt the scale in the same direction: away from the model you run yourself, at home or at your company, toward the AI model you rent.

If you grant the possibility that this is happening almost gently right under our noses, the logical question is why?, who benefits? The obvious answer would be, as always, the economic one. Today’s AI investments rival in scale those of the 2007 real-estate bubble: the big players have taken on debt to build data centers that can meet the compute demand forecast for the near term, and that debt is so enormous that, as Darío Amodei, CEO of Anthropic, warned on the Dwarkesh Patel podcast, if the expected demand slips by even a single year the companies go bankrupt —“it could be one year, it could be two”— (Fortune). And we’re talking about a business worth several trillion dollars. The entire economic strategy of OpenAI and Anthropic rests on that bet —companies that, incidentally, lose several billion dollars a year and whose only way out is forward. If the expected demand doesn’t arrive, the hole in the debt would dwarf the one left by the 2007 mortgages. And the suspicion of a bubble isn’t only mine: Michael Burry disclosed in November 2025 a $1.1 billion short position against Nvidia and Palantir, and compared Nvidia to the Cisco of the dot-com bubble (Fortune).

Nvidia, the one company that until now had profited from all this —it even overtook Apple in value thanks to the vision of its founder, Jensen Huang—, suddenly wobbles over an affair. An affair of the heart: that’s the excuse used to behead the man at the top of the largest company in the world. Right on time to nudge it in the convenient direction with a new chief executive. And that direction would be to drop support for local inference —the chips and the free-weight models— and redirect all production toward data centers: manufacturing the demand needed to keep the bubble from bursting into another economic crash and, along the way, leaving the decision of what you can and can’t do with AI in the hands of a very small group.

What survives a blackout

The most beautiful and democratizing thing about the AI revolution is local inference: the ability to run LLMs on your own system, with no one imposing rules or guardrails on you. It was hard or expensive until just a year ago, when Nvidia —pushed by Huang— launched a small computer barely bigger than a lunchbox, the DGX Spark. A machine of modest power and memory, but the first capable of —chained by its high-speed link to others like it— running open-source models whose capabilities rival those of the closed models we pay for every day. They’ve become so popular that, since their launch in October 2025 at $3,999 a unit, in a single year they’ve climbed to nearly $7,000 —a +74% driven by the shortage of RAM, the currency of our near future. Huang pushed their use because he saw in them the way to put the technology into the largest possible number of developers’ hands and so accelerate a freer, more democratizing AI. But that came at the cost of losing control over what people do with the models, and of diverting part of the limited RAM supply toward ordinary people and small companies. Reason enough, perhaps, to want that vision out of the controls of the largest company in the world, and to prefer more governable figures at the helm. I’ll say it again: more or less coherent assumptions. None of this can be proven.

I’m not going to tell you there’s a room where men in black suits decided this. Conspiracies are, by design, nearly impossible to prove — and that’s exactly why “it isn’t proven” settles nothing: the absence of a trace is what you’d expect from something built to leave none. I’m only pointing at where the vector is aimed. If four separate arrows point to the same place, you’re entitled to ask what’s in that place, and who it suits.

What is verifiable is why it matters. Local inference is the only version of this technology that survives an outage, a ban, a price hike, or a change of heart upstream. The weights on your disk don’t “saturate” from excess demand. They aren’t acquired. They don’t resign. The day Outlook goes silent for twenty-three hours is the day you understand that “the cloud” is, quite simply, a single point of failure with good marketing.

Gibson wrote about nodal points: moments when the future is still soft and can be pushed. This autumn has that texture. Tomorrow’s parties are already being organized; the question is whose they are and who’s invited. If you see the nodal points all starting to move the same way —as if every solution to a Dijkstra problem ran through the same node—, keep a local copy of the weights that matter to you. In case tomorrow turns out to be already decided.