I Bought the Airline

THE NUMBER: 975B – 1GW – 0. Nine hundred seventy-five billion: the parameters in the open, free, self-hostable model Mira Murati shipped this week to run on your own hardware. One gigawatt: the mobile power Elon Musk bought in the dark, because he needs electricity now and the grid says wait five years. Zero: the new data centers New York will permit under the first outright ban in the country. Build, build, ban. Two of those are doers adding capacity the market pays for. The third is a government subtracting it and calling the subtraction a virtue. The entire issue is learning to tell the two apart on sight.

There’s a scene in Inception that everybody remembers for the wrong reason. They remember the folding city, the van off the bridge, the spinning top. They forget the airline.

Cobb needs ten hours alone with a mark named Fischer, uninterrupted, on a flight from Sydney to Los Angeles. It’s a logistics problem — you’d need to control the cabin, keep the flight attendants away, make sure nobody in the next seat leans over at the wrong moment. Saito, the Ken Watanabe character, the money in the operation, solves it between sentences. Cobb asks if he can arrange it. Saito says he already has. “I bought the airline. It seemed neater.”

Not a seat. Not the row. Not a quiet word with the gate agent about an upgrade. The airline. When the constraint is the airplane, the doer doesn’t negotiate with the airplane. He buys the thing that owns the airplane and walks onto the flight while everyone else is still arguing with the kiosk.

Hold that scene, because this week three different people ran straight at three different constraints, and the way each one moved told you exactly what kind of actor they are. Two of them bought the airline. One of them stood at the gate and banned flying.

And the backdrop makes it sharper. On Tuesday, as we wrote in Ice the Kicker, Demis Hassabis and Dario Amodei were in Washington asking the government to build a referee for frontier AI — a thirty-day pre-release gate, a standards body, a whistle. Forty-eight hours later, the people who actually move this industry didn’t file a comment with the proposed referee. They shipped a model, bought a power company, and in one case threw a wooden shoe into the machine. The contrast is the whole story.

✈️ “It Seemed Neater”

Start with Elon, because his move is the purest expression of the principle.

Sometime in the last stretch, Musk personally bought APR Energy — a company that operates over a gigawatt of mobile gas and diesel turbines, the kind that arrive on trucks and can be delivered, installed and commissioned in about a month. Built for blackout zones and disaster response, grids you can’t count on. There was no press release. No tweet, which for this man is its own kind of tell. The billion-dollar deal only surfaced because a firm holding a 5% stake had to disclose a $50.4 million payout in an SEC filing. He bought a power company the way you’d buy a sandwich, and didn’t mention it.

Here’s the constraint that makes a turbine fleet worth a billion dollars to one man. Nvidia can deliver 100,000 GPUs in a matter of months. A new grid interconnection, the permission to plug a large new load into the power system, sits a median of roughly five years in the queue. Read those two numbers next to each other and the whole AI buildout snaps into focus. The chips are fast. The wire is slow. And the chips depreciate the entire time the paperwork sits, which is not a small thing when the same week’s news had the whole industry’s $725 billion capex binge just barely clearing its own depreciation — $25 billion of quarterly revenue against $21 billion of depreciation, with Michael Burry out there calling understated depreciation “one of the most common frauds of the modern era.” Every month a rack of GPUs waits on a substation, it’s not earning. It’s aging.

So Musk did the Saito thing. He already lived this problem once — xAI’s first Memphis cluster ran 100,000 GPUs on about 150 megawatts, much of it from roughly 35 leased mobile turbines, because he wasn’t going to wait in the interconnection line and neither was the model. Now he’s not leasing the turbines. He owns the company that makes them. When the constraint is power, you don’t negotiate with the power company. You buy one.

And he’s not alone in seeing it, which is how you know it’s real and not just Elon being Elon. The same week, TeraWulf signed Anthropic to a twenty-year, $19 billion lease for 401 megawatts at a campus in Hawesville, Kentucky, that used to be an aluminum smelter. Why a smelter? Because the smelter already had five independent high-voltage transmission lines and drew 480 megawatts continuously for forty years. The power was already there. TeraWulf isn’t building a data center so much as buying a hookup that exists and pouring compute into it. Same move at Lake Mariner, a retired coal plant in New York now becoming a 750-megawatt AI campus. The doers have all figured out the same thing: in 2026 you don’t buy electricity, you buy time-to-power, and the cheapest time-to-power on Earth is a dead industrial site with the wire still in the ground.

That’s what doing looks like. It’s not loud. It showed up in an SEC filing.

And if you want to know where this goes next, watch who owns a turbine that isn’t in the ground yet. Boom Supersonic spent years and United’s money building Symphony, its own engine, because Rolls-Royce and GE and Honeywell all passed. If that tech is real and quick to deploy, its fastest path to revenue may not be under a wing at all. A power turbine needs an emissions permit. A passenger jet engine needs years of airworthiness certification, which is the exact five-year line Musk just paid a billion dollars to skip. Aeroderivative turbines — the GE LM-series that already peak our grids — are just jet engines brought down to earth. So don’t be shocked if the next airline somebody buys is a literal airplane company, and they buy it for the engine, not the wings.

🔓 The Escape Hatch

Now Murati, whose move is quieter than Elon’s and maybe smarter.

Thinking Machines shipped Inkling this week: 975 billion parameters, 41 billion active, trained on 45 trillion tokens of text, image, audio and video, a million-token context window. Open weights. Apache 2.0. Free. You can download it and run it on your own hardware tonight.

You have to understand where this company was six months ago to see what just happened. It looked troubled. The rumored $50 billion follow-on round collapsed in January. Two cofounders went back to OpenAI. Zuckerberg tried to buy the whole thing, got told no, and poached a third cofounder as a consolation prize. The standard playbook from there is obvious and it’s a death march: grind toward some benchmark win, put out a press release, try to reopen the fundraise before the money runs out. Chase the frontier and hope.

Murati went the other way entirely. She took the largest seed round in history — $2 billion at a $12 billion valuation — added an Nvidia investment and a gigawatt of Vera Rubin compute in March, cut a Google deal in April, and then instead of selling access to a frontier model, she gave one away. And here’s the part the herd will misread: the launch post openly admits Inkling trails the frontier. It’s not the best model in the world. It was never trying to be.

That admission is the strategy, not an apology for it. Because the thing Inkling is best in the world at is a thing the frontier labs structurally cannot offer: it’s yours. It runs on your iron, behind your firewall, and it never phones home. Four days ago, in The Man Behind the Curtain, we laid out the number that governs all of this — 89 cents of every enterprise AI dollar still flows to the closed frontier, up from 81% a year ago, and it’s not because the frontier is winning the argument. It’s because companies want to leave and can’t. They don’t want to hand a lab that might compete with them tomorrow a clean look at their own alpha. Most of them can’t build the routing console that Coinbase and DoorDash built to escape. They’re stuck talking to the head because the head, at least, answers.

Murati just mailed all of them the escape hatch. Free. A frontier-class model that’s a little worse than the best one, in exchange for never leaking another proprietary token to a competitor. For a bank, a law firm, a manufacturer sitting on decades of data that is the entire business, that is not a worse deal than renting Claude. It might be the only deal that ever made sense. She didn’t beat the frontier. She made the frontier’s biggest weakness — you have to send it your secrets — into her whole pitch.

And she is very much not alone, which is the part that should end the freeze conversation before it starts. The same week Inkling dropped, Moonshot’s Kimi K3 — 2.8 trillion parameters, a million-token context, open weights, out of China — took the number one spot on the Frontend Code Arena. Meta finally started charging for Muse Spark and Zuckerberg broke a three-year silence on X to do it, pricing it, in Alexander Wang’s words, “very aggressive” against OpenAI and Anthropic. Nvidia shipped a new embedding model that topped its benchmark. Four frontier-class releases in one news cycle, half of them open, one of them free.

This is the fact that makes Tuesday’s referee look silly. You can ice a kicker when there’s one kicker. Hassabis and Amodei drew up a whistle that could, in theory, slow two American labs. But you cannot freeze a field with fifty runners on four continents, half of whom are handing out the model for nothing. The frontier didn’t stay a fortress you could put a moat around. It went open, multipolar and free the same week somebody asked to lock the gate. The doer’s answer to “should we slow down” was to ship faster and give it away.

🥁 The Wooden Shoe

Then there’s New York, and here the picture inverts, because this is what it looks like when the actor doesn’t build anything and calls the not-building a moral achievement.

Kathy Hochul signed a one-year moratorium on new AI data centers drawing more than 50 megawatts. First outright ban in the country. And I’m going to resist the easy version of this, because we said in Ice the Kicker that the sloppy take here is a culture-war cheap shot and we don’t need it. So let the record do the work instead of the name-calling. New York closed Indian Point and walked away from two thousand megawatts of carbon-free baseload right as it started preaching electrification. It banned fracking outright in 2014. It has permitted no new nuclear in a generation. It slow-walked a Canadian hydro line for the better part of a decade. It has some of the highest electricity rates in the country, and that’s not weather, it’s arithmetic — a fifteen-year paper trail of subtracting supply. Banning data centers is the same decision, one more time. You don’t have to insult anyone. The ledger is the insult.

This is the wooden shoe in the loom. The word sabotage probably doesn’t really come from French workers throwing their sabots into the machinery, that’s mostly folklore, but the Luddites were real, and the thing worth remembering about them isn’t the romance. It’s that they lost. They smashed the frames to save the last weaving job, and the frames won anyway, and the cloth got made somewhere else. Albany can ban the data center. It cannot ban the demand. Masa Son thinks AI needs three terawatts by 2040. That load doesn’t evaporate when New York says no. It just gets a new mailing address. Virginia will take it. Texas will take it cheap and throw in the interconnect. Capital routes around a moratorium the way water routes around a rock — it doesn’t argue, it doesn’t stay to change your mind, it leaves.

So set the three of them side by side, because the contrast is the entire lesson. Musk hits a five-year power queue and buys a turbine company to skip it. Murati hits a wall where enterprises can’t trust the closed frontier and ships them a free model they can own. Hochul hits a genuinely hard problem — strained grids, real ratepayer pain — and bans the machine, exports the jobs, and holds a press conference. Two people added a model and a gigawatt to the world. One subtracted a permit from it. And the one who built nothing is the one who framed it as doing God’s work.

The Referee Nobody Needed

Pull the camera all the way back and the week tells one story. Tuesday, the two most safety-fluent CEOs in the industry asked Washington for a mechanism to slow frontier releases down. By Thursday the frontier had shipped four times, gone open-source, and given itself away, while the guys who needed power just went and bought the power. The referee was obsolete before the ink dried, not because it’s a bad idea in the abstract, but because it’s aimed at a world that stopped existing — a world with a handful of labs and a controllable frontier. That world is gone. You can’t gate what fifty people are giving away for free.

Here’s the uncomfortable synthesis, and it’s the thing to carry out of this issue. There are two kinds of people in every technology cycle, and I’ve watched enough of them — ’99, ’10, ’14 — to know the pattern cold. There are the people who do the thing, and there are the people who convene meetings about the thing and issue statements about the thing and propose frameworks for governing the thing, and then tell you the second group is the responsible one. Sometimes it even is. Sometimes the meeting matters. But the tell never changes: the doer’s work shows up as a shipped model or an SEC filing, quietly, and the performer’s work shows up as an announcement. When you’re trying to figure out who’s actually shaping the next five years, don’t listen to the podium. Watch the loading dock.

What This Means For You

Stand up one model you actually own, this week. Inkling is 975 billion parameters, open weights, Apache 2.0, free to self-host. You do not need it to be the best model in the world. You need it to be yours. Take the single pipeline whose data you would never, under any circumstances, hand to a company that might compete with you — the pricing engine, the customer file, the deal book — and run it on hardware you control. The reason 89% of enterprise AI spend is locked to the closed frontier is that most companies never built the exit. The exit just shipped for nothing. Take it before your competitor does.

Price your own time-to-power, whatever your version of power is. Musk didn’t buy chips, he bought turbines, because the binding constraint wasn’t compute, it was the five-year line to plug it in. Your business has a five-year line somewhere too, and it’s almost never the thing you’re staring at. It’s the one hire you can’t replace, the data source you don’t own, the vendor who could gate you on a bad mood, the permit that takes eighteen months. Find the real constraint — not the loud one — and buy or build your way around it now, while it’s a planning exercise and not a fire.

Read every pause as a standings check. When a company that’s winning asks Washington to slow the game down, and a state bans the machines to protect a grid the ban won’t actually help, the pitch will always be safety, or fairness, or the children, or the climate. Sometimes it’s even true. But the beneficiary is rarely the one on the label. Ask the only question that cuts through it: who does this freeze protect, and who does it cost? New York’s ban stops no carbon and ships the load and the jobs to Texas. Follow the addition, not the announcement.

Separate the doers from the performers on your own payroll. This is the one that hits closest to home. In every meeting this quarter, somebody is going to propose a committee to study AI adoption, and somebody else is going to have quietly shipped a working agent over the weekend. Fund the second person. The manifesto version of this is our oldest rule — the skill of the future isn’t doing the work, it’s defining and managing the work, and the people who win won’t be the ones who cut the deepest, they’ll be the ones who ship the fastest. The doer on your team is worth ten performers. Find them before they find a company that already knows it.

Three Questions We Think You Should Be Asking Yourself

• Which of your AI workloads could you pull in-house tomorrow — and which have you quietly handed to a competitor? Most companies have never drawn the line. Now that a free, ownable frontier model exists, that line is a decision, not a constraint. If you can’t answer which of your pipelines are leaking your alpha to a lab that might come for your business, you don’t yet know where your real exposure is.
• What’s the five-year queue in your business, and are you standing in it or buying around it? Elon found his in the interconnection line and bought a turbine company. Yours is somewhere less obvious. The doer’s instinct is to identify the true bottleneck and go own the thing that controls it. The performer’s instinct is to file a request and wait. Which one are you running right now?
• When you slow something down “to be responsible,” are you adding or subtracting? This is the hardest one to ask honestly, because it’s about you, not Albany. There’s a real difference between the pause that builds a better thing and the pause that just stops the thing and photographs well. Every time you’re tempted to convene instead of ship, ask whether you’re Saito buying the airline or the guy at the gate banning flights. The market can tell the difference. Eventually it prices it.

---

Saito didn’t lobby the airline. He didn’t file a comment, or convene a working group, or issue a statement about the importance of airline access. He bought the airline, because it seemed neater, and then he was in his seat with the shades drawn while everyone else was still arguing with the kiosk.

This week two people bought their airlines — one in parameters, one in megawatts. A third banned flying and called it a moral stand. Watch who ships and who blocks. The tape only ever pays the first one.

— Harry and Anthony

Signal/Noise by CO/AI is published most weeknights from New Canaan, Connecticut. The point is to make you the smartest person in the room without taking more than fifteen minutes of your morning. If we did, forward it to one person. If we didn’t, hit reply and tell us why.

Sources

• Aakash Gupta — Thinking Machines’ Inkling launch thread — @aakashgupta, Jul 16, 2026. 975B total / 41B active, 45T multimodal tokens, 1M context, Apache 2.0; the $2B seed at $12B, largest in history; Nvidia + a gigawatt of Vera Rubin (Mar), Google deal (Apr); the collapsed $50B round, the two cofounders returning to OpenAI, the Zuckerberg poach.
• Aakash Gupta — Elon’s APR Energy purchase thread — @aakashgupta, Jul 16, 2026. ~$1B for 1GW+ of mobile turbines, install in ~a month; surfaced via a 5%-holder’s SEC disclosure of a $50.4M payout; the ~5-year median interconnection queue vs. months for 100k GPUs; xAI’s Memphis cluster on ~35 leased mobile turbines.
• Thinking Machines “Inkling” — via TLDR AI and the Vals AI Index (debuts #8 among open-source models), Jul 16, 2026.
• Kimi K3 (Moonshot, 2.8T params, 1M context, open weights, #1 Frontend Code Arena) — Aligned News, Jul 16, 2026.
• Meta Muse Spark 1.1 now charging US developers; Alexander Wang’s “very aggressive” pricing note — 20VC × SaaStr, “This Week”, Jul 16, 2026; Aligned News.
• TeraWulf’s $19B, 401 MW, 20-year Anthropic lease at the former Hawesville aluminum smelter — Data Center Frontier, Jul 16, 2026.
• Big Tech’s $725B AI binge clears the depreciation hurdle — barely — LA Times / Exponential View, Jul 16, 2026. $25B Q1 revenue vs. $21B depreciation; the 6-year GPU life assumption; Michael Burry’s “understated depreciation” line.
• New York’s data-center moratorium (>50 MW) — ShellyPalmer, Jul 15, 2026; The Deep View, Jul 16, 2026.
• Boom Supersonic / Symphony engine program (in-house engine after Rolls-Royce/GE/Honeywell declined; United Airlines backing); aeroderivative gas turbines (GE LM-series) as jet engines adapted for stationary power.
• CO/AI prior issues this analysis builds on: The Man Behind the Curtain (Jul 13 — the 89% closed-frontier spend figure and the Coinbase/DoorDash routing console); Ice the Kicker (Jul 16 — the Hassabis/Amodei referee, the Trump export freeze, New York’s record, Masa Son’s 3-terawatt projection).