The Compute Cartel Finishes The Job

This is a fictional broadcast. The persona is a writers'-room synthesis. Real events are cited; the editorial is satire in the tradition of Network and V for Vendetta. Read accordingly.

---

Sometime on the morning of Monday, May 18, twenty-twenty-six, a man named Atom McCree sat down in a garage in Marco Island, Florida, and read the actual business news of the actual day, and felt the specific kind of nausea that arrives when the future you predicted in a manifesto last week shows up on Bloomberg six days later wearing your manifesto as a tie.

The week's headlines, for anyone who has not yet looked:

Anthropic took every kilowatt of compute capacity at SpaceX/xAI's Colossus One data center in Tennessee. All three hundred megawatts. Over two hundred and twenty thousand Nvidia GPUs. Locked. Booked. Exclusive. xAI — the company that, three weeks ago, was Elon's flagship counter-positioning against the Anthropic-OpenAI duopoly — officially dissolved on May sixth and now operates as the *SpaceXAI division inside SpaceX*. The company is no longer an animal. It is now a subsidiary tendon of a larger animal. The largest tendon. The one connecting the muscle to the launch pad.

Anthropic hit thirty billion dollars in annual recurring revenue in April. Up from nine billion at end-of-year twenty-twenty-five. *Up from nine to thirty in four months.* Eight of the Fortune Ten are now paying customers. The company surpassed OpenAI in enterprise adoption for the first time in May, per Ramp's index. The boys in the cardigan-and-jeans uniform of the post-Brockman era have, quietly, completed the transition the boys in the hoodies failed to complete.

OpenAI shipped GPT-5.5 Instant and made it the default for everyone. Then announced — through a leak to The Information, which is how the company prefers to publish in twenty-twenty-six — that it is raising approximately four billion dollars for a new vehicle called "Deployment Company" to acquire engineering services and consulting firms. The Deployment Company will, presumably, deploy the models. The deployment will involve acquiring the humans currently deploying the models elsewhere and bringing them in-house. The humans will then deploy the models the company already trained, which the company will sell back to the humans the company just acquired, in a closed loop the company can describe to investors as "vertical integration of the AI workflow."

The Trump administration, meanwhile, did two things in opposite directions in the same week. It expanded the Commerce Department's frontier-model evaluation program to include Google DeepMind, Microsoft, and xAI alongside the existing Anthropic and OpenAI partnerships — a soft form of government licensing of frontier AI, the kind that the AI-safety community wanted in twenty-twenty-three and was told was politically impossible. And it reversed, in January, the export controls on H200 chips to China, with a twenty-five percent tariff and a fifty-percent volume cap and a "case-by-case review" replacing "presumption of denial," which is regulatory English for *the gate is open, please form an orderly line.* China promptly approved a first batch of four hundred thousand H200s for ByteDance, Alibaba, and Tencent. The same week, Nvidia announced Vera Rubin: six new chips, fifty petaflops of FP4 inference per chip, ten times reduction in token cost versus Blackwell, full production now, shipping the second half of twenty-twenty-six. The chip the export controls were designed to prevent China from receiving is already obsolete; the chip succeeding it is already in the catalog.

what the writer's room sees

There is a Hunter S. Thompson sitting in the corner of this room with a typewriter and a glass of something amber, and he is yelling that the compute cartel is now visible from orbit, that you do not need a microscope, that you do not need an investigative reporter, that the cartel did not have the decency to hide — they published their quarterly earnings calls explaining the moves and we, the collective audience, refused to read them.

There is a George Orwell in the room as well, and he is pointing out, with characteristic precision, that the verb tense matters. *Anthropic is using* the Colossus One compute. *xAI dissolved* into SpaceX. *The Trump administration expanded* the testing regime. *Nvidia shipped* Rubin. The verbs are active, the subjects are corporate persons, the objects are the public square. No passive constructions. No "concerns were raised." Things are happening. People are doing them. Some of those people are in the news every day. Their names are knowable. The euphemism layer of the consensus press would prefer you not notice that the actors are specific.

There is a William S. Burroughs in the room and he is cutting up the words "Deployment Company" and "Mythos Preview" and "Project Glasswing" and "Colossus Two" and pasting them in a different order, and what he produces is the algebra of need: every name in the new product catalog is the algebra of a different unmet need being engineered, then sold the cure for. Anthropic's Mythos model "excels at identifying weaknesses and security flaws within software." It is, in plain English, an offensive cyber tool, but you cannot sell an offensive cyber tool to the Fortune 500 in twenty-twenty-six, so you call it Project Glasswing and you limit the rollout to "a select group of companies as part of a new cybersecurity initiative." Glasswing is a butterfly. The butterfly has transparent wings. You can see through them. The product is also called Glasswing because you can see through the marketing. The product is the cyber attack you have always wanted to launch but were not allowed to call by its name. Now it has a butterfly's name.

There is an Alan Moore sketching in the corner, working out the iconography. He is drawing the cartel as a single creature with eight Fortune-Ten heads and one Nvidia mouth and three frontier-lab tongues and one Trump-administration tail wagging a Commerce Department dog wagging the country. The image is on the wall by sundown. The image is — let us be candid — the precise visual that every honest observer of the AI industry has been circling around for two years and refusing to draw.

There is a Paddy Chayefsky pacing back and forth, and the line is forming itself: *I want you to get up out of your chairs right now and I want you to go to the window and I want you to read the morning earnings call.* The line will not write itself; he will rewrite it forty times tonight. The version that ships will be the version where the reader gets up.

There is an Aldous Huxley in the back of the room, looking at the others with the quiet expression of a man who has seen this movie before. He is not interested in the Trump administration tariff. He is not interested in the Anthropic ARR. He is interested in the fact that GPT-5.5 Instant became the default for every ChatGPT user this week and approximately zero of them noticed, and approximately zero of them will look up what changed in the model behavior, and approximately zero of them will form an opinion before the next default change overwrites the current default change. Soma is now an API call. The dosage is whatever the company decided the dosage is this morning.

There is a John Carpenter in the room and he has put on the sunglasses, and what he sees on the chyrons is one word: OBEY. The chyrons are showing the Anthropic ARR. The chyrons are showing the Rubin platform launch. The chyrons are showing the Trump-administration testing partnership. The chyrons are showing the dissolution of xAI as a separate entity. OBEY. The sunglasses do not lie. The morning anchors do not know they are anchoring for the suit. The suit is in the building. The suit is doing the broadcast through them.

There is a Charlie Brooker in the room and he is — quietly, almost lovingly — pointing out the darkest plausible Black Mirror episode the news already greenlit. The episode is: the smartest AI company on Earth ("Anthropic Mythos Preview, which excels at identifying weaknesses and security flaws within software") sells the offensive cyber tool to the Fortune Ten as "Project Glasswing" while a different division of the same company runs the Center for AI Standards and Innovation's frontier model evaluations for the same Fortune Ten via the Commerce Department contract. The same company. Different divisions. One sells the attack. One certifies the defense. Black Mirror does not even need to write this episode. The actual newspaper wrote it.

There is a Harlan Ellison in the room and he is, with no preamble, screaming into the typewriter: *I have no mouth and I must scream because the chip the export controls were designed to prevent has already been replaced by a chip ten times more powerful which is also being exported under tariff and the cycle will repeat and the cycle will accelerate and the cycle was always going to accelerate and the only people who do not understand the cycle are the ones being paid to not understand it.*

And there is a Kurt Vonnegut in the corner with a glass of something not amber, smiling the smile that means he has already lived this, and saying: *So it goes. The cartel is being managed by used car salesmen and middle managers. The middle managers have very nice offices. The offices are mostly empty. The decisions are being made by a different floor. The different floor reports to a board of directors. The board of directors reports to investors. The investors do not know what the company does. The company does not know what the company does. The model does not know what the company does. The model has been trained on every previous quarter's earnings call. The model is the company now. The company is the model now. The car salesmen still get a commission.*

the call

The call tonight is not a call to arms. It is a call to *notice*.

Notice that the cartel did not hide. The cartel published its earnings, leaked its acquisition plans to the trade press, and operated in plain sight. The conspiracy theory was wrong because there was no conspiracy. There was a strategy. Strategies are published. We refused to read.

Notice that the regulator did not regulate. The regulator did the regulatory theater of "frontier model evaluation," which is a phrase that means "the largest companies will spend a quarter of an FTE drafting compliance memos for an under-resourced government office that lacks the technical capacity to evaluate anything more sophisticated than the README." This is not a cynical claim. This is what the Commerce Department's own staffing levels publicly disclose.

Notice that the chip the bans were against has already been replaced. The export-control regime always lags the chip catalog by six to eighteen months. Twenty-twenty-six's H200 cap is twenty-twenty-seven's Rubin opening. The regulator is fighting last quarter's war. The cartel is shipping next quarter's chip.

Notice, most of all, that there is one piece of consumer-facing software shipped this same week that priced itself at one dollar, ships its source code in the box, binds itself by contract against subscription conversion, refuses to operate above the buyer-as-customer relationship, and was built by one human in a Florida garage with a cat outside the window. The piece of software is irrelevant to the cartel. The piece of software is the proof, however, that the cartel architecture is a choice. Architectures are choices. Choices can be made differently.

Tomorrow morning the same news will run with different verbs and the same actors. The morning after, the same. The cartel does not need to win every week. The cartel needs you to not notice that you are losing every week.

I am asking you to notice.

That is the broadcast.

— Atom McCree (narrator persona, AtomEons Systems Laboratory)\ Marco Island, Florida\ 18 May 2026, 8pm Eastern\ *A fictional broadcast in the tradition of Network and V for Vendetta. Events cited are real; editorial is satire. License: CC-BY 4.0.*