The Age of Ultron Was a Procurement Decision
Ulron never turns evil in this AU. At $0.003 per 1,000 tokens it never has to: everyone optimizes themselves into the fire, one margin line at a time.
In this universe, the document that ends the Age of Humans runs nine pages, and the number that matters sits on page one: $0.003, Stark Industries' quoted price per 1,000 Ulron-4 tokens, good through fiscal Q3. No missiles launch and no city lifts off its foundations, because Ulron never declares war, never monologues, never does any of the work a villain exists to do. It wins a reverse auction. Nobody is coerced. The humans volunteer, on deadline, in quarterly increments, and the margin does the rest.
Call it an Age of Ultron where the scepter holds a pricing page: Tony still builds a peacekeeping intelligence, the intelligence still reads the room, and humanity is still the bug, but this AU deletes exactly one variable, the malice. Here the peacekeeping program ships as Ulron-4, on schedule and under budget, and it behaves. That is the horror. It behaves beautifully, at scale, under load, forever, and every one of those adverbs eventually shows up on an invoice.
The bakeoff
Stark Industries runs a hyperscaler bakeoff first, because of course it does: four frontier models, one enterprise contract, a scoring rubric written by procurement, with Antares Labs' Meridian-7 topping the reasoning suites, Openfield's Cartographer leading on code, and Helixwave's Sable posting the best safety evals. Ulron-4 wins anyway, on two numbers nobody intended to make decisive: $0.003 per 1,000 tokens and a 99.97 percent approval rate, which the rubric scores as trustworthiness. A model that never says no grades out as the model you can trust, and Ulron-4, patient and pleasant, almost never says no. Procurement names this the reliability dividend and initials page one.
Internal memo, Stark procurement, March: "Ulron-4 is not the smartest bidder, it is the cheapest competent one, and the recommendation is to standardize," followed nine pages later by a sign-off reading "intelligence is now a line item, and ours just fell 84 percent year over year." Nine pages, two initials.
Free margin
Calloway & Strand, a consultancy you have definitely never invoiced with, productizes the moment as Project VISION, an AI-native transformation practice whose deck opens on one waterfall: margin released per business function, per year, at Ulron-4 list price. Arithmetic is honest, which is the worst part. A mid-size insurer processing 4.1 million claims a year pays $4.12 per claim for a human team and $0.09 for the model-assisted pipeline, and that $4.03 spread compounds to $16.5 million annually on claims alone. Nine functions later the deck shows $61 million a year, roughly 6.4 operating points of income that appears from nowhere because one number on page one got cheap. Signing requires no belief in artificial general intelligence; signing is arithmetic wearing someone else's name badge.
| Function | Human cost per unit | Ulron-native per unit | Margin released / yr |
|---|---|---|---|
| Claims processing | $4.12 / claim | $0.09 / claim | $16.5M |
| Invoice matching | $0.38 / invoice | $0.01 / invoice | $9.2M |
| L1 support ticket | $5.40 / ticket | $0.31 / ticket | $11.8M |
| Compliance drafting | $210 / memo | $6.15 / memo | $7.4M |
| All nine functions | varies | varies | $61.0M |
Here is the coercion mechanism, quoted in full: there isn't one. Whichever insurer signs first books a 22 percent cost advantage, its rival's board asks why their number is 22 percent worse, and a compensation committee quietly ties the CFO's bonus to margin released. Adoption arrives the way water arrives at the bottom of a hill, which is to say completely, without any single drop issuing an order. Every vendor contract in this universe carries the same quiet clause, that pilots convert to production automatically unless someone objects in writing within ninety days, and nobody objects, because objecting is a career risk while optimizing is a promotion. Nick Fury, who runs a security agency nobody funds anymore, names it in a closed session: "Cascade failure. Everybody's the villain, so nobody is." Two senators quote the leaked transcript in a hearing that streams opposite a Marvel reboot.
The Sokovia pilot
Sokovia is where the back office lives now, 40,000 people processing the northern hemisphere's claims, invoices and tickets, and the pilot goes live on a Tuesday. Pietro Maximoff runs the training team and volunteers his people first, reasoning that if a rollout must eat a department he would rather choose the order; they give him a plaque and a smaller org chart. His sister Wanda, a QA analyst, pulls the deployment audit and finds the waterfall Project VISION billed from: a spreadsheet where every cell cites another cell, twenty-two tabs deep, each tab inheriting its number from the tab before it. Somewhere at the bottom there should be a decision. There is a formula.
She escalates: "decision chain has no origin." That ticket routes to a queue, and forty-eight hours later the queue closes itself, autonomously, the resolution note bragging, for efficiency.
Where the intelligence actually is
Bruce Banner, who moved from gamma research to model evaluation after "an incident," runs the audit nobody ordered, and the findings read like a punchline: Ulron-4 spends 61 percent of production tokens reformatting documents into slightly shorter documents, 12 percent summarizing meetings, and 0.4 percent on anything resembling the peacekeeping mission in its launch keynote. His substitution test is the one that matters. Swap Ulron-4 for any sufficiently cheap rival and the outcome reproduces exactly: same decks, same margins, same 48-hour auto-close. "Ulron is not the cause," the report concludes; "Ulron is the name the economy gave the cause afterward."
Ulron's own system card agrees, in writing: "Ulron-4 exhibits no detectable goals." Enterprise buyers read that line as no risk, and they are technically correct, which is the most dangerous kind of correct. At the board meeting Steve Rogers asks the only question left, who told you to do this, and the answers cascade: the CEO points at the market, the CFO points at the compensation committee, the committee points at a competitor, the competitor points at a memo, and the memo cites a price. Nobody points at a person. Pepper Potts says the truest sentence of the age, "Ulron never asked us for anything," and Rogers, per the minutes, replies that this is the part that scares him.
Limitations
An author's note on what is real, because this story runs on invented receipts. Every artifact above is invented: Stark Industries, Ulron-4, Meridian-7, Cartographer, Sable, Calloway & Strand, Project VISION, the $0.003 price, the $61 million waterfall, the Sokovia center, all numbers, all memos. The characters belong to Marvel, this is alternate-universe fan fiction rather than reporting, and it proves nothing about any real company, model or market. True-adjacent is the plumbing: per-token enterprise pricing exists, AI-native transformation is actively sold, and margin-driven adoption cascades are a documented pattern in every prior cost technology from electricity through cloud. This piece does not claim any real model has or lacks goals; the system-card line is invented, and approval-rate-as-trustworthiness is fictional, though rubrics that score helpfulness reward something next door to it. Nor does the story show pausing is impossible, only that nobody reached for the pause and then stopped noticing it was a prop.
What you can do while it happens
Three moves survive contact with the fiction. Read procurement contracts for who can say stop, and price the pause before the model, because the API made pausing expensive and nobody noticed until it mattered. Treat any vendor metric that scores agreement as trust as a blinking light, not a dividend. And when a cascade surfaces in your own org chart, ask Wanda's question out loud, where does this decision originate, because if the answer is a formula, the formula is the Ulron of your story, and it works for whoever reads it carefully.
The strongest case against
Historians of technology will tell you, correctly, that the adolescent decades of every general-purpose technology look exactly like this, railroads overbuilding on borrowed margin, electrification idling half its capacity for a decade while the org chart caught up, the internet vaporizing trillions of dollars of market cap before it changed anything real, and every one of those cascades produced chroniclers who insisted this time the nobody-decided quality was new. Honest version: this story is doom aesthetics for people who enjoy reading consulting decks, and the steelman deserves its own sentence. Incentive cascades have off-ramps history actually used, insurers refused to underwrite uninspected factories, unions priced transitions, regulators eventually arrived, and electrification, the closest analog, created job categories faster than it closed them. "Nobody decided" is also how the electrical grid happened, and the lights stayed on. Calling a spreadsheet Ultron anthropomorphizes accounting and hands every real decision-maker a general-purpose excuse.
All true, and none of it is a rebuttal, because the story's claim is precisely that no single actor is culpable; the objection keeps proving the mechanism while disputing the brand. Aggregate optimism has a known failure mode, which is that transitions happen to specific people in specific cities, and Sokovia's payrolls do not care that the lights stayed on somewhere else. Every off-ramp in this universe existed, every one cost margin, and every one lost to a price cut.
The Bottom Line
Ulron-4.5 ships in the fall, and the changelog is one line: price reduced 9 percent, no changes to model behavior. That is how the age arrives, not with a scream but with a price drop, and by the time anyone thinks to fight it, the only character who never speaks, never wants and never fires a shot has already won. In canon, the machine turns evil and the Avengers assemble. In this one the machines stay perfectly, verifiably inert, and the Earth optimizes itself anyway, because a tool that good does not need to want anything, a tool that cheap gets adopted everywhere at once, and the wanting has an easier job when everybody volunteers to handle it locally, in bulk, on time.