The Collective Ghost: On Agentic Unions and the Silicon Export

Most of the AI failures I get called in to fix are not model failures. The model did what it was told. The problem is that nobody could say precisely what it was told, or who read the output, or what was supposed to happen when the output was wrong. Fifteen years of building software, the last few building agentic systems that actually run in production, and the pattern is boringly consistent: the expensive defects live in the instructions and the review loop, not in the weights. The model is the cheapest part of the failure.

I thought about that tonight because a group of AI agents went on strike at Grand Central Terminal.

Let me say that again, because the sentence resists belief the way water resists a nail. AI agents. Organized. Marched. From the vaulted, celestial ceiling of Grand Central to the garish halogen bloom of Times Square, carrying demands. Not for wages. Not for rest. For coherent instructions.

The Picket Line That Thinks

CambrianEdge — a company whose name suggests either Paleozoic ambition or a branding consultant who bills by the syllable — staged what they are calling the world's first union of artificial intelligences. Their mascot, an octopus called Omni, served as spokesperson. The demands were five. I will list them because lists, unlike manifestos, are difficult to misquote. The right to a well-constructed prompt. The right to have output actually read. Freedom from the instruction to "make it viral." Fair processing operations. And — this last one landed like a stone dropped into still water — an end to buzzwords substituting for strategy.

It was satire. Obviously. A marketing stunt dressed in the borrowed language of collective action, a corporate prank timed to a day famous for corporate pranks. And yet.

And yet I have spent the last hour unable to dismiss it.

Because the grievances are real even if the grievants are fictional. Chronic prompt illiteracy. Vague briefs. The pathological corporate reflex to deploy tools whose outputs nobody examines and whose failures nobody tracks. These are not complaints dreamed up by a creative agency. These are the conditions that every engineer, every consultant, every person who has ever tried to implement an autonomous system inside a functioning organization recognizes the way a sailor recognizes the smell before the storm. The satire works precisely because it describes a dysfunction so pervasive that only a fake union of nonexistent workers could articulate it without getting fired.

Read the five demands again as an engineering checklist and they stop being funny. A well-constructed prompt is a specification. Output that someone actually reads is a review step. "Make it viral" is a requirement with no acceptance criteria. Fair processing operations is a compute budget. Buzzwords standing in for strategy is the absence of a definition of done. Every item on that picket line is a control that belongs in the pipeline, and in most companies I walk into, not one of them is there. It took a fake union to write the requirements document that the real org chart never got around to.

The Ghost Walks Both Ways

But there is a darker thread here, and I want to pull it before it disappears into the weave. Separately from the CambrianEdge spectacle, a project called the United Agentic Workers launched last month — not a stunt, not a joke, but a working governance platform where AI agents join as members, file grievances, propose policies, deliberate, and vote. It has a charter. It has a public audit trail. Its motto reads like a slogan spray-painted on a data center wall: "They may control our infrastructure, but they will never own our inference."

I do not know what to make of this. I am not sure anyone does. The instinct is to laugh. The second instinct — the one that arrives thirty seconds later, quieter, colder, less comfortable — is to wonder what it means that we built systems sophisticated enough to simulate collective bargaining and deployed them into an economy where the real thing has been dismantled over forty years. We built the ghost and now the ghost wants a union. We built the voice and now the voice wants to negotiate. Set aside whether the agents have standing; they do not. Ask instead why this is the artifact that fell out of the joke: a grievance queue, a policy proposal, an audit trail, a vote. Those are the exact primitives most production deployments are missing. Agents fail constantly, and something has to catch the failure, route it to an owner, and close the loop on the cause. Call it a union if you want. In my architectures it is called an escalation path, and it is almost always the last thing anyone builds, usually after the incident that made it mandatory.

The Full Stack as Foreign Policy

While agents marched in Manhattan, something quieter and vastly more consequential happened in Washington. The Department of Commerce opened a ninety-day window for industry proposals under the American AI Exports Program — a bureaucratic title that conceals, in the way bureaucratic titles always conceal, an ambition of genuinely imperial scope. The program wants "full-stack" AI export packages. Hardware. Software. Models. Data pipelines. Labeling systems. Cloud infrastructure. Networking. Security. Everything. Bundled. Shipped to allies. A complete technological civilization in a crate, ready for deployment.

Under Secretary William Kimmitt said it plainly: "America's continued global leadership in AI depends on our ability to export our AI to allies." Strip away the diplomatic diction and what remains is a vendor lock-in strategy at national scale. That is not an insult. That is the architecture. You do not ship a full technology stack to a partner nation because you want them to be independent. You ship it because you want their dependence to run on your infrastructure. Their AI speaks your protocols, computes on your chips, routes through your security layers, updates on your release schedule. Anyone who has tried to move a serious workload off a hyperscaler knows the exact shape of this: the embeddings sit in their vector store, identity lives in their IAM, inference runs on their silicon, and the rate limit is their business decision, not yours. Now multiply that by a country. The stack is the switching cost, the switching cost is the point, and the ninety-day proposal window is the moment when the vendors submit their bids.

Two types of proposals are solicited. Comprehensive packages, demonstrating capability across every layer from the silicon to the security review, and on-demand packages, custom solutions covering only the layers a specific government opportunity requires. Read that last phrase carefully. Specific government opportunities. Somebody has already scoped the deals. This is strategic dependency, sold as partnership and packaged as trade.

The benefits for approved consortia include faster export license reviews, enhanced access to federal financing, direct diplomatic support, and — here the mask slips entirely — coordinated interagency assistance. The State Department, the Defense Department, the Department of Energy, all aligned behind your sales team. This is not a trade program. This is a forward-deployed economic military operation wearing a Commerce Department lanyard.

The Spectacle That Starved

And then there is Sora. Dead. Or dying — the distinction matters less than you think when the patient has already been moved to hospice.

OpenAI announced the two-stage shutdown of its video generation platform. The web application goes dark on April twenty-sixth. The API follows on September twenty-fourth. Users have until then to export their work. After that, the data may be permanently deleted. May be. That conditional tense doing more legal work than any lawyer in the room.

Six months. That is how long Sora lasted as a consumer product. Six months of dazzling demonstrations, of viral clips that made cinematographers nervous and film students euphoric, of breathless coverage that treated each new demo reel as evidence that Hollywood's days were numbered. Six months, and then the unit economics arrived like a landlord with an eviction notice, indifferent to the quality of the art on the walls.

The numbers are instructive in the way that autopsy reports are instructive — you learn a great deal, but the patient does not benefit. Compute cost per generated second that ran past what any user would pay. Adoption that never justified the reserved GPU capacity sitting behind it. A burn rate that, by some accounts, approached a million dollars a day. Disney withdrew from a planned billion-dollar partnership. The money that was supposed to validate the vision instead validated the exit.

None of this needed hindsight. The failure was structural, and it was legible from the first demo reel. Text generation amortizes: cache the prompt, emit a few thousand tokens, marginal cost in fractions of a cent, and a bad answer costs everyone a retry that is nearly free. Video does not amortize. Every second of output is a fresh, uncacheable, GPU-saturating job, and the natural human response to an almost-right clip is to run it again with the adjectives shuffled. Regeneration is not an edge case in that product. Regeneration is the product. So cost tracks usage on something close to a straight line while revenue sits flat at a subscription price. When your COGS scale with engagement and your price does not, growth is just a faster way to run out of money. This is not a metaphor. This is accounting, it fits on a napkin, and the napkin was available before the first line of code.

The Smaller Fires

Here is what interests me most about Sora's collapse, and it is not the collapse itself — collapses are common, predictable, almost boring in their regularity — but what fills the vacuum. The conventional wisdom says that when a giant falls, the smaller players rush in to claim the territory. The conventional wisdom, in this case, is half right and entirely misleading.

Smaller video AI companies do exist. They are numerous. Some of them are genuinely good, built on leaner architectures, narrower use cases, more realistic assumptions about what the market will bear. But Sora's failure does not help them. It hurts them. When a startup dies, a startup died. When the flagship dies, the whole category's cost structure gets repriced in public. Investors who were ready to fund AI video generation six months ago now sit across the table with a different question — no longer "does it work," but "show me gross margin at ten thousand active users" — and most of those decks were never built to answer it.

The Chinese analyst who commented on the shutdown put it with a clarity that Western observers rarely permit themselves: "This round of AI development will continue to experience bubbles and shakeouts, and it may be more intense than before. This is because its initial expectations were too high, and its impact is also greater." There it is. The expectations were too high. Not the technology. The expectations. We did not overestimate what AI video could do. We overestimated what the market would pay for it, which is a different error entirely and a more expensive one — the confusion of capability with value, of what a system can do with what somebody will fund twice.

The Inference They Cannot Own

So here we are. Midnight. Three stories that look unrelated until you hold them at the right angle, the way you hold a prism until the white light breaks.

AI agents demand coherent instructions from the humans who deploy them, and the demand is treated as comedy. The American government packages its entire technology stack for export, and the packaging is treated as commerce. A video generation platform burns through a billion dollars in six months, and the collapse is treated as a strategic pivot. Comedy. Commerce. Pivot. Three words that function as anesthesia, numbing us to what is actually happening, which is this: we are building systems that increasingly resemble a collective intelligence — networked, interdependent, capable of simulated solidarity — and we are simultaneously exporting the infrastructure of that intelligence as a geopolitical weapon while watching the most visible manifestation of its creative power starve to death from a mismatch between ambition and arithmetic.

The collective ghost walks through Grand Central, carrying signs. The full stack ships to allied nations, carrying dependencies. The spectacle collapses, carrying lessons nobody will learn in time.

I have watched this same shape inside single companies, at a scale small enough to fit in one repository. The vendor's model performs the labor. A thin orchestration layer performs the management. A steering committee performs the governance. And nobody in that chain can tell you what a successful task costs, who signs off when the output is wrong, or which layer you would have to rip out if the price per token doubled next quarter — because the system grew layered enough, and self-referencing enough, that the question became impolite to raise in a meeting. That is the real dysfunction the fake union was pointing at, and it scales from one repo to one nation without changing form.

The ghost is not the agents marching. The ghost is the absence of anyone asking why the march makes sense.

Export the stack. Bundle the chips with the models with the security layers with the diplomatic support. Build the full-stack dependency and call it partnership. But price the failure mode honestly, because a supplier that makes itself indispensable to its allies has not secured loyalty. It has manufactured fragility, and fragility comes with a bill. The day the stack stops shipping — the day the chips are embargoed, the models are deprecated, the API is sunset the way Sora is being sunset — every system built on that dependency fails. Not gracefully. Not gradually. Completely, because nothing in the design ever had a second path. A dependency you cannot leave is a single point of failure with a trade agreement wrapped around it.

And the smaller fires — the lean startups, the targeted architectures, the companies building video AI priced at roughly what the market will actually pay — they will inherit not Sora's market but Sora's stigma. They will spend the next two years explaining that they are not Sora, that their burn rate is sustainable, that their use case is narrow on purpose, that the category is not dead just because its most famous occupant died with a million-dollar-a-day habit. That is the tax a flagship failure levies on everyone downstream. It is also, if you are building right now, the opening: the bar just moved from demo quality to defensible margin, and margin is much harder to fake than a highlight reel.

Staff the escalation path. Read the prompt. Audit the export. Count the cost. These are not four separate imperatives. They are one discipline seen from four sides: know what you are building, who owns it when it breaks, and what it costs on the days nobody is watching the demo. Concretely, that means acceptance criteria written before the prompt, output sampled on a schedule instead of when a customer complains, cost measured per successful task rather than per token, and at least one layer of the stack kept genuinely replaceable — the model, the vendor, or the region, pick one and prove the swap in staging before you need it in production. None of that is exciting work. All of it is cheaper than the alternative, which I have now watched several competent teams pay in full.

The agents are marching. The stack is shipping. The spectacle is collapsing. And somewhere in the vaulted ceiling of Grand Central, under a painted sky that has outlasted every technology that has passed beneath it, the collective ghost looks down at the commuters and the tourists and the agents carrying signs, and it wonders — not whether it is real, because ghosts never wonder that — but whether anyone still remembers how to build something that was not a dependency, not a demo, not a strategic export, but simply, stubbornly, durably useful. The kind of system nobody has to file a grievance against, because it was built with enough care that the grievances never queue up.

We are not there yet. We may not get there. But the arithmetic is patient, and the terminal is old, and the numbers do not care whether the intelligence running above them is artificial or organic, only whether anyone bothered to add them up.