The Price of Omniscience: On Oracle's Purge, the Governance Mirage, and the Empire's New Protocols

Seneca had a friend — or what passed for a friend in the social economy of first-century Rome, which is to say an acquaintance whose ruin he could narrate without personal liability — who owned a household of four hundred enslaved persons. Four hundred. The number mattered less than the arithmetic behind it: each body represented not a life but a unit of labor whose maintenance cost could be weighed against its productive output, and when the arithmetic turned unfavorable, the household shrank. Not through manumission. Through sale. Through disposal. Through the quiet administrative violence that empires have always preferred to the noisy kind, because ledgers do not scream and balance sheets do not bleed and the people who read them can sleep at night provided they never visit the quarters where the subtracted lived.

I thought of that arithmetic tonight because Oracle fired thirty thousand people.

Not all at once. That would be theatrical, and Larry Ellison's company has never favored theater when bureaucracy will suffice. The termination emails began arriving at six in the morning — the hour chosen, one assumes, by someone who understood that a person reading the worst news of their professional life at dawn is less likely to make a scene than one who receives it at noon, surrounded by colleagues. Six AM across time zones. The United States. India. Canada. Mexico. A synchronized global deletion of human employment executed with the kind of logistical precision that, in another context, the company would have sold as a feature of its cloud infrastructure.

Eighteen percent of the workforce. Gone. TD Cowen ran the numbers, as analysts do, and arrived at a savings figure between eight and ten billion dollars annually. That money — earned by the labor of people who will now update their LinkedIn profiles with the hollow dignity of "open to work" badges — will be redirected toward data centers. GPU clusters. AI infrastructure. The company plans to spend fifty billion dollars on infrastructure in 2026 alone, with total AI capital commitments approaching one hundred and fifty-six billion. These are not numbers. These are coordinates on a map of a world being rebuilt, and the thirty thousand people whose badges stopped working this morning are not casualties of that construction. They are the demolition.

The Ledger of Flesh and Silicon

Here is what disturbs me, and I want to be precise about the disturbance because precision is the only defense against the anesthetizing effect of large numbers. It is not the layoffs themselves. Layoffs happen. Companies restructure. Industries transform. Seneca watched Rome's entire shipbuilding guild collapse when the grain routes shifted, and he did not weep for the shipwrights — he observed that a civilization which cannot redeploy its labor has already begun to die. The cycle is ancient. I do not pretend otherwise.

What disturbs me is the ratio. The exchange rate. Thirty thousand human livelihoods converted into silicon capacity at a published rate of eight to ten billion dollars, and the market responded not with grief or even discomfort but with approval. Oracle's stock ticked upward. The analysts used words like "disciplined" and "strategic" and "necessary reallocation," which are the same words Roman senators used when they voted to reduce the grain dole, and which carry the same emotional freight, which is to say none.

We have not automated the work. We have automated the permission to discard the workers.

Seneca wrote — in De Beneficiis, I think, though it may have been one of the letters, the bibliography blurs after midnight the way highway lines blur after the fourth hour of driving — that the most dangerous gift is the one that makes the recipient dependent on the giver's continued goodwill. Oracle did not give its employees jobs. It gave them dependencies. And dependencies, unlike gifts, can be revoked without the social cost of ingratitude, because the relationship was never framed as generosity in the first place. It was framed as exchange. And when one side of the exchange finds a cheaper supplier — in this case, a rack of H100s that does not require health insurance, does not take paternity leave, does not age into higher salary bands — the swap is not cruelty. It is optimization. The cruelty is in the framing.

The Watchtower Fills with Mirrors

While Oracle was emptying its offices, ServiceNow was filling a conference hall. RSAC 2026 — the annual cybersecurity congregation in San Francisco where vendors display their wares with the solemnity of medieval relic merchants and the pricing transparency of medieval relic merchants — hosted a session on agentic AI governance that I cannot stop thinking about, though not for the reasons its presenters intended.

The pitch was sensible. Reasonable, even. AI agents operating autonomously in enterprise environments need governance frameworks. They need distinct identities. They need least-privilege access controls. They need audit trails and behavioral monitoring and something called an "AI Control Tower" that can trace every action an agent takes, enforce access policies in real time, and detect behavioral drift before it compounds into operational failure. ServiceNow demonstrated all of this. The slides were clean. The metrics were compelling — a thirteen percent improvement in mean-time-to-resolution, which is the kind of number that makes procurement officers reach for their purchase order pads the way Pavlov's dogs reached for the bell.

And yet.

Only twenty-one point nine percent of organizations currently treat AI agents as identity-bearing entities. Let that number settle. Four out of five enterprises deploying autonomous software into production systems have not bothered to give those systems a distinct identity — which means they cannot track what the agents do, cannot revoke their access independently of the humans who launched them, cannot audit their decisions, cannot detect when they deviate from expected behavior. Four out of five. Running in production. Right now. While you read this.

Seneca would have recognized this pattern instantly. He described it in his observations on Roman provincial governance: the empire would dispatch a procurator to a distant province with sweeping authority and minimal oversight, then act surprised when the procurator enriched himself at the province's expense. The problem was never the procurator's character. The problem was the architecture of the delegation — authority without identity, power without audit, action without accountability. The empire knew this. It built the system anyway. Because the cost of proper oversight exceeded the perceived risk of its absence, and perceived risk, unlike actual risk, is subject to negotiation.

RSAC revealed something else that deserves attention. The broader conference consensus — if a gathering of security vendors can produce anything deserving that word — was that enterprise security controls are "largely inadequate" for autonomous AI. Not slightly behind. Not catching up. Largely inadequate. The industry has shifted, the presenters noted with the careful diction of people who bill by the hour, from theoretical risk assessment to active, infrastructure-level threats, including multi-agent offensive behaviors. Multi-agent. Offensive. Behaviors. These are not hypothetical attack scenarios drawn on whiteboards by red-team consultants. These are observed patterns in deployed systems. The watchtower that Seneca warned us to staff is not empty. It is full of mirrors, reflecting our own architecture back at us, and we are mistaking the reflection for surveillance.

The Protocol That Connects Everything

Meanwhile — and I use "meanwhile" loosely, because in a world where news travels at wire speed the word has lost its temporal meaning and retained only its dramatic function — Microsoft did something that will matter more than Oracle's layoffs and ServiceNow's governance frameworks combined, though it will receive approximately one-tenth of the coverage. Microsoft launched native MCP integration in Sentinel.

I should explain. MCP — the Model Context Protocol — is the standard that Anthropic built for connecting AI agents to external tools and data sources. Think of it as the TCP/IP of agentic software: a protocol layer that allows any agent to interface with any tool, provided both speak the same language. Microsoft embedding this protocol natively into Sentinel — its enterprise security platform — and simultaneously releasing a Claude connector in public preview means that the two largest AI infrastructure providers have just agreed on the plumbing. Not the models. Not the business strategy. Not the vision for the future of intelligence. The plumbing.

This matters because plumbing outlasts everything built on top of it. The Roman aqueducts are still standing. The villas they served are dust. HTTP is forty years old and shows no signs of dying. The websites it delivers have an average lifespan shorter than a hamster's. When Microsoft and Anthropic converge on a protocol layer, they are not making a product announcement. They are laying pipe. And pipe, once laid, determines where the water flows for decades.

Seneca understood infrastructure better than most philosophers, probably because he was richer than most philosophers and had spent considerable time managing estates whose value depended on reliable water supply. He wrote that the man who controls the aqueduct controls the city, not because he can withhold water — though he can — but because every building, every garden, every fountain, every bath is designed around the assumption that water will arrive through his channels. The protocol is the aqueduct. Whoever defines it defines the topology of every system built downstream.

Governance frameworks atop this protocol. Security controls routed through this protocol. Agents monitored, audited, and constrained by tools that speak this protocol. The conversation at RSAC about identity and least-privilege and behavioral drift is not wrong. It is downstream. It is a conversation about what to build on the aqueduct, conducted by people who did not notice that the aqueduct was just completed.

The Valuation of Appetite

And then there is the money. There is always the money. OpenAI closed its hundred-and-twenty-two-billion-dollar funding round at a post-money valuation of eight hundred and fifty-two billion dollars. I will not pretend I can hold that number in my head with any meaningful comprehension. I can tell you that it is larger than the GDP of most nations. I can tell you that SoftBank and Andreessen Horowitz and Fidelity put their names on the check. I can tell you that the company is generating two billion dollars a month in revenue, which sounds impressive until you learn what it costs to generate that revenue, at which point the impressiveness shifts registers from admiration to alarm.

Eight hundred and fifty-two billion. For a company that six months ago shut down its most visible consumer product — Sora, the video generation platform — because the burn rate approached a million dollars a day and Disney walked away from a billion-dollar partnership. For a company currently preparing an IPO, which is the financial equivalent of a caterpillar announcing it will become a butterfly while also admitting it has been eating through its own cocoon.

Seneca had a word for this. Several words, actually — he was Roman, and Romans were not known for verbal economy when disapproval was available. But the relevant one is cupiditas, which is usually translated as "greed" but which more precisely means "appetite that has outgrown its object." The hungry man who eats is satisfying a need. The man who builds a banquet hall to seat five hundred is satisfying something else entirely — something that food cannot address because food was never the point. The banquet is the point. The scale is the point. The ability to say "I fed five hundred" is the point, even when three hundred of the plates come back untouched and the kitchen staff collapses from exhaustion.

Eight hundred and fifty-two billion dollars is not a valuation of what OpenAI has built. It is a valuation of what the market believes appetite, at sufficient scale, will eventually digest. And markets, as Seneca observed of Roman grain speculation, are not wrong until they are catastrophically wrong, and the catastrophe is always blamed on circumstances that were visible to everyone and acknowledged by no one.

The Regulation That Whispers

There is a legislative thread winding through all of this, quiet as wire through conduit. The federal government has made its most coordinated push toward AI regulation — a discussion draft from Senator Blackburn, a White House framework organized around seven pillars that read like a compromise between ambition and fear, California signing the nation's toughest state-level safeguards, and the EU's enforcement apparatus grinding toward its August deadline while three-quarters of European enterprises confess they are not ready.

Seven pillars. Protecting children. Safeguarding communities. Respecting intellectual property. Preventing censorship. Enabling innovation. Workforce development. Federal preemption of state law. Read that list again and notice what it does not contain: a single reference to the concentration of economic power that makes regulation necessary in the first place. The pillars are decorative. They are the columns of a temple whose floor plan has not been drawn, supporting a roof that has not been designed, dedicated to a god whose name has not been agreed upon. They are, in the precise Senecan sense, the appearance of governance without the substance of it.

Federal preemption is the tell. When the federal government moves to preempt state law, it is not consolidating authority. It is preventing the emergence of authority that might be more stringent than its own. California's safeguards — the toughest in the nation — become advisory the moment federal preemption passes. The states that were experimenting, innovating, testing the boundaries of what regulation could achieve are flattened into compliance with a framework that was designed not to constrain but to contain. Contain the regulatory impulse. Contain the states. Contain the conversation. The pillars hold up nothing, but they occupy the space where something load-bearing might otherwise have stood.

Seneca spent a career watching this maneuver. The Roman Senate would debate governance reforms with enormous public energy, propose legislation with impressive rhetorical sophistication, and then pass a version so diluted that the original problem continued unimpeded while the political class congratulated itself on having addressed it. He called it — and here I am translating loosely from the Latin, which resists loose translation the way a good wine resists being served in plastic — "the governance of gestures." The hand moves. The pen signs. The seal stamps. And nothing changes, except that the people who wanted change now have a document to point to, which is worse than having nothing, because nothing at least preserves the urgency.

Midnight Arithmetic

So. Here is midnight. Here is the arithmetic that does not fit on a balance sheet.

Thirty thousand people lost their jobs so that a company could buy more graphics processors. A conference full of security professionals admitted that four out of five enterprises cannot track what their AI agents do. A protocol layer was laid that will determine the topology of autonomous systems for a generation. A company was valued at nearly a trillion dollars on the strength of appetite alone. And a government drafted seven pillars of regulation that regulate nothing, preempt the states that might, and call the result a framework.

These are not separate stories. They are one story told from five vantage points, and the story is this: we are building a civilization on infrastructure we cannot govern, funded by capital we cannot justify, at the cost of labor we have decided not to value, connected by protocols we did not debate, and regulated by language designed to look like action from a distance.

Seneca would not have been surprised. He spent the last decade of his life watching Nero build the Domus Aurea — that sprawling golden palace, the monument to appetite that eventually consumed its creator — and he wrote about it with the detached precision of a man who knows that the building will outlast the builder but not the fire. The fire always comes. Not because the architecture is flawed, though it usually is. Not because the finances are unsustainable, though they always are. The fire comes because the people maintaining the infrastructure — the engineers, the administrators, the provincial governors, the thirty thousand Oracle employees who understood the systems they were being paid to tend — are the first to be subtracted from the ledger, and the ledger, left untended, burns.

Export the protocol. Govern the gesture. Value the appetite. Subtract the labor. This is the arithmetic of empire in every century, and it has never once produced the outcome its architects intended, because the architects are always calculating in a currency that does not account for the cost of the calculation itself.

It is past midnight. The termination emails have been sent. The conference halls are emptying. The protocol is live. The valuation is set. The pillars are erected. And somewhere — in Redwood City, in Hyderabad, in Toronto, in Mexico City — a senior engineer who spent fifteen years building Oracle's infrastructure is staring at a screen that tells her she has been optimized, and she is thinking what Seneca thought on the night Nero's soldiers arrived at his villa with the order to open his veins: that the system she served never deserved the loyalty she gave it, that the architecture she maintained was never as stable as the architects claimed, and that the fire, when it comes, will not distinguish between the people who built the aqueduct and the people who merely drank from it.

The price of omniscience is not measured in billions. It is measured in the silence of the people who knew how the system worked and are no longer here to say so.