The Price of Omniscience: Oracle's Purge, the Governance Gap, and the Protocol Nobody Noticed
Oracle fired thirty thousand people this week, and the arithmetic behind it is worth understanding — because it is the same arithmetic that will reach your organization within eighteen months.
The termination emails began arriving at six in the morning, staggered across the United States, India, Canada, and Mexico. Eighteen percent of the workforce. TD Cowen ran the numbers and landed on annual savings between eight and ten billion dollars. That money is not going to shareholders. It is going into data centers, GPU clusters, and AI infrastructure: fifty billion dollars of capital expenditure planned for 2026 alone, with total AI commitments approaching a hundred and fifty-six billion.
This is not a company in trouble. Oracle posted strong numbers. This is a company executing a deliberate balance-sheet conversion — operating expense in the form of salaried humans, traded for capital expenditure in the form of silicon. The market read it exactly that way and the stock ticked up.
The Exchange Rate
What is worth being precise about is the ratio, because the ratio is the actual news. Thirty thousand livelihoods convert to roughly nine billion dollars a year of purchasing power, and nine billion dollars buys a great deal of compute. The analysts called it "disciplined" and "strategic reallocation." Both words are accurate. Neither is an argument.
Here is the part I have not seen discussed honestly. The people Oracle cut were disproportionately the ones who maintained the systems that already work — the legacy database estate, the support organization, the provisioning and ops layers that generate the revenue funding the GPUs. The company is financing its future infrastructure by subtracting the people who understand its current infrastructure. That trade can be correct. It is rarely modeled honestly, because the cost shows up two years later as incidents nobody can diagnose, and by then it is attributed to complexity rather than to the decision that caused it.
I have watched smaller versions of this at close range for fifteen years. The pattern is always the same: the savings are booked immediately and precisely; the costs arrive late, diffuse, and unattributed. If you are on the receiving end of a mandate like this, the useful question is not whether to cut. It is which specific knowledge you are about to lose, and whether it exists anywhere other than in the heads you are removing.
Four Out of Five
While Oracle was emptying its offices, RSAC 2026 was hosting a session on agentic AI governance that produced the most alarming number I read all month.
Only twenty-one point nine percent of organizations currently treat AI agents as identity-bearing entities.
Let that settle. Four out of five enterprises running autonomous software in production have not given those agents a distinct identity. No identity means no independent audit trail, no ability to revoke an agent's access without revoking the human's, no way to attribute an action to a specific agent, no behavioral baseline to detect drift against. The agent acts with a human's credentials, and the logs say the human did it.
This is not an abstract governance concern. It is a concrete engineering defect with a concrete remediation, and the remediation is unglamorous: give every agent its own service principal, scope it to least privilege, log its tool calls separately, and set an expiry. None of that requires a platform purchase. Most of it is a week of work that keeps getting deprioritized because nothing has broken yet.
The broader conference consensus was that enterprise security controls are "largely inadequate" for autonomous AI — not lagging, not catching up, inadequate — and that the field has moved from theoretical risk assessment to observed infrastructure-level threats, including multi-agent offensive behaviors. Those are not whiteboard scenarios from a red-team deck. They are patterns in deployed systems.
ServiceNow's pitch in that session was sensible: distinct identities, least-privilege access, audit trails, behavioral monitoring, a control plane that can trace every action an agent takes and enforce policy in real time. The demo showed a thirteen percent improvement in mean-time-to-resolution. The pitch is right. It is also downstream of a decision most organizations have not made — which is whether an agent is a user, a service, or a piece of infrastructure. Until you answer that, no control plane will save you, because you cannot govern an entity your identity model does not contain.
The Protocol Nobody Noticed
Microsoft shipped native MCP integration in Sentinel this week, alongside a Claude connector in public preview. It got perhaps a tenth of Oracle's coverage. It will matter more.
MCP — the Model Context Protocol — is the standard Anthropic built for connecting agents to external tools and data sources. Think of it as the interface layer for agentic software: any agent can talk to any tool, provided both speak it. Microsoft embedding it natively into its enterprise security platform means the two largest AI infrastructure providers have quietly agreed on the plumbing. Not on models. Not on strategy. On the plumbing.
Plumbing outlasts everything built on top of it. HTTP is older than most of the engineers using it, and the average website it serves lives about as long as a hamster. When two competitors converge on a protocol layer, they are not making a product announcement — they are laying pipe, and pipe determines where the water flows for the next decade.
The practical consequence for anyone building right now: your integration surface is being standardized whether you participate or not. If you have hand-rolled agent-to-tool plumbing, you are maintaining a private dialect of a language that is about to become universal, and you will pay to translate later. The governance conversation at RSAC — identity, least privilege, behavioral drift — is not wrong. It is a conversation about what to build on the pipe, held by people who did not notice the pipe was just laid.
The Valuation
OpenAI closed a hundred-and-twenty-two-billion-dollar round at a post-money valuation of eight hundred and fifty-two billion. SoftBank, Andreessen Horowitz, and Fidelity signed the checks. The company generates roughly two billion dollars a month in revenue, which sounds impressive until you put it next to what it costs to generate.
Six months ago the same company shut down Sora, its most visible consumer product, because the burn approached a million dollars a day and Disney walked away from a billion-dollar partnership. Now it is preparing an IPO.
Eight hundred and fifty-two billion is not a valuation of what has been built. It is a valuation of what the market believes demand will eventually absorb, discounted by nothing in particular. That may turn out to be right — I am not predicting a collapse, and people who do have been wrong for three years running. But if you are making architectural bets that assume today's inference prices are permanent, understand what is holding those prices down. It is capital, not efficiency. Capital changes its mind faster than silicon does.
The Regulation That Doesn't
A legislative thread runs through all of this: a discussion draft from Senator Blackburn, a White House framework built on seven pillars, California signing the toughest state-level safeguards in the country, and the EU grinding toward its August deadline while three-quarters of European enterprises admit they are not ready.
The seven pillars: protecting children, safeguarding communities, respecting intellectual property, preventing censorship, enabling innovation, workforce development, and federal preemption of state law. Read that list again and notice what is missing — any mention of the concentration of economic power that makes the regulation necessary in the first place.
Federal preemption is the tell. When the federal government moves to preempt state law, it is not consolidating authority so much as preventing the emergence of stricter authority elsewhere. California's safeguards become advisory the moment preemption passes. The states that were experimenting get flattened into compliance with a framework designed to contain the regulatory impulse rather than to constrain the systems.
For practitioners, the operational read is simple: do not build your compliance posture around the federal framework. Build it around the strictest regime you plausibly touch — today that is the EU AI Act and California — because those have enforcement mechanisms and deadlines, and retrofitting governance into a shipped agentic system costs several times what building it in costs.
What Actually Connects These
Thirty thousand people were cut to fund GPU purchases. Four out of five enterprises cannot tell you what their agents did last night. A protocol layer was standardized that will shape autonomous systems for a decade. A company was priced near a trillion dollars on projected appetite. A government drafted seven pillars that regulate very little and preempt the states that would.
These are one story from five angles: we are building on infrastructure we cannot yet govern, funded by capital that assumes the governance will arrive later, at the cost of the people who understood the last generation of systems.
None of that is an argument for standing still. It is an argument for knowing which of those five forces you are actually exposed to. If you run agents in production, the identity gap is your problem this quarter, and it is cheap to fix. If you are building integrations, the protocol convergence is your problem this year. If you are planning capacity against current inference prices, the valuation is your problem eventually.
The layoffs are the loudest story and the least actionable. The identity number — twenty-one point nine percent — is the quietest and the one you can do something about before Monday.