Break While You Build

What Time It Is, Part III

Data centers this, data centers that. Every week produces another headline about grid strain, water draw, speculative overbuild, debt-financed hyperscale expansion, transformer shortages, local permitting fights, and politicians suddenly discovering that the “cloud” is not a metaphor but a building full of machines plugged into a power plant.

Fine. Good. Let it burn a little.

Asking whether all this is “good” or “bad” already concedes the field. That is spectator talk. A development arrives, obviously real, obviously large, obviously rearranging the material landscape, and the only available response is to stand at a safe distance and render a verdict over it, as though history were a restaurant dish that can be sent back to the kitchen for being overcooked.

History offers no such service. The buildout is here. The compute is here. The models are here. The capital has already moved. The state will posture, the utilities will panic, the locals will object, the consultants will multiply, the debt will wobble, and the machines will keep arriving anyway. The operative question is simpler: How can these developments be used in our favor?

That question continues the prior argument.

The first piece in this series argued that there is no coherent national “scene” from which a political future can simply be recovered. The second argued that meaningful productive entry into much of American life has been enclosed by charters, licenses, inspection regimes, permitting timelines, zoning, compliance overhead, and incumbent protection. The practical effect is simple enough: there are a great many things one is constantly told to build, almost none of which one is actually permitted to build at meaningful scale.

So the next question was always going to be: Where, precisely, has the wall cracked?

Artificial intelligence is one answer, though in a narrower and more material sense than the usual chatter allows. Leave aside “start an AI company,” “prompt engineering,” and the usual parade of men with haunted eyes announcing the future of productivity from a stage made of LED panels and venture debt. The real opening is more concrete.

At the individual level, AI adoption has been almost frictionless. A writer, programmer, analyst, researcher, operator, or mediocre middle manager can sit down in front of a model and get immediate leverage. A single person can couple judgment to action very quickly. The machine drafts, searches, summarizes, compares, reformulates, translates, organizes. Even when it fails, it fails in a way that produces useful pressure on the user’s own thinking. The individual can use it because the individual can decide to use it.

The solitary user already possesses the one thing the enterprise withholds from itself: immediate discretion. He can decide that a draft is good enough, that an answer is useful enough, that the machine may be trusted for this small next move. He risks his own embarrassment, his own time, maybe his own job. The blast radius remains human-scale. The institution is haunted by scale. Every workflow touches another workflow. Every record mutates another record. Every convenience becomes precedent the moment it works twice.

Inside enterprise systems, the same tool suddenly meets jurisdiction. Value appears only when the model reaches from language into records, approvals, payment status, inventory, claims, scheduling, code deployment, procurement, maintenance logs. Every useful crossing touches a boundary someone already owns. The model hovers over the institution like a clerk without keys. It can see enough to tantalize and never enough to govern the thing.

That blockage only looks technical at first glance. For thirty years the valley has sold the same catechism: cybernetics, markets, countercultural incense, and the promise that history will sort itself out if enough tools are shipped. Product replaces politics. Choice among platforms replaces argument over ends. The pitch remains libertarian at the surface and managerial underneath. Everybody gets to feel emancipated while somebody else quietly fixes the permissions. The anti-state music still plays even when the whole performance depends on subsidies, research grants, procurement, easements, tax deals, grid upgrades, and friendly sovereign power.

The same mechanism leaves the rest of the economy feeling dead from the inside. The large firm is an accumulated political settlement among departments, vendors, databases, legal teams, compliance offices, procurement rules, internal empires, overlapping chains of custody, and managerial habits nobody entirely understands but everyone has learned not to disturb. A bureaucracy is a memory machine with stamps. The model enters a bureaucracy, not a frontier.

Once a model has been threaded into that environment, vocabulary hardens into infrastructure. “Approval,” “exception,” “customer,” “risk,” “ship,” “escalate,” “safe,” “authorized” cease to be descriptive terms. They become gates. Whoever fixes the schema fixes the range of admissible action. Legal, security, IT, procurement, compliance, HR, and middle management are fighting over operational meaning, over which actions will exist for the system at all. What looks like workflow design is often a machine-readable ordering of what can count within the firm.

Here lies the enterprise AI puzzle. Intelligence can be rented by the month. Agency must be granted by an institution that no longer trusts its own interior. Every serious use case therefore arrives as a jurisdictional dispute.

That dispute is rarely staged honestly. The official language is always narrower than the real conflict. One hears about governance, safety, change management, data quality, acceptable use, model risk. All real enough. Yet beneath those proper administrative nouns sits a cruder problem: no one wants to discover, in a way that can be audited later, how much of the institution still depends on undocumented habits, tolerated workarounds, personal memory, and people who know where the body is buried in a spreadsheet nobody was ever supposed to use as a system of record. AI is threatening partly because it is powerful, but more because it is illuminating. To make the model useful, the firm must expose itself to itself. That exposure can be more frightening than the model’s mistakes.

The result is predictable. Enterprise AI gets defanged into a sandbox. It assists, suggests, summarizes, decorates, drafts talking points, generates the sort of bloodless internal memo no human should have had to write in the first place. The line is drawn at action. Firms announce “AI integration” while carefully ensuring that nothing structurally important is permitted to happen. The Californian sermon promises liberation. The actual delivery mechanism is another layer of controlled access.

This should sound familiar, because it is the internal analogue of the external permission economy described in the previous essay. There, the barriers were charters, licenses, federal inspection, zoning, permitting, capital requirements, compliance architecture, and incumbent capture. Here, the barriers are permissions, audit rules, system fragmentation, liability exposure, vendor lock-in, records obligations, semantic drift between departments, and the fear of making legible just how little command the institution actually has over itself.

Same disease. Different scale. The old ideology promised that technology would dissolve politics into networks and markets. In practice it carried the fight over power deeper into the machine.

Spectator questions miss the point. The builder’s question is narrower and more serious: Where does this institutional friction generate a market that did not exist before?

Large institutions cannot simply hand authority to a model and hope for the best. If the AI misprices risk, violates a contract, leaks customer data, fabricates a compliance answer, discriminates in hiring, approves the wrong payment, miscites a policy, gives bad medical guidance, or writes a workflow change that later becomes evidence in litigation, someone must be responsible. Operationally. Legally. Financially.

This creates fear inside the enterprise, but fear is often just a market waiting to be named properly. The sales pitch around AI depends on a cultivated amnesia about institutions: every inherited procedure is treated as deadweight until the first bad output, the first bad trade, the first fabricated note, the first compliance failure, the first lawsuit. Then everybody rediscovers that old arrangements, however ugly, were carrying accountability somewhere in their wiring.

The new entrant can enter through the liability layer. He does not need to be the foundation-model company, own the whole enterprise software stack, or displace the incumbent’s core business. He makes AI action legible, auditable, reversible, permissioned, insurable, and limited to the domains where a human institution can actually tolerate delegated machine behavior.

That means audit trails. Review systems. Permission layers for agents. Escalation chains. Model-behavior logging. Risk-scored automation. Compliance wrappers. Chain-of-custody records for AI decisions. Warranty structures. Sector-specific deployment standards. Unglamorous work. Real work.

Liability does not merely slow AI adoption. Liability creates demand for an intermediary class capable of translating raw model capability into institutionally acceptable action.

This is where a great deal of stupid conversation dies. Men keep asking whether “the AI” will replace lawyers, claims adjusters, paralegals, compliance analysts, underwriters, schedulers, coders, case workers, procurement staff, as though historical change arrives by one object displacing one profession in a neat row of dominoes. What arrives first is a reorganization of supervision. Before a machine takes the chair, it changes who has to sign, who has to review, who owns the exception queue, who must explain a bad outcome to someone with subpoena power. The labor shift begins as a liability shift. That is one reason the implementation layer has room to exist.

Every large organization is a ruin pretending to be a system. Legacy software. Half-finished migrations. Custom databases nobody wants to document. SharePoint graveyards. Vendor APIs with ancient authentication schemes. Local spreadsheets functioning as shadow sovereigns over departments that officially deny their existence. Email chains standing in for process. Ticketing systems disconnected from records systems. Records systems disconnected from billing. Billing disconnected from operations. Operations disconnected from reality.

An AI system is only as useful as the systems it can safely touch. The implementation problem is straightforward to state: How do you build connective tissue between intelligence and institutional action without blowing a hole in records control, legal exposure, or operational continuity?

There sits the business.

The entrant who solves it is selling orchestration, translation, mapping, normalization, retrieval, connector architecture, workflow repair, and actionability. He is building the narrow bridge between a model that can reason and an institution that cannot move.

At this point ontology re-enters by the side door. The problem is not merely that one system cannot talk to another. The problem is that they do not agree on what the objects are. What one database calls a customer, another calls an account holder, another calls a party, another calls a user, another calls an entity, another calls a beneficiary; another silently splits into household, contact, legal person, device ID, tax record, and billing relationship. A workflow is not jammed only because the pipes are disconnected. It is jammed because the nouns are unstable. Any firm that can stabilize the nouns long enough for action to travel across them is doing more than integration work. It is imposing a temporary world inside the enterprise, a practical ontology of what exists there and what may be done to it. That is why these boring connective businesses can become strategically important so quickly.

This also explains why so many enterprise demos feel fraudulent. They present the machine as though it were floating above the firm in a layer of immaculate intelligence, untroubled by nomenclature, entitlement conflicts, stale records, duplicate entities, contradictory clocks, or the ordinary administrative ugliness through which large organizations actually know themselves. The ugly layer is the decisive one. Intelligence without alignment to records, permissions, and categories remains ornamental. The market opening lies where ornamental intelligence becomes boringly reliable enough to touch the file.

The opportunity sits here. The enclosed economy still makes frontal assault difficult. You are not just casually going to start a bank, replace a hospital chain, rebuild domestic manufacturing from a garage, or out-charter the chartered. The better path runs by adjacency. Leave the charter where it is and build the AI compliance, fraud-review, underwriting-support, document-processing, or back-office coordination layer the bank cannot construct without help. Leave the hospital chain where it is and build the workflow, records, scheduling, claims, triage, or liability-management layer it cannot safely improvise. Leave the factory floor where it is and build the procurement, inventory, maintenance, supplier-intelligence, or quality-control layer that lets actual productive intelligence touch operations. Leave the insurer in place and build the review, comparison, documentation, or exception-handling layer that lets machine reasoning operate inside a regulated environment without detonating it.

Adjacency has another advantage: it lets a small entrant accumulate the one thing incumbents can no longer fabricate on command: situated operational knowledge. The bank knows banking, yes, but often not its own process architecture except as inherited sediment. The hospital knows medicine, yes, but not necessarily where scheduling, claims, notes, billing, triage, consent, and follow-up break against each other in the daily weather of the place. A focused entrant working in the seam can learn the seam faster than the institution can relearn itself. That knowledge compounds. By the fifth or sixth deployment, the entrant has something much closer to a transportable method than a one-off service engagement. That is how a small intermediary becomes hard to dislodge.

Call it entry by adjacency rather than frontal assault. The cartel state is very good at preventing new actors from replacing incumbents directly. It is, at least for the moment, much worse at preventing new actors from becoming necessary to incumbents who can no longer metabolize change on their own.

There is the crack in the wall.

And it is historically unusual because the same complexity that protected incumbents for decades is now turning against them. For a long time, institutional sclerosis could function defensively. Slow processes, compliance overhead, vendor lock-in, internal complexity, fractured permissions, and bureaucratic obesity all worked as moats. New entrants were kept out because the terrain was too expensive, too slow, too opaque, or too regulated to cross cleanly.

AI changes the polarity of that complexity.

Now the incumbent’s moat is also a prison. He has the data, but not in usable form. He has the workflows, but not clearly mapped. He has the customers, but not the flexibility to reorganize around a new productive tool. He has software contracts, but not interconnection. He has compliance departments, but not liability architecture for AI action. He has budget, but not agency.

So for the first time in decades, the enclosed economy actually needs help from outside itself.

Read the hysteria around data centers carefully. The panic proves that the technology has reached the real economy. Once a thing starts demanding land, power, cooling, substations, permits, legal review, public resistance management, workflow redesign, liability structures, and new organizational forms, it has ceased to be a novelty. It has become an industrial fact.

The naïve observer sees friction and concludes failure. The more serious observer sees friction and asks where the bottlenecks will produce a new class of firms.

The social question returns here.

The relevant unit for this opening is probably a small group. Giant firms move too slowly. The isolated genius breaks too easily. The work described here requires trust, discretion, speed, technical competence, tolerance for ambiguity, and the ability to move inside broken institutions without becoming trapped in their brokenness. It requires people who can see a workflow, a records chain, a liability boundary, an access problem, a permissions map, and an implementation failure as one connected terrain rather than as separate departmental concerns.

That kind of action is usually taken by small groups before it is taken by large organizations. A few serious people who trust one another can often move through a frozen institution more effectively than a hundred internal stakeholders who do not. A compact group can stay close to the actual problem. It can identify the exposed gate, the unguarded supply line, the process nobody inside the castle understands anymore, and it can build exactly enough connective tissue to make the thing move.

This is one reason the club argument from the earlier essays returns with more force here than in most ordinary business writing. Enterprise software people love to pretend that talent is a market like wheat. Post the role, run the process, sort the candidates, grant the options, and capability arrives. Sometimes it does. More often, the decisive work requires preexisting trust of a kind no hiring funnel can manufacture quickly. Someone must be willing to let another man touch a sensitive workflow, map an ugly process, speak plainly about hidden dependencies, and fix a thing without using the occasion to build an empire of bullshit around it. That is not just technical trust. It is moral trust, or near enough.

The market recognizes this only after the fact. Then the language changes: founder-market fit, domain intimacy, embedded teams, technical depth, execution culture. A swamp of euphemism. What they are circling is older than venture capital: a compact formation of people who know one another’s strengths, keep confidence, distribute spoils in a way the group can bear, and retain enough shared memory to move fast without constant internal litigation. That is the startup form when it ceases to be theater.

Call it a startup if you need the modern word. I care more about the form. The form is older than venture capital and more durable than most corporations: a small, trust-bound body organized around competence, risk, shared upside, and repeated proof of reliability. No mass coalition. No national faction. No scene. A band.

The club form earns its keep here. The club is where operational trust accumulates before the market names the need. It is where men learn who can actually do what, who keeps confidence, who folds under pressure, who can explain a technical system without lying about its limits, who can enter a room full of institutional nonsense and come back with the one real lever. That kind of knowledge does not appear on a credentialing rubric. It accumulates through repeated contact.

The state can regulate a charter, a permit, a plant, a license, a filing, a securities offering, a payroll tax account. It has a harder time regulating the prior social technology out of which competent action emerges: a small body of people with shared interests, growing trust, operational memory, and enough seriousness to recognize when a wall has cracked.

Which returns us to the beginning.

“Data centers good or bad” is unserious because it treats a strategic development as a morality play. History does not require our approval. It requires that we notice where a channel has opened.

The model companies will do what they do. The utilities will do what they do. The state will do what it does worst. The incumbents will resist, posture, procure, delay, and eventually capitulate in the most expensive way available to them. None of that is our decision.

Our question is narrower. Where is intelligence trying to become action, being blocked by liability, interconnectivity, permissions, and institutional fear? What is the smallest competent formation that can enter that gap and make itself necessary? What can be built in the cracks before the cracks are paved over?

That question is worth asking now.

Look what time it is.