There's a scene in There Will Be Blood I keep returning to. Not the obvious one. The earlier one, on the ridge, where Daniel Plainview is reading land that looks like nothing to everyone else. The slope of the strata. A cadence in the silence. A shape nobody else sees. Twenty years of looking. His body knows things the industry doesn't.
Anderson painted from a type.
The type is still working.
What's happening in AI right now is the same kind of landscape. Dense with signs most people aren't reading yet. The chip story is real. Nvidia wins. The hyperscalers either win or get eaten. Fine. That's two signs. There are more.
Five, actually. The first three change how you see the boom. The last two change what you do about it.
A hundred years after Plainview, Chase Lochmiller stood on a different piece of Texas. Abilene, population 120,000. Flat, windy, not on anyone's data center map. Wind farms had been overbuilt across that corner of the grid. Tax credits paid developers to produce power whether anyone wanted it, and the transmission lines couldn't carry the power out. Prices went negative. The electrons were stranded, screaming for a buyer.
While everyone else put data centers in Northern Virginia, Chase went to Abilene. He signed a deal for two buildings and a substation that would eventually carry a gigawatt. The power draw of Denver. Eighteen months later, eight buildings are on that ground. A 350-megawatt natural gas plant. Nine thousand workers on site in a town of 120,000. A 5,000-car parking lot full every day. The campus powers what Oracle and OpenAI are calling Stargate, the largest AI compute cluster in the world.
Chase read the land. The majors didn't.
That has always been the operator's move. The Plainviews. The wildcatters. The people who saw something in the dirt before the industry knew where to look. The oil boom minted Exxon. It also minted the Plainviews. One got the headlines. The other got the ground. AI is the new oil boom, and the Plainviews are reading a different set of signs this time.
The capex chart everyone is looking at says hyperscaler AI spend will cross $650 billion through 2026. Bigger than the Manhattan Project. Bigger than the interstate highway system. Second only to US defense as a concentration of capital in one direction.
Most analysis of that chart is about the chips. Chase's is different. He said it at Stanford last month. He walked through Cobb-Douglas, the production function that says GDP growth is a function of labor, capital, and technology. Labor has always been the slow variable. You can't just buy more of it. You wait twenty years for more humans to grow up. Capital moves fast. Technology moves fast. Labor is stuck at birth-rate speed.
Except now. A data center is a factory that stamps out workers. A GPU cluster is a labor force that didn't exist before you bought the chips. His line: "For the first time in history, you can change delta-L without waiting twenty years."
$650 billion is the largest hiring spree in history, disguised as a concrete pour. Every hyperscaler CEO has been saying some version of this for a year. Nadella. Pichai. Zuckerberg. Huang. Agents. Digital workers. Synthetic employees. Pick the flavor.
Now look at the cost stack per megawatt. About $60 million all-in. Around $20 million goes into the building. Steel, concrete, chillers, electrical. Around $40 million goes into chips and networking. And $4.7 million of that $60 million is trades. Electricians. Welders. Pipe fitters. Concrete guys pouring at two in the morning. Chase has nine thousand people on the ground holding this thing together, and his biggest bottleneck isn't compute or power. It's hands. "We don't have enough electricians. We don't have enough welders. We don't have enough plumbers."
The industry spending $650 billion to replace human labor is spending a quarter of its construction cost on human hands. You cannot build an empire of synthetic workers without calluses.
Everyone knows AI is coming for white collar work. The useful question is which white collar work, and from which direction.
AI eats the middle. Pattern work. Scheduling, dispatching, invoicing, summarizing, routing, first-draft-anything. The clipboard part of every profession. The part a competent person can do after reading a manual. AI is going to chew through that across every industry that has a middle. That part isn't a debate.
The ends are different.
On one end: the substrate. Nvidia, the chip designers, the scaled training infrastructure. That layer keeps compounding.
On the other: the thing the customer actually buys. The moment someone shows up at your grandmother's house to replace a valve at 11pm. The fifteen seconds when a doctor finishes the exam and decides whether to tell you the worst or schedule a callback. The minute a teacher notices a kid has stopped talking and stays five minutes late to ask why.
Coordination is the right word for the first end. For the second, we'll call it service, though service is a worn-out word. Here is what I mean: the version of work that requires a self. Someone who can be affected by the situation in front of them. Someone whose attention has a cost and who spends it anyway. The plumber who stays an extra hour because the water is still leaking somewhere he can't see. The nurse who answers the call bell before the light finishes its second blink. The lawyer who rewrites the paragraph she isn't being paid to rewrite because it isn't right yet.
You have seen the absence of it on the customer side. You have been routed through the AI chatbot that couldn't route you. You have watched the automated loop send you back to itself. What irritated you was the absence of a self on the other end.
AI can't do service because service runs on self-abandonment. You have to have a self to give one away. The chatbot doesn't. The agent doesn't. They can simulate the language of service. The cost of producing it is beyond them.
The deeper objection is that simulation will close the gap. Character.AI and Pi produce user attachment today. In five years, the argument goes, the chatbot will console you better than the nurse. Maybe in low-stakes contexts. The thesis fails at the moments when the stakes force the customer to check whether the presence is real. Grief. Pain. Fear. Money on the line. A child not breathing right. A parent losing a long fight. Those moments strip the interaction to one question. Is there actually someone here with me. As synthetic presence becomes abundant, the verification premium on real presence goes up. Time to Humanity named that dynamic. This is where it gets paid.
Cybersecurity already ran the experiment. Companies used to run red teams of a hundred-plus humans trying to break their own systems. AI changed that. One engineer now runs the first pass against every known attack vector overnight. The field didn't shrink. It concentrated. The humans still in the room are doing the creative precision work, the what does an actual adversary try that isn't in any playbook yet work. AI does recall. Humans do precision. Value-per-head went up.
The same split is coming to every service industry with a coordination layer and a presence layer mashed together.
| Industry | Coordination layer (AI eats) | Presence layer (survives) | Split applies? |
|---|---|---|---|
| Home services (HVAC, plumbing) | Dispatch, scheduling, diagnostic intake, billing | The technician who actually shows up and listens | Yes |
| In-home eldercare | Care plan logistics, compliance documentation, scheduling | The aide who sits at the bedside at 3am | Yes |
| Veterinary care | Scheduling, diagnostic imaging review, inventory | The vet who tells you your dog isn't coming home | Yes |
| Tax preparation (commodity) | The entire workflow, end to end | There isn't really one | No |
Wherever the customer pays for presence, AI handles recall and humans own what's left. Where they don't, AI eats the whole thing.
The obvious objection: coordination automation historically collapses price. Thumbtack, Angi, HomeAdvisor. Every platform that consolidated fragmented home services commoditized the operator and transferred savings to the customer or the platform. Technicians stayed cheap. No presence premium.
The pattern is real. The cause is different.
The platforms forced operators into price competition because the operators didn't own the customer. Savings went to the marketplace. Commoditization came from the platform structure.
Look inside operator-owned businesses that automated their own coordination. PE has been rolling them up for a decade. Wrench Group. Service Experts. ARS/Rescue Rooter. Southern Home Services. Home Brands. The platforms running smart diagnostics and predictive-maintenance subscriptions post 15 to 25 percent EBITDA margin expansion over two to three years. Multi-trade consolidators command 30 percent higher lifetime value than single-trade. Subscription revenue runs 28 percent of total at top-quartile firms. The savings showed up as margin because the operator kept them.
Veterinary has the same sign. Mars paid $9.1 billion for VCA in 2017. JAB Holdings owns NVA. KKR took VetStrategy private. Between them they run hundreds of clinics on tech-enabled practice management. Vet fees have climbed 6 to 8 percent a year since, well above general inflation at 3.1 percent. Customers paying more for the consolidated category.
Direct primary care is the cleanest version. Ninety dollars a month, paid as a subscription. 413 patients per doctor against 2,500 in a traditional insurance-based practice. Fewer patients. Higher fee. More time with each one. Revenue per patient lifts because the subscription replaces insurance reimbursement. Margin lifts because the billing overhead is gone.
The common structure: the operator owns the customer, uses technology to strip coordination cost, redirects the freed margin into more time per customer. The thesis is already running in real businesses. The signs are on the ground.
The pattern scales to B2B. The AI majors will do the infrastructure layer at trillion-dollar scale. But the work of turning raw model output into usable product (evaluation, fine-tuning, domain-specific data, quality at the edges) lives closer to the use case than the majors can reach. They'll try to absorb it. They'll fail. Every extraction boom produces a service layer the majors cannot eat.
In 1926, two French brothers named Conrad and Marcel Schlumberger started running electrical logs on oil wells nobody else could read. A hundred years later the company that bears their name is worth $50-plus billion, and the oil majors still have to hire it. Exxon tried to absorb that layer. Shell tried. BP tried. None of them managed. The service company lives closer to the physical situation at the well.
Each boom rhymes with the last in a way the business press mostly misses. The obvious part is capital deployed fast and trillion-dollar winners. The less obvious part is what gets exposed in the wake.
Edwin Drake hit oil in Titusville, Pennsylvania in 1859. By 1911 Standard Oil was being broken into 34 companies that became Exxon, Chevron, Mobil, and a dozen more worth trillions combined. In between, oil didn't just build an industry. It exposed the parts of the economy held together by muscle and timing. Whaling collapsed because kerosene was cheaper. Railroads restructured around refined fuel. Shipping moved from sail to diesel. Chemicals got invented. When the reorganization finished, you could see what had been load-bearing all along and what had only been just-getting-by.
AI is doing the same thing at a different level. It exposes the part of the economy held together by human presence pretending to be coordination. Most of what was called customer service was coordination. Forms, scripts, escalations, handoffs, hold music. That was always legible to software. The part that was actually service, the part that required a self, is being surfaced for the first time as its own category. You can see the outline because AI is eating the thing next to it.
Service is the only asset AI cannot synthesize. To synthesize it, you would need a self to give away.
When customers feel the difference, the business can charge for it. Here are the marks where the four signs cluster.
-
i
Physical presence that can't be faked remotely. Someone has to be in the house. In the operating room. On the actual site. The smart thermostat can flag a failing furnace. It can't pull the old unit out of the cabinet and hand the homeowner a quote in the driveway.
-
ii
Emotional stakes at the point of contact. Healthcare. Eldercare. Legal under duress. Veterinary. Funeral services. Home repair in crisis. Nobody googles best chatbot for a pet that is dying. They want the vet in the room.
-
iii
Trust that transfers through a specific person. Not a brand. Not a platform. I have a guy is the sign. I won't see anyone else is the sign. A technician leaving and taking the book of business with them is the sign.
-
iv
High exception density. Seventy percent pattern, thirty percent judgment. AI handles the pattern. Humans own the exception: the basement flooding at midnight, the rash that might be meningitis, the estate where one sibling wants to sell and three don't.
-
v
A fragmented category where no one has claimed "actually good." Home services, dental, home health, small business accounting. Sectors where customer expectation is low because everyone has been mediocre for a generation. The first operator to claim the quality slot owns an outsized share.
-
vi
Repeat-customer economics. Lifetime value is the unit. Five years of a good plumber is worth more than fifty one-offs. Churn kills the model. Retention compounds it.
The ratio ties them together. Presence at roughly 20 percent of labor hours, 80 percent of what the customer experiences. Outside that range, probably not the right hunt. Presence as the whole product means nothing to automate. Presence as incidental means nothing to protect.
The operators who read the new signs will own ground the majors can't drill.
Where the thesis doesn't fire: pure information work. Commodity manufacturing. Fully remote service where the human is a voice on a phone that could be swapped for a better voice. License-locked labor markets where savings can't redeploy into more humans per customer. Ultra-high-end luxury, where presence is already priced and abundant.
That is the Plainview map in 2026. The land is often cheap. The operator who uses AI to strip coordination and concentrate presence will find themselves on ground the majors cannot drill.
The first-order effects are the capex story everyone is writing. The second-order effects are the operator story, the roll-ups, the margin expansion, the new brands in categories that have been mediocre for a generation. Those are already visible.
The third-order question is the one nobody is asking. What happens when the status hierarchy inverts. The LSAT built an elite funnel toward legal abstraction. H-1B visas prioritize software engineers over electricians. The four-year college economy is staked on the premise that you climb out of physical work, not toward it. The average age of a U.S. electrician is north of 50. All of those structures assume the premium lives where it used to. None of them are ready for what happens if it doesn't.
Somewhere tonight a 52-year-old plumber is updating his dispatch software at a folding table in his garage. His customers will never see that. They will see that he shows up faster. That he listens to the basement for ten seconds and names the joint that's leaking. That his hand on a copper line tells him whether the system is still under pressure. That he stays the extra hour the silence after the repair asks for.
His body knows things no model can name. A hundred years later, the story rhymes.
Same instinct.
Different signs.