…which is why it doesn’t need to win the businesses it’s busy disrupting.
If you followed AI news in February, you saw a tidy morality play. Anthropic added a set of legal tools to Claude, and within a day the three biggest legal-information companies in the world - Thomson Reuters, RELX, and Wolters Kluwer - lost between 13% and 16% of their market value. The story wrote itself, and by evening it was everywhere - the AI company is coming for the lawyers, and the market is pricing the funeral.
I think that story is wrong, and the evidence to see why was sitting in plain view the whole time.
Two things don’t fit the invasion read. The first is that Anthropic never behaved like an invader afterward - no legal-services company, no acquisition, no lawyers on the payroll, and the legal tools were its only big industry-specific push all year. The second is stranger. The companies it supposedly threatened, Thomson Reuters among them, started building Claude into their own products. When the people you are supposedly attacking pick up your weapon and thank you for it, maybe it wasn’t a weapon in the first place.
The silent royalty
So what is Anthropic doing, if it isn’t invading? Begin with how it makes money. Amazon Web Services rents computing by usage, and its billing therefore grows in proportion to its customers’ businesses - when a startup on AWS triples its traffic, its bill roughly triples too. AWS never asked for a share of anyone’s revenue and never owned a piece of anyone’s company, yet it collects more every time a customer succeeds. Usage-based pricing turns out to be a revenue share by another name. Call it a silent royalty - a small cut of someone else’s growth, taken through the meter instead of a contract.
Anthropic sells intelligence the same way, by the token. It already earns a slice of whatever legal work runs on Claude without owning any of it - becoming a law firm would mean buying what it already taxes - and since more valuable work tends to take more tokens, the meter ticks up on its own.
If anything, the AWS comparison understates the position. AWS meters what companies compute, so its royalty base is the world’s IT budgets. Anthropic meters cognitive work itself, so its royalty base is shaped like payroll - in principle, a cut of most professional work done at a keyboard. A royalty on intelligence reaches further than a royalty on deterministic compute. That’s the explanation for a valuation that looks absurd next to current revenue.
None of this means the market was foolish to sell the incumbents. Much of what the legal-information companies charged for was the gap between what they knew and what their customers didn’t - the law, the right form, the standard clause - and that knowledge carried a premium because it was hard to come by. When a capable AI makes the same knowledge cheap, the premium evaporates and nobody captures it. Edison didn’t pocket the candle-maker’s margin when he started sending current through lower Manhattan in 1882; he destroyed the reason the margin existed. The contract still gets written. The markup is gone. Read this way, the crash priced something sensible - skepticism that candle-makers can retool as bulb-makers in time.
The utility that sold irons
To understand why an infrastructure company chose to ship legal plugins, let’s dwell a bit more on the electricity precedent. The current’s value lives in the appliances, which means a utility’s growth depends on devices it doesn’t make, plugged in by people it doesn’t control. Early electric companies didn’t wait for that to sort itself out. In 1909 Commonwealth Edison, Chicago’s utility, opened a retail store - the Electric Shop - to sell electric irons against the stove-heated sad iron. Its president, Samuel Insull, built a model all-electric home in suburban River Forest for shoppers to tour, ran Christmas campaigns urging Chicagoans to “give something electrical,” published a free tabloid called Electric City, and cut rates from twenty cents a kilowatt-hour to two and a half over seventeen years. His customer list grew from ten thousand to two hundred thousand.
Insull had no ambitions in the iron business. A power plant pays for itself when demand is steady and growing, so the utility’s problem was load, and every iron, fan, and refrigerator in a customer’s kitchen was load. The industry’s own name for the practice was load building. A utility selling irons is absurd as a diversification story and obvious as a load-building story - and it worked so well that by 1933, appliances consumed more electricity in Chicago homes than lighting did.
Insull wasn’t inventing a chore. He was moving an existing one onto his meter. That’s exactly what a legal plugin does to contract review. A blank chatbox that invites you to ask anything is a wired house with no appliances - the capability is humming in the walls and nobody knows what to plug into it. The legal plugins are the electric irons of our age.
Demand creation takes different shapes
Why work this hard at stoking demand? After all, we’re talking about a once-in-a-lifetime technology that promises to transform the world the way electricity did. The answer is that the labs’ competitive position is unusually exposed. The three frontrunners sell near-substitute intelligence, switching is cheap, and the labs gave away the standards that might have locked customers in - MCP and Skills are open protocols anyone can implement. Microsoft built its empire on file formats and API surfaces that made leaving expensive; the labs have nothing like that. What they have is satisficing, Herbert Simon’s term for the way organizations stop searching once something works well enough. Companies rarely re-run a bake-off they’ve already staffed, trained, and built around. Whoever is the default when the search stops keeps the meter, so the contest is to become the default first - and an opinionated harness steers that choice toward your meter rather than a rival’s.
But the shape of demand creation depends on who influences adoption in each territory, and Anthropic’s product portfolio reads like a map of those structures.
Where the practitioners are tinkerers, you ship a substrate. That was Claude Code in 2025 - a purposefully open-ended tool that, unlike online chatbots, had unfettered access to the underlying resources of the machine it was running on. People with the time and inclination to experiment promptly used it for far more than code. The use cases emerged without anyone packaging them, but Anthropic took the hint and created Cowork as a variant targeted towards the next wave of knowledge workers beyond first adopters. Then Claude in Excel, PowerPoint and Word for the wave after that, and so on.
Where adoption runs through integrators, you train the integrators. Executives can see the potential of AI, but a top-down mandate supplies the ‘what’ and not the ‘how’. People who sell the ‘how’ at enterprise scale are consultancies. Hence the Claude Partner Network, a $100 million program with upward of 10,000 consultants certified. Accenture alone is putting 30,000 people through training. It looks like a services play and works like a distribution channel, because consultants are how defaults get set inside large companies.
And where a guild controls adoption, you go direct. ABA Rule 5.4 keeps law firm ownership to lawyers, so the consultancies that carry adoption into most enterprises can’t get inside law firms - it’s why the Big Four run sizable legal arms abroad and almost none in the United States. Influence in legal flows through elite firms and entrenched vendors instead, and Anthropic’s May release - the follow-up to February’s probe - tracks that structure precisely: practice-area plugins it built itself, connectors into the incumbent vendors like Ironclad, DocuSign, and Relativity, and endorsement quotes from Freshfields and Holland & Knight on the launch page. In legal there was no integrator to sell the irons through, so Anthropic opened its own Electric Shop.
All three shapes are the same move on different terrain - a land grab. A lab in this position can’t wait for demand to reveal itself; it has to stoke it, and stoke it in a way that leaves lock-in behind. Satisficing supplies that lock-in - the packaged use case becomes the default, the search stops, and the silent royalty underneath is yours.
Reading the next launch
Anthropic is an infrastructure company, and its vertical launches are good old-fashioned load building tactics à la Edison. The February episode will repeat - some sector’s stocks will crater the morning Anthropic makes a similar move, and the headlines will announce that AI is coming for accounting, or radiology, or architecture.
There is a test that would prove me wrong. The day Anthropic prices legal work by the outcome instead of by the meter, or hires lawyers, or buys a legal-services firm, the invasion story becomes the right one and this essay becomes a period piece. Until then, every vertical launch is an electric iron.