How OpenAI and Anthropic Take Over the $7TN Professional Services Market

OpenAI owns part of the company buying up America's accounting firms. It was the first investor in Harvey, legal's biggest AI company. Its engineers already work inside law firms, banks and hospitals. And when Anthropic released a legal plugin for Claude at the end of January, roughly $285 billion came off software, data and professional services stocks in a matter of days.

The labs are not selling to professional services anymore... they are moving in. The first half of this edition is about how, because there are exactly three ways to do it and two are already running. The second half is about what happens to the rest of us if this goes unchecked... because the endgame here is not a product launch. It is market power.

Professional services... legal, accounting, consulting, IT services, engineering, advisory... is a roughly $7 trillion market globally. IT services alone is around $1.5 to 2 trillion. Legal is about a trillion. So is consulting. Accounting, which I mapped last week, is $650 billion.

Now put that against the entire global software market: roughly $900 billion. Every SaaS company ever built is fighting over a pot seven times smaller than the work that surrounds it. And for the labs specifically: OpenAI's revenue this year is measured in tens of billions, selling into markets measured in trillions. Every board in San Francisco can do that division.

I have spent two editions mapping who fights for these markets... the tools, the AI-native firms, the incumbents, and in accounting, the buyers. This edition is about the player standing behind all of them: the one selling the intelligence everyone else runs on, which has started keeping the work for itself.

OpenAI holds equity in Thrive Holdings, the vehicle rolling up CPA firms... 48 bought so far, with OpenAI's head of applied research in a joint role and its engineers building tax software inside the acquired firms. The results were in last week's edition: one accountant's tax season went from 180 hours to 15.

Microsoft, Amazon, OpenAI and Anthropic have committed over $9 billion combined to building services and forward-deployed engineering businesses... their engineers physically embedded inside client operations, selling outcomes rather than seats.

Anthropic launched Claude for Legal in May... twenty-plus integrations into the legal stack, twelve practice-area plugins, and firms like Freshfields and Quinn Emanuel using it on live matters. Fortune put it well: no longer just the model underneath Harvey and Legora, but a direct participant in legal work. And you already know how the market reacted... that January selloff was this launch's opening act. Thomson Reuters had its worst day on record. RELX had its worst since 1988. Investors were not confused about what it meant.

There is a smaller example one layer down that I find telling, and I mean it as an observation about gravity, not a criticism. ElevenLabs built the best voice models and sold them to everyone... including the wave of companies building AI call-centre agents on top. Then it released its own agents product, one layer closer to the finished work. Nobody did anything wrong there. It is just what happens in this stack: whoever owns the layer below eventually offers the layer above. The labs sit at the very bottom of the stack. The same gravity applies, with trillions attached.

The polite objection is that the labs would never compete with their own customers... law firms and consultancies pay them millions in API fees, and you do not burn that channel.

I do not buy it, for two reasons.

First, the prize dwarfs the channel. Legal spends single-digit billions a year on AI tools, inside a trillion-dollar work market. If owning the work is possible... and accounting is currently proving that it is... then tools revenue is not the business. It is the scouting expense.

Second, restraint only works if everyone restrains, and they will not. If Anthropic builds a $50 billion services business, OpenAI is not going to sit politely on the sidelines protecting its API relationships. Nor the reverse. Every lab knows every other lab is looking at the same pie. The first mover forces everyone's hand.

"Then clients will revolt and switch models." Switch to whom? Every closed lab is running the same play. What the channel conflict really decides is the order they move in... the quiet approaches first, the loud one last.

The stake.

Capital, models, engineers and infrastructure, exchanged for equity in service businesses. This is OpenAI and Thrive today, and it is the cloud playbook grown up... hyperscalers have traded credits for startup equity for a decade. The client notices nothing. The API customers cannot even object... the lab owns a slice of a buyer, not a competitor with a logo.

The FT reported OpenAI put up no cash at all... the stake, which grows as the acquired businesses perform, was paid for entirely in engineers and expertise. And OpenAI's own COO has said it is likely the first of a series.

The roll-up.

The lab or its vehicle buys firms outright and rebuilds the production line underneath them, keeping the brand, the clients and the talent. The acquired firm gets the newest models, embedded engineers, effectively unlimited tokens, and its entire client history as fuel. Accounting is the working prototype. Still quiet... the client sees the same firm, suddenly faster and cheaper.

The firm.

The loud one, and so far a thought experiment... no lab has done it.

Imagine Anthropic Law: elite lawyers hired in-house, the best models on earth at zero marginal cost, a brand every general counsel already knows, and a marketing budget no law firm can match. Cut prices ten or twenty times and watch the market reprice. Notice what this is not... it is not replacing lawyers. The firm would be full of excellent, AI-native lawyers.

They will not replace the humans. They will replace the firms.

Regulation slows this in law specifically... non-lawyer ownership is barred in most of the US, though Arizona, Utah and the whole UK offer side doors. Accounting has fewer locks. Consulting has none. Friction here is not a wall, it is a sequencing instruction... which is why the firm comes last, after the stakes and the roll-ups have done the quiet work.

OpenAI was Harvey's first investor... a $5 million seed in November 2022, before most lawyers had heard of generative AI. Harvey's founders pitched Sam Altman directly, got early access to GPT-4, and built legal's biggest AI company on the lab's models and the lab's money. Today Harvey is worth $11 billion and OpenAI has participated in round after round since.

Sit with that. OpenAI owns a piece of the tool standing between it and the legal market. If Harvey crosses into services and wins, OpenAI owns part of the winner. If the labs go direct and the tools get squeezed, OpenAI loses a $5 million position while gaining a trillion-dollar market. Harvey's investors, meanwhile, priced it at $11 billion... a bet that you can build a lasting business on land your own supplier both owns and covets. The next year of Harvey's choices will say more about this market than any funding round.

All three routes share one engine, and it deserves to be named plainly.

We pay the labs to learn our industries.

Every prompt, every workflow, every edge case a law firm runs through the API is tuition... paid by the customer, banked by the lab. Then the forward-deployed engineers arrive and the learning gets deeper: they sit inside client operations, absorb the know-how, and build infrastructure around the real problems. The client pays for that too.

So when a lab crosses into a vertical, it arrives holding cards nobody else at the table has. It can undercut prices drastically, because its marginal token cost is zero. It can give its own service arm the best models first. It can reserve infrastructure for the businesses it owns. And it holds the pricing lever over every external competitor... the firms bidding against Anthropic Law would be paying Anthropic for the privilege.

Ask the question every enterprise should be asking: would you let a future competitor see your usage, learn your know-how, and build your infrastructure? I cannot think of a better Trojan horse. This one invoices monthly.

We have seen the film once. Amazon built the platform everyone sold on, watched what sold best, then launched its own versions against its own sellers. That was retail. Professional services is seven times the software market, with margins retail never had.

Worth remembering what Amazon's squeeze created, though: Shopify. The merchants who refused to build on their competitor's platform needed an ally, and Shopify grew enormous by arming them. The same seat is open in AI... a model provider that credibly commits to never entering services could become the Shopify of this story, the arms dealer for every firm that will not build on a landlord. Watch who claims that position.

Here I want to say something bigger than markets. I believe capitalism is the spine of our society... it is what lets strong companies get built by resourceful, disciplined people, and it is worth defending with everything we have. Especially now, when AI can move the ground further in a year than it used to move in a generation.

Demis Hassabis, one of the great scientists of our era, wrote last week that with AI "we've essentially found a way to make sand think." He is right, and it is miraculous. But every big technology arrives twice... first the warmth of the fire, then the burnt villages. To my eyes, the burnt-villages version of frontier AI is not only the sci-fi scenarios. It is monopoly: a handful of labs that own the intelligence layer extending that ownership into every market that runs on intelligence... which is, increasingly, all of them. Innovation does not survive that.

One force is positioned to check this, and it had a remarkable week. Moonshot AI's Kimi K3... a Chinese model the company says it will open-source... scored highest of any model on Artificial Analysis's independent run of Harvey's own legal benchmark, ahead of the best closed models. On a second independent run it placed fourth, still ahead of several closed flagships. The caveats are real: the weights are not public yet, some numbers are self-reported, and one evaluation flagged serious hallucination rates. But the direction is hard to miss. Open models are reaching the frontier, and a firm that can self-host one owes the labs nothing... no tokens, no telemetry, no tuition.

Earlier I asked: switch to whom? For the closed labs, that question has no good answer... they are all running the same play. Open weights are the one real exit, the alternative that breaks the dependence entirely. Which is exactly why the fight over them decides everything else in this edition.

The customers, meanwhile, are getting restless. Palantir's Alex Karp... who sells the alternative, so read him accordingly... told CNBC this month that the enterprises he deals with are livid: paying for tokens while worrying the labs will end up with their IP and, in his words, replicate their business. His sharpest jab was a question: if the models are as valuable as claimed, why charge for tokens at all... why not take a share of the value created? Here is my honest answer: because tokens were step one. The labs were not ready to take the value share... they needed the know-how, the data, the infrastructure practice first. Now they have all three. Owning the businesses is how they take the share. That is what this whole edition has been describing.

Which leaves two futures.

In one, open models stay a real competitor... legally deployable, commercially usable, good enough that no firm is hostage to a lab that also competes with it. In the other, the labs' position hardens into monopoly. And if that happens, it will not happen by accident. It will happen through regulation with good intentions on the label.

To be clear, I am not against the Frontier AI Standards Body that Hassabis proposed last week. Testing frontier models for catastrophic risk is sensible, and few people alive are more qualified to design it. But the road to hell is paved with good intentions. A body that decides which models may be deployed in the US market can, if captured, become the mechanism that keeps competition out... nuclear risk, bio risk and privacy make unimpeachable cover stories. Done right, it protects the public. Done wrong, it protects the incumbents. The difference sits in who staffs it, who funds it, and whether open models get a seat or a cell.

If that sounds paranoid, consider what Dean Ball, OpenAI's head of strategic futures and formerly a senior AI policy advisor in the Trump administration, sketched publicly last week: no need to ban Chinese open-weight models, he wrote... just have federal agencies issue soft-law warnings that create enough regulatory fear that enterprises back off on their own. The warnings, he added, "needn't be that well justified." He has since said he was predicting what the administration will do, not proposing it, and the backlash reached inside the White House. Forecast or confession, the playbook is now written down in public, by a man whose employer benefits from it.

If you sell services: build the moats that force a lab to buy you rather than compete with you. Proprietary data they cannot reach without you. Regulatory licences. Long client contracts. Operational knowledge that lives in your people. And stay model-independent... build so you can swap providers in a week, because a firm that can walk away is a firm nobody can squeeze. Being acquired by the winner is a strategy. Being commoditised by them is the default.

If you sell tools: cross into services before the squeeze closes. Sitting between the labs and the work is the worst seat in the theatre, and it gets worse every quarter. Own the outcome, or get priced like a feature.

The sequence runs quiet to loud. Stakes first... already running. Roll-ups second... already running in accounting, and within twelve months pointed at legal and insurance brokerage, as I said last week. The firm comes last: my call is that within twenty-four months, a frontier lab launches or majority-owns a professional services firm, under its own brand or at arm's length.

Two more, dated. Within twelve months, at least one major vertical AI tool announces its own services arm or gets acquired into one. And on policy: within a year, expect coordinated regulatory pressure against Chinese open-weight models in regulated industries... not bans, but vague official warnings designed to chill adoption. The tell will be the vagueness. If the justifications are thin and the beneficiaries are obvious, you will know which way the fork went.

If I am wrong, all of it will be obvious by the dates named. Hold me to them.

The first two editions mapped single markets. This one named the force that flattens all of them... every tool, native, incumbent and consolidator in this category is negotiating, knowingly or not, with the same landlords.

The first four editions are weekly, as promised, and the news is not slowing down.

Subscribe, and if you think the labs will stay in their lane... reply on LinkedIn and tell me why.

See you next Tuesday.