The new geometry
of M&A.

JP Morgan paid Andrew Carnegie $480 million for Carnegie Steel in 1901, and welded it onto 200-plus smaller mills to build US Steel, the first billion-dollar corporation in history.

In 2026 dollars, that math runs north of $17 billion. Morgan bought the entity, and the entity contained the asset. Buying one got you the other.

That logic ran for a century: acquirers paid for entities, entities held the value, and the company was the unit of consolidation. Integration meant absorbing the whole thing, books and boilers and people and lawsuits all stitched together inside the parent.

The legal scaffolding assumed the same shape; corporate law treated the firm as a fictional person with a continuous existence, capable of owning property, employing workers, signing contracts, and being acquired as a unit. Antitrust law inherited that assumption. The Sherman Act of 1890 and the Clayton Act of 1914 both presupposed that competitive harm flows through the buying and selling of corporate entities, and a century of jurisprudence followed the same furrow.

In March 2024, Microsoft inverted it.

The company paid $620 million for a non-exclusive license to Inflection AI’s models, plus $30 million to settle legal claims arising from mass hiring. Mustafa Suleyman and Karen Simonyan led roughly 70 of Inflection’s 90 employees across the street to Microsoft, and Suleyman became CEO of Microsoft AI.

Inflection — the legal entity — kept breathing, and its investors got their $1.3 billion back. But everything that mattered had already left the building.

This is the new geometry: the acquirer no longer needs the company to get the capability; the capability detaches from the corporate form and walks out under its own power.

The gut and release acquisition has hardened into the standard operating procedure of AI dealmaking. In June 2024, Amazon ran the same play on Adept AI; roughly two-thirds of the staff joined Amazon, the founders included — while the company kept a husk and a payment.

In August 2024, Google paid around $2.7 billion for a non-exclusive license to Character.AI’s technology and hired Noam Shazeer and Daniel De Freitas back into the mothership. Shazeer is the same researcher Google originally lost when he co-authored “Attention Is All You Need” in 2017 and then left to build his startup. Google paid him to bring his team home.

In 2025, Google did it again with Windsurf for around $2.4 billion. The founders went to Google, while the remaining employees were stranded at a leaderless company until Cognition rolled up the wreckage.

In June 2025, Meta took a 49% non-voting stake in Scale AI for $14.3 billion and pulled in Alexandr Wang to run a new superintelligence group, with Meta returning voting rights to Wang to keep Scale operationally neutral.

Add Nvidia’s reported $20 billion arrangement with Groq, plus a few smaller ones, and the total clears $40 billion in two years — none of them technically acquisitions as we’ve understood them.

Microsoft paid roughly $9 million per Inflection employee; Google paid closer to $90 million per Character researcher. The headline numbers look absurd against revenue, against models, against any conventional valuation logic; they make sense only if you understand what was actually being bought and sold.

Two pressures inverted the geometry, one competitive and one regulatory, and they fed each other.

AI capabilities concentrate in small groups of researchers in a way that violates most assumptions of corporate M&A. The model weights are increasingly fungible, and the infrastructure can be rented from any of the hyperscalers. What can’t be moved is the cluster of 8 (or 80) people who actually know how to train a frontier model and ship it.

The Hart-Scott-Rodino Act requires merger filings above a threshold that sits in the low nine figures. Acquisitions trigger antitrust review, but licensing agreements and individual hires don’t, at least not directly. The structure that Microsoft pioneered with Inflection routed around the trigger by making sure no single transaction looked like a merger.

Compare the path real acquisitions have to walk. IBM’s $6.4 billion buyout of HashiCorp closed in early 2025 only after a second-request investigation by the FTC and a Phase 1 review by the UK’s Competition and Markets Authority; and HPE’s $14 billion acquisition of Juniper Networks went all the way to a DOJ lawsuit and a controversial settlement. Both deals took the long way around the regulatory perimeter, but the license-and-acquihire takes a shortcut through the wall.

By mid-2025, the DOJ had opened a formal investigation into Google’s Character.AI deal, and the FTC had issued a staff report arguing that pseudo-acquisitions could constitute unfair competition by starving rivals of engineering talent. In March 2026, FTC Chair Andrew Ferguson told Bloomberg the agency intends to investigate acquihires and reverse acquihires for HSR violations.

But the regulators are late to the orgy. The deals that mattered closed in 2024 and 2025, the talent has been re-housed and the market has already been restructured. Whatever the FTC decides about the next license-and-acquihire, the precedent from the last 12 has taken root.

Lina Khan’s FTC made deals harder in a way that pushed acquirers toward exactly these structures. Khan’s strategy of treating mergers as presumptively suspicious, combined with a willingness to litigate even cases the agency expected to lose, created an 18-month penalty box for anyone trying to buy a company outright; license-and-hire structures dodge the penalty box entirely.

Andrew Ferguson’s FTC, sworn in alongside the Trump administration in January 2025, has signaled it’ll look at the structures while relaxing pressure on traditional deals. Gail Slater, who ran the DOJ Antitrust Division alongside him, resigned in February 2026 after reportedly clashing with AG Pam Bondi over the HPE/Juniper settlement. The internal politics are messy, but the direction of travel is clear. Traditional M&A will get waved through faster, and the workarounds will draw closer scrutiny.

But each administrative transition creates an arbitrage window. The dealmakers who closed Inflection, Adept, Character, and Windsurf understood Khan’s enforcement posture and structured around it. The next wave of acquirers will read Ferguson’s signals and structure around those. Whatever shape the regulator settles on, the structurers will already be one move ahead, because the structurers only need to be right about the present deal and the regulators have to be right about every deal.

The shape may bend back partway, but it won’t snap.

A merger absorbs everyone. Common stock converts, options vest, bondholders get paid out, employees join the acquirer or get severance, and the whole population rides the deal together.

A reverse acquihire breaks that solidarity by design.

The principals win — both the founders and the early investors whose preferred shares get bought back near par when the licensing fee circulates through the corporate husk. Named senior researchers win too, with Big Tech executive jobs and eight-figure sign-on packages waiting on the other side of the transaction.

Recent rank-and-file employees lose, especially anyone who joined in the last two years. Their unvested options can’t convert because nothing’s being acquired. The company they joined still technically exists, but it’s been hollowed out. Without the founders, without the top researchers, without the model roadmap, without anyone authorized to make decisions, what’s left runs on autopilot until the cash burns down.

When Google pulled Windsurf’s founders and core team in 2025, the rest of the staff sat at their desks waiting for instructions that no one was empowered to give. Cognition eventually showed up and bought the remains, which was a kinder outcome than most. In other cases, the husk runs out of money with zero fanfare.

The acquirer’s whole reason for using a license-and-hire was to avoid paying for the parts of the company that weren’t the talent. The non-talent parts include, mathematically and morally, the recent hires.

In the old geometry, a company was the container for a capability. You couldn’t separate the team from the corporate entity without enormous friction. Non-competes, IP assignments, equity vesting, and cultural inertia all worked together to keep the substance bound to the form. M&A was the legitimate way to break the binding.

In the new geometry, the binding has weakened. The team can leave intact, the IP can be licensed across the boundary, and the corporate form can be disposed of as a separate question. What remains of “the company” is mostly a registration certificate, some commercial contracts, an empty cap table, and a forwarding address.

Modern corporate theory, going back at least to Ronald Coase’s 1937 paper “The Nature of the Firm,” treated companies as bundles of contracts that exist because internal coordination is cheaper than market coordination. The firm has boundaries because crossing those boundaries costs something. What the license-and-acquihire reveals is that, for a particular kind of high-talent, low-asset business, those boundary costs have collapsed. The team can be lifted across the boundary in an afternoon. The IP can follow on a license. The friction Coase wrote about is no longer enough to hold the firm together against a determined external bid.

Companies still matter. They raise money, sign customer contracts, hire, and litigate. But for a particular slice of frontier-model AI startups, the company has become the wrapper around the asset, and the wrapper is disposable. The asset is the cluster of researchers and the model weights they trained, and those can be moved.

If this generalizes, and the early signs are that it already is, past pure AI labs into developer tools (Windsurf) and data infrastructure (Scale), then the question every founder and every late-stage investor needs to ask is: what part of this company would survive if the founders walked?

The buyer is going to ask, and the answer is the actual valuation.

The license-and-acquihire structure has had two good years and is now under active investigation by both US antitrust agencies. The next iterations are already visible in outline.

Compute is the new collateral. Microsoft’s deeply tangled relationship with OpenAI and Amazon’s $8 billion stake in Anthropic aren’t acquisitions either, but they bind the startups to the hyperscalers in ways that make independent exit nearly impossible. The capability flows back up the stack as model usage, in exchange for compute credits that flow down. No HSR filing required, and no clean line on the org chart that a regulator can point at.

Talent scarcity // arbitrage will keep widening. The price of frontier-model researchers is set by whoever’s most desperate. Meta’s $100 million sign-on offers in 2025 set a floor that OpenAI, Anthropic, xAI, and Google now have to meet. The license-and-acquihire was, in part, a way to package those payments inside a corporate transaction so the cap table absorbed some of the cost. As that path narrows, the payments will move directly to individuals. The talent market will start to resemble professional sports, with signing bonuses, non-compete carve-outs, luxury-tax dynamics, and franchise-tag standoffs that nobody in corporate finance has names for. At least, not yet.