AI Roll-Ups Are the New PE Playbook

If you're a founder thinking about your exit in 2026 and you're not paying attention to what's happening in AI M&A, you're missing the most important shift in the dealmaking landscape since SaaS itself.

The headline number is the £42 billion or so spent on AI-related acquisitions globally in the last 18 months. That's not the interesting part. The interesting part is the structure. Almost none of it has flowed through traditional M&A. Most of it has flowed through a hybrid structure that didn't exist three years ago: the licence-and-acquihire.

Microsoft-Inflection, $650 million. Amazon-Adept. Amazon-Covariant. Google-Character.AI, around $2.7 billion. Google-Windsurf, July 2025. Meta-Scale AI, June 2025, around $14 billion for a 49% stake plus the founder. Each of these deals re-engineered M&A to do something the buyer couldn't do through a traditional acquisition: get the team, get a non-exclusive licence to the technology, and leave the company structure intact for everyone else.

Below that headline pattern, a second pattern is happening more quietly: traditional PE firms are running AI roll-ups in vertical applications. Healthcare AI. Legal AI. Compliance AI. Manufacturing AI. Same playbook, different sector.

Here's what's actually being acquired, what's being paid, and what UK founders should know before they accept the next term sheet or sign the next LOI.

The licence-and-acquihire: a new deal structure

Big Tech can't buy AI startups outright. The reason is regulatory. After Microsoft-Activision and the global pile-on against Big Tech consolidation, no major AI lab can be acquired by Microsoft, Google, Amazon or Meta without inviting a multi-jurisdictional competition review that would last 18 months and probably block the deal.

So the buyers got creative. The licence-and-acquihire structure works like this:

• The buyer pays a large sum (between $500m and $2.7bn in the deals we've seen) for a non-exclusive licence to the startup's AI models.
• The buyer hires the founder and most of the technical team into a new internal AI division.
• The original company continues to exist as a shell, with its existing investors paid out from the licensing fee, and (in some cases) a residual product business.

Microsoft-Inflection (March 2024): around $620m for the licence, around $30m for the staffing arrangement. Most of Inflection's 70 employees, including co-founders Mustafa Suleyman and Karén Simonyan, joined Microsoft to form Microsoft AI. Inflection itself remained as a corporate entity, pivoted to enterprise AI. Investors got paid out at par or above. Microsoft got the team and the IP without owning the company. The competition regulators got nothing to review.

Google-Character.AI (August 2024): broadly the same template, around $2.7bn. Co-founders Noam Shazeer and Daniel De Freitas re-joined Google. The remaining team and product continued. Investors were paid out.

Amazon-Adept (June 2024) and Amazon-Covariant (August 2024) followed the same pattern.

For the founder of a frontier AI lab in 2024-25, this became the dominant exit. The traditional M&A path was effectively closed. The licence-and-acquihire path was the only one Big Tech would underwrite.

Who wins and who loses in a licence-and-acquihire

The structure isn't neutral. Some parties walk away with what they expected. Others don't.

Founders typically win. They get a large personal cheque, often from a sign-on package at the acquirer plus their share of the licensing fee. They get to continue working on the same problem with substantially more compute. The strategic ambition stays alive.

Early investors typically win. The licensing fee usually gets them out at par or better, which on a venture portfolio basis is a respectable outcome.

Late-stage investors typically lose. The licensing fee is sized to cover the cap table at a specific valuation. If they invested at the most recent peak, they're lucky to get their money back, often less. Any preference stack favours seed rounds in this structure.

The company itself loses. The shell continues but is denuded of its founder, its top talent, and its core IP licence has been sold non-exclusively (which means the buyer's competitors will, in time, license the same technology). The remaining business has to find a new strategy in a market that's just been validated by Big Tech's buy-in. Most don't survive.

For founders, this is usually fine. For the people working at the company who weren't part of the acquihire group, it's often a slow-motion redundancy.

Below the headlines: the AI vertical roll-up

While Big Tech has been acquihiring frontier labs, traditional private equity has been quietly running the same playbook one tier down, in vertical AI applications.

The thesis is straightforward. AI infrastructure is commoditising fast. The actual margin in applied AI is in vertical applications: an AI workflow tool for radiologists, an AI contract review platform for in-house legal teams, an AI claims assessor for insurance carriers. Each of these is a small-to-mid-cap business with $5-30m of ARR, decent retention, and a defensible moat in domain expertise rather than model superiority.

PE firms are buying these businesses for two reasons. First, they have the cashflow profile of vertical SaaS but with an AI premium attached: revenue multiples of 7-12x for the best assets versus 4-6x for the equivalent SaaS without AI features. Second, they can be rolled up. Buy three or four vertical AI players in healthcare. Combine them. Sell the platform to a strategic in five years for a multiple of the combined entity.

The buyers running this play in 2025-26 include Hg, Vista, Thoma Bravo, KKR's tech franchise, EQT and a long list of smaller PE houses with sector-focused funds. The deal sizes are usually $50m–$500m of enterprise value, which makes them invisible compared to the Big Tech acquihires but, in aggregate, far more capital is flowing this way than into the headlines.

What commands a premium right now

Not all AI businesses are getting bid up. Three categories are commanding genuine premiums:

1. Frontier AI talent. The acquihire structure exists because the talent is the asset. If you're a founder of a small frontier lab with ten serious researchers and a credible model, the price per head in the public deals has run between $5m and $40m. That's the highest price-per-head in the history of tech M&A.

2. Vertical AI with proprietary data. A medical imaging AI with FDA approval and a labelled dataset built over five years is much more valuable than a general-purpose imaging model with the same accuracy. Data moats translate directly into multiples. Proprietary, hard-to-replicate, regulated data is the strongest moat in applied AI right now.

3. AI-native workflow tools with paying enterprise customers. A code review tool, a sales agent, a legal drafting platform, all with $10m+ ARR and growing 100%+ year on year, can credibly command 15-25x ARR. The market hasn't seen multiples like this since the 2021 SaaS peak. They're sustainable for businesses that can show real productivity uplift to customers.

What sells at fire-sale

The other side of the bifurcation is just as important. Three categories of AI business are selling at distress prices in 2025-26:

1. Foundation model wrappers. Businesses that built a thin layer on top of GPT-4 or Claude, raised at high valuations in 2023-24, and have since seen OpenAI and Anthropic build the same feature directly into the API. The cap tables are ahead of the value. Most of these are quietly being shut down or sold for the team.

2. Generic AI consumer apps. Image generators, AI assistants, chatbot products without a specific vertical or workflow. The market has consolidated around the platforms (ChatGPT, Claude, Gemini, Midjourney). Standalone apps are subscale and hard to sell.

3. AI businesses with no enterprise traction. Anything that has raised at high valuations on the back of consumer growth, demo videos, or pilot deals that didn't convert to paid contracts. By the second half of 2025, the market started pricing pilots as zero. Buyers want signed annual contracts, not LOIs.

What UK founders should learn

Five takeaways for any UK founder thinking about an AI strategy and an eventual exit.

1. The exit market for foundation-model competitors is closed. If you're trying to build a frontier lab and your exit thesis is “Microsoft will buy us,” understand that the only structure available is licence-and-acquihire. The cheque can be large, but you don't own a company at the end of it.

2. Vertical wins over horizontal. The cleanest exits in 2025-26 are vertical AI businesses with domain expertise, real customers, and recurring revenue. PE buyers are active. Strategic buyers in your sector are active. You don't need OpenAI to bid on you for the deal to be life-changing.

3. Data moats matter more than model moats. A 12-month head start on a labelled dataset specific to your industry is a stronger moat than a 12-month head start on a model. Models commoditise in months. Data moats compound over years.

4. Your enterprise customer count is your valuation. Two paying enterprise customers are worth more, in 2026, than 100,000 free consumer users. The buyer pool that pays premium prices for AI businesses is buying recurring revenue, not engagement.

5. Don't over-raise. The most expensive mistake in AI right now is raising at peak valuations and watching the market re-price your equity downward over 18 months. Take less, dilute less, get to profitability earlier. Optionality on exit is worth more than headline valuation in fundraising.

Where this goes next

Two things are likely in the next 12-24 months. First, regulators will start to look at licence-and-acquihire deals as functional acquisitions and probably try to apply merger control to them. The CMA, the FTC and the European Commission have all signalled this. The era of Big Tech using L&A as a regulatory workaround may be ending.

Second, the vertical AI roll-up trend will accelerate. PE has more dry powder than at any point in the last decade and is actively looking for AI exposure that doesn't require building from scratch. If you run a vertical AI business with $5m+ ARR, the inbound from PE in 2026 will be heavier than at any time in the past three years.

The opportunity for founders is to build the right kind of business for the right kind of buyer. That means recurring revenue, real customers, defensible data, narrow vertical focus, and a cap table that hasn't mortgaged your exit upside. The buyers are coming. The question is whether your business is shaped for the deal that will actually be on offer.

– Adam