The Future of AI Models Shows Its Hand

I round up the most relevant AI-in-finance news, the deals being done, who's rolling out what, and what's actually working on the front lines.

Google is paying $10 million for the internal business data of Spirit Airlines: the emails, spreadsheets and calendars of a company that stopped flying in May. Mercor bid $7.5 million for the same thing.

Elsewhere, Blackstone and Hellman & Friedman have put a 160-person AI engineering team inside their portfolio companies, and Wall Street's invite-only fishing camp spent the weekend arguing about whether any of the spending pays.

Also inside: Anthropic's retention climbdown, Wispr Flow at $2 billion, Nvidia's $6 billion open-weight bet, and Harvey's memory update.

But first, my take on what Google actually bought, and where I think models go next.

In today's Acquisition Intelligence:

From The Trenches

The Future of AI Models Shows Its Hand

Google is paying $10 million for a dead airline's inboxes.

Spirit Airlines shut down in May under high debt and fuel costs and has been selling assets in bankruptcy since. Reuters reported last week that Google is buying the internal business data for $10 million: employee emails, Teams messages, spreadsheets and calendars, plus marketing, productivity and operations data. Mercor, an AI data company that assembles verified industry material for the frontier labs, bid $7.5 million for it. The customer records get stripped out before completion, so nobody here was bidding on passengers.

What Google Is Actually Buying

Not an airline. Years of how one was run, in the form that work actually leaves behind: what got escalated and what got ignored, how crews were rostered, where the maintenance money went, which decisions were argued over and which were waved through.

That's material no amount of general web text contains, because none of it was ever published. Feed enough of it into a model and the model stops being a clever generalist about aviation and starts knowing how the business runs.

Then Google turns round and sells that understanding back to the airlines still flying. The data is the input and a model that knows aviation is the product.

The Next Evolution of Models

Aviation is first because a large operator went bankrupt and its records came up at auction. The logic works anywhere somebody can assemble enough operating history to train against, and there are plenty of routes to that beyond a bankruptcy sale.

This is the next evolution of the models themselves, and it's slightly funny to watch, because the industry oscillates constantly between horizontal winning and vertical winning. I don't think that's really the argument here. What Google is buying is an understanding of how a business actually works day to day, and that understanding feeds two things at once: the model, and the implementation of it inside a company.

If you're wondering whether anyone else reads it this way, look at what else got bought in the same week. Oakley Capital took majority control of Graphwise, a knowledge-graph business whose entire job is making a company's own data legible to a model. Which, yes, is what we spend our days doing. Google is buying the raw material and a sponsor is buying the machinery that turns it into something usable.

Where This Goes Next

Think of the aviation model as a floor. Once Google has something that understands how airlines run, it can go considerably deeper with any individual airline willing to plug its own operations in. The generic industry model gets built once, and the valuable work starts after that, one customer at a time.

In the SaaS era this is where the story ended badly for the buyer. Everyone bought the same product, everyone got the same capability, and nobody moved relative to anybody else. I don't think models play out that way. A model is a starting point you push further with your own material, so what you do after you buy it matters considerably more than which one you bought.

Everyone Trains Their Own

Which is why I think we're at the beginning of large enterprises training their own models. The shape of it is a capable base model taught how one particular business actually operates: its processes, its exceptions, its history of what worked and what got abandoned. That's where I think the next real innovation lands, and it's a bigger shift than most of the current agent conversation.

All of it has one hard prerequisite. You can't train a model on how your firm works unless how your firm works exists somewhere a model can reach. Your deal history, your operating numbers, the things you passed on and why, the reason a customer left in 2023, organised and linked rather than scattered across systems that never spoke to each other.

Most firms are nowhere close. The deal history is in inboxes, the operating numbers are in spreadsheets on somebody's desktop, and the institutional memory leaves when a partner retires. Getting a firm past that is most of what we do at DealSage.

Google has just put a price on a decade of that for a company that no longer exists.

What The Builders Are Saying

Two posts worth your time this week. Follow both accounts.

@spakhm (Slava Akhmechet, founder of RethinkDB, formerly Stripe)

The post: on the ground, he writes, there's no such thing as AI adoption in the enterprise. Transcription nobody reads, autocomplete nobody uses, document search worse than Claude, with engagement metrics he calls atrocious. Real adoption sits in vertical products like coding, where it genuinely works. The rest is executive-pressure experimentation that produces slop and gets abandoned. All of which makes him extremely bullish, because compute demand is already enormous while barely anything works.

Why this matters: the bull case doesn't need the rollouts to succeed. Demand looks like this while the workshops and the transcription tools achieve nothing.

My take: I agree with him. Winning with this takes hard, long, gruelling work, and there are no quick wins on offer. The firms getting anywhere are the ones that accepted that a year ago and started on the boring part.

@willchen500 (Will Chen)

The post: responding to an FT opinion piece on legal AI, Chen argues a thin wrapper is perfectly survivable in enterprise software, because the goal is not the best product but getting through the law firm procurement process "by greasing up the right people". Get in with aggressive token subsidies, hike prices later. What's defensible is the capital funding sales and distribution, not the technology.

Why this matters: a vertical AI company can look like a software winner while the advantage really sits in distribution, with the early economics running on investor money.

My take: I agree. Most of the early AI winners in enterprise have no differentiation beyond clearing security review before anyone else did. Go and ask someone at JP Morgan or Wells whether they're actually using Rogo and you'll see the point.

News Digest

Wall Street's Summer Camp Got Spooked

Camp Kotok is an invite-only fishing weekend in Grand Lake Stream, Maine, that veteran money manager David Kotok has run for financiers, economists and policymakers for around 25 years, under Chatham House rule. The WSJ went this year and found the AI trade had followed everyone into the woods. Half an hour into a talk on chip design, the speaker was asked whether the trillions going into AI will ever earn a return. He said he can't answer that, and neither can Wall Street.

The details:

Why it matters: this is the room that prices the AI trade, and it has gone from confident to openly unsure in twelve months while continuing to buy.

My take: these are people who can price a capex cycle in their sleep. Every one of them is treating AI as a position to size. Almost nobody in that room seems to be asking what it does to their own business. I get the same thing talking to old colleagues in banking, who have a firm view on Nvidia and no view at all on what their analysts will be doing in two years. The returns here go to firms that change how they work, and timing the trade is a different game entirely.

Read more

Private Equity Is Putting AI Engineers Inside Its Portfolio Companies

Blackstone and Hellman & Friedman have built a roughly 160-person team of AI engineers with Anthropic and started deploying them into the businesses they own, per the WSJ. It sits inside a $1.5 billion joint venture called Ode, backed also by Apollo, General Atlantic and Goldman Sachs, with the three lead parties each committing around $300 million.

The details:

Why it matters: the largest sponsors in the world have concluded that getting AI into a business means putting engineers on site to map how the work happens, rather than buying software and hoping.

My take: they're leading with new revenue lines and saying explicitly that cost savings aren't the point, which is the opposite of what most people expect from a sponsor. It matches what we see: the operational upside is bigger and easier to underwrite than the headcount story. This is exactly what we do for everyone who isn't owned by Blackstone or Hellman & Friedman. If the biggest sponsors in the world are staffing it internally, the question for a mid-market firm is why the work is worth doing for a $5 billion portfolio company and not for yours.

Read more

Other Interesting Things I've Read or Seen This Week

Anthropic plans to change its enterprise data retention policy (Reuters, August 20) - Enterprises still hold 30 days of traffic, now with the option to keep it on their own cloud. (Turns out somebody read the terms.)

Nvidia is spending $6 billion to build a US alternative to Chinese AI (WSJ) - Nvidia licenses Poolside's technology, invests $1 billion at a $12 billion pre-money and takes most of its engineers into the Nemotron open-weight project. (The most expensive acquihire in history, arranged by the landlord.)

How Harvey builds AI around domain expertise (WSJ CIO Journal) - Token consumption on the legal platform went from 1 trillion a month in January to 14.5 trillion in June. (Either product-market fit or a very expensive habit.)

Wispr Flow valued at $2 billion on demand for AI voice-to-text (Reuters, August 17) - A $280 million Series B nearly triples the valuation in nine months. (Dictation, the killer app nobody had on the card.)

Accounting AI startup Rillet reaches unicorn status (Fortune, August 18) - $100 million at a $1 billion valuation for an ERP that puts agents inside the general ledger. (The books close themselves, and the auditor still signs.)

Private equity has a software debt problem (Axios Pro, August 19) - Software buyout debt is maturing exactly as AI erodes the SaaS economics it was underwritten on. (The multiple was the collateral.)

90 per cent of executives say AI hasn't boosted productivity, and some are still cutting jobs (Fortune, August 22) - Most senior managers report no measurable gain, while firms keep citing AI as a reason for cuts. (Both can't be true, and the layoffs are the ones happening.)

Nvidia and Wall Street firms strike an AI financing deal targeting $500 billion (WSJ, August 18) - Apollo, Blackstone, BlackRock, Brookfield, Goldman and KKR are building platforms to help customers afford AI hardware. (Vendor financing, wearing a very good suit.)

Stripe clinches a $7 billion-plus deal for OpenRouter (Bloomberg, August 16) - The payments company buys the model-routing marketplace. (Whoever sits between buyer and seller does well again.)

Acquisition Intelligence is a weekly newsletter on AI in M&A for finance professionals, private equity investors, investment bankers, corp dev teams, and deal-makers.

For questions, feedback, or to share what you're seeing in the market, reply to this email.

P.S. I'm Harry, co-founder of DealSage. The work in this issue is the work we do: we get a firm's deal history and a portfolio company's operating systems into one structured place, then build against what's actually there. Reply here if you'd like to see what that looks like on one of yours, or have a look at dealsage.ai.