What Google Actually Bought From A24

Last month Google invested roughly $75 million in A24 as part of a research partnership with DeepMind. It’s the first time Google has taken equity in a film studio, so the coverage was extensive, and most of it orbited one question: did A24, the studio built on taste, just sell out to the machines? I think that misses the more interesting part of the deal, which is what Google agreed not to buy.

The deal explicitly does not give Google access to A24’s content library or its data. No training on the catalog, no scraping the archive. What DeepMind gets is a multi-year, non-exclusive arrangement where its researchers sit inside A24’s production process and build tools alongside the people doing the work. Google paid $75 million for a seat in the room, and A24 kept everything the room has ever produced.

That looks like a strange trade until you notice how many other buyers are lining up for the same asset right now, at very different prices.

The Cost Moved, So the Buyers Followed

I argued in “SaaS Is Dead, Long Live SaaS” that AI collapsed the cost of writing software while the cost of owning it, the maintenance and the integration, stayed roughly where it always was. That piece left a question open. If writing software is now cheap and integrating it is still expensive, the money in the industry has to pool at the integration layer, and somebody is going to capture it. Judging by this year’s deal flow, the somebody is whoever can actually see the workflow they’re integrating into.

The problem integration keeps running into is that every business process exists in two versions. There’s the documented version, the one in the process manual and the org chart. Then there’s the version that runs, which lives in one specific person’s spreadsheet and in a judgment call the senior underwriter makes on Thursdays for reasons she’d struggle to explain but is usually right about. Nobody involved is being careless. For a hundred years the only people who needed to understand how a firm actually worked were its employees, and they learned it by standing next to each other, so the knowledge never had to survive in a form an outsider could read.

AI changes who needs to read it. A model can automate the documented version of a process and fail completely on the real one, and from the outside there’s no way to tell which version you’re looking at. It’s like buying a house from the listing photos… everything looks great until you want to knock a wall down and need to know which ones hold the roof up.

Model quality is abundant now. Frontier releases leapfrog each other every quarter, and open-weight models keep pulling the price of a token toward zero. Access to how work actually happens inside organizations hasn’t gotten cheaper, because there’s no market where you can buy it, and that scarcity connects three stories that otherwise look unrelated.

Three Buyers, Three Prices

The first buyer is the frontier labs, and they’re renting. The forward deployed engineer was Palantir’s odd, expensive habit for a decade, and now it’s the standard playbook… OpenAI went further this spring and launched an entire deployment company, then bought a consulting firm to staff it with 150 forward deployed engineers on day one. Consider what that admits. A company whose entire pitch is automation is paying humans to sit on-site, at a cost that scales linearly with headcount. The model can’t see the workflow from the API side, so someone has to go find out which walls hold the roof up. They do it anyway, which tells you what the access is worth.

Alex Karp said as much on CNBC two weeks ago, in his way. He went after the labs’ pricing model, arguing that enterprise CEOs are livid because they’re paying for “tokens that create no value.” His sharper argument was about what the pricing reveals. A vendor who could reliably make you a billion dollars would ask for a percentage of the billion, and the labs meter usage instead. Karp has an obvious interest in saying this. Palantir’s whole business is embedded engineers priced on results, and the interview came two days after Palantir announced a sovereign AI deal with NVIDIA, so he was selling the alternative while he criticized the incumbents. The criticism still lands. Token pricing is what you charge when you can’t see far enough into a customer’s business to underwrite a result. There’s a duller explanation available too. Usage metering is how every commodity input gets sold at scale… AWS charges for compute by the hour, and nobody reads that as an admission that Amazon can’t see your business. Charging for outcomes means attributing outcomes, customer by customer, and attribution doesn’t scale the way a meter does. Both readings can be true at once, and the rest of the deal flow suggests the market is at least partly pricing Karp’s. His other complaint cuts the same direction. Every API call ships a little of the customer’s proprietary process knowledge to the provider, what he calls the weights and alpha of the business, which is exactly why enterprises ration that access, and why it commands a premium whenever it’s actually for sale.

The second buyer is Google, and it bought a window. Seventy-five million dollars for a multi-year arrangement that isn’t even exclusive. If training data were the prize, the content library is the obvious ask, and it’s the specific item Google agreed not to touch. What remains once you remove the films and the data is the process itself… how a studio famous for refusing to explain its own success decides what to make and how to make it. A24 has spent a decade declining to tell Hollywood how it works, and it just sold Google a view of the how while keeping the what. And because the deal isn’t exclusive, A24 can sell the same view to the next buyer, since letting someone watch how you work doesn’t use anything up. The incentives are interesting. Veo can train on everything the public internet has to offer, and none of it explains how a working production would want to use the tool, because that knowledge only exists inside working productions.

The third buyer is private equity, joined lately by venture firms running a private equity playbook, and they’re buying the whole building. Several of the largest venture firms have raised dedicated vehicles to buy accounting firms and insurance brokerages outright and rebuild them around AI from the inside. The deals are already large. Long Lake agreed in May to take American Express Global Business Travel private for $6.3 billion, and General Catalyst and Trian closed their $7.4 billion take-private of asset manager Janus Henderson at the end of June. The logic is simple. Sell an AI tool to an accounting firm and the firm keeps the margin the tool creates. Own the accounting firm and you keep it, which means you’re buying at a services multiple and underwriting software margins, and the spread between those two numbers is the whole trade. Owning also ends the sales problem, since nobody has to make a career-risk call on your software when you own the company. What it doesn’t fix is legibility. The underwriter’s Thursday judgment call is exactly as undocumented after the acquisition as before, so the roll-up math still depends on an extraction step none of these deals is old enough to have demonstrated.

So the labs are renting access through payroll, and Google bought a window with equity. The buyout funds went furthest and are buying the companies outright. Each step up costs more and buys deeper, more permanent access to the same input. Buyers this different don’t converge on one asset unless it’s been mispriced, and the spread between hiring an engineer and writing a $6 billion check means the market hasn’t agreed on the right price yet.

Who’s Left Holding the Tools

Now run the same lens over the application-layer AI startup, the “AI for X” company. Strip the pitch down and it usually reads: we understand how work happens in industry X, and we’ll automate it. The workflow knowledge is the product, which is an awkward position when the workflow knowledge is what every better-capitalized player in this story has decided to stop buying through intermediaries. The startup acquires it the hard way, deal by deal, through nine-month sales cycles and pilots that may never convert, and what it ends up holding is a rented view of workflows it doesn’t own. The PE firm that buys forty insurance brokerages ends up with the same knowledge, and it also gets the revenue and the margin, and it never has to ask anyone’s permission to deploy. The startup that spent two years learning how mid-market brokerages process claims built something genuinely valuable… for a buyer class that would rather own the brokerage.

I don’t think this kills the application layer, but it sharpens the diligence question, which was always the useful kind of question anyway: does this company own its workflow, or rent a view of it? Ownership at the application layer does exist. It can look like proprietary data that gets better with every customer, or a position so embedded you’ve become the system of record, the way nobody ever rips out their payroll provider. Sometimes it’s a domain with enough regulatory friction that nobody upstream wants to own the building. “We integrate with their workflow,” on its own, is renting, and renting is a fine way to start and a bad thing to still be doing when the landlords show up with acquisition capital.

If a model ever reconstructs a firm’s real workflow from its exhaust alone, the documents, tickets, screen recordings, and meeting transcripts, without a human embedded in the room, then the premium on access collapses and the roll-up math gets much worse. The early evidence would look like a published eval where a model rebuilds a company’s actual process from tickets and screen recordings, exceptions included, at a fidelity an embedded engineer would sign off on. I haven’t seen one yet, but some people are betting on exactly that outcome.

The labs themselves may be among them, and the instrument they’ve chosen says so. Paying for access through salaries is the cheapest and most reversible commitment available, the one you pick when you expect a scarcity to be temporary. Google went longer and locked up a multi-year deal, and the buyout funds went longest of all, holding companies they can’t easily unwind. Lined up that way, the three prices read like a term structure of bets on how long workflows stay illegible, and the buyers with the best information about model progress are sitting at the short end. A24 kept everything it has ever made and sold a look at how it makes things, and one of the most sophisticated AI organizations on earth decided the look was the part worth $75 million. When the people building the models tell you what they can’t see, believe them. When they rent instead of buy, believe that too, because they’re also telling you how long they expect the blindness to last.