The AI Application Land Grab

We predicted at the beginning of this year that there would be a slew of AI application acquisitions. We’re of course happy to say that we were right, although things have turned out to be even crazier than we expected. (That is becoming something of a trend for everything in AI these days!) We’ve obviously been keeping a close eye on the news, and we originally were going to wait to see how the dust settled — but things have gotten too interesting for us to keep quiet for much longer.

Source: GPT-4o.

Because things are changing so quickly, we don’t yet have a pristine thesis for what the implications of all these acquisitions will be. It’ll probably be a few years until we can really say. (Just think about how long it took for people to realize Instagram → Facebook was a huge deal.) Today, we thought we’d share a grab bag of reactions to these acquisitions. For our purposes, the acqui-hires like Scale AI → Meta and Windsurf (part 1) → Google count as a part of this list (more on why below). Let’s dive in!

Anti-trust concerns are real, and acqui-hires are here to stay

We first saw the zombie acquisition/acqui-hire trend start in 2024 with Microsoft hiring the Inflection AI founders and Google hiring the Character AI founders. Back then, we might have been able to convince ourselves then that this was a blip on the radar, but with the news in the last month about Scale AI and now Windsurf has us convinced that this is a key trend. Whether you believe this is good or bad, the obvious explanation is that concerns about anti-trust are driving this behavior. You might have been able to argue that the change in administrations earlier this year should have prompted a change in behavior, but with the previous administration’s FTC attempting to reverse the Instagram acquisition, no acquisition can really be considered safe — and so, the disincentive to acquire a market leader is strong.

This trend of course makes for some weird headlines — and probably will lead to a Barbarians at the Gate-style book in 20 years based on last weekend — but it has some longer-term implications as well. As founders, the one that stands out to us is that if key employees continue to be poached while everyone else is left behind, there will be increasing concerns for new employees joining growth-stage startups. (This is less of a concern at the early stages.) While it’s impossible to blame anyone for taking a strong offer, it’s also in everyone’s incentive to avoid this becoming the default expectation. Suddenly and abruptly finding yourself at a headless startup is certainly a bad feeling.

The incumbents (obviously) think they’re behind

The motivation for these acquisitions (whether total or partial) is obvious: The incumbents see an opportunity and think that they’re behind. We’d break down the acquisitions into two categories. The first category is the one that we were focused on in our predictions for 2025 — tech giants that see obvious places where they can’t compete on product and feel the need to supplement their talent. In Meta’s case, it’s obvious that Llama 4 wasn’t competing with the state-of-the-art in frontier models, and in many other cases — e.g., Moveworks → ServiceNow — the product and talent advantage the acquiree had was obvious.

What’s more interesting for us to see Google hire the founders of Windsurf without acquiring the company. The implications here are much more interesting. First, Google believes that they wouldn’t be able to recreate what Windsurf has built in-house, and more interestingly, they believe that the expertise of the founders + license rights to the IP will be sufficient to compete against Cursor (and Cognition, Augment, etc.). That wouldn’t have been our bet. Bard certainly didn’t compete with OpenAI a couple years ago — so the fact Jules doesn’t really compete with Claude Code, Cursor, or Windsurf today wouldn’t have presented itself as a major concern. But this is clearly a long-term play about building a foundation that lasts beyond achieving feature parity, and out bet is that the team that’s jumping ship likely has strong incentives to stick around for a longer period.

Data is just as important as AI

While the slew of AI application and infrastructure acquisitions has been the major focus in reporting, there have been just as many interesting data acquisitions: DataStax → IBM, Neon → Databricks, Crunchy Data → Snowflake, and so on. Many of these acquisitions were explicitly framed as being focused on the AI developer experience, and we believe this is more than just an investor relations tactic.

We’ve discussed many times how important finding the right data at the right time is, and as AI applications become more and more dynamic, data management is going to become a critical part of every application (e.g., see Snowflake’s attempt to acquire Redpanda earlier this year). There’s not much else to say other than that there’s probably more to come on this front as data management companies continue to grow.

Who’s not being bought says as much as who is

While there’s plenty of investment, there seem to be comparatively few acquisitions on the infrastructure side of things. There have been a few — Predibase → Rubrik, W&B → CoreWeave, Lamini → AMD (acqui-hire) — but the leading startups in areas like LLM inference have stayed relatively constant.

It’s hard to know from the outside exactly why this is the case, but there are two possible hypotheses. One is that the startups are doing so well that they’re rebuffing acquisition offers from the major players — but given the dollar sums that are being thrown around, this seems relatively unlikely. What seems more plausible is that the incumbents feel confident enough about the infrastructure they’re building that they don’t see a need to go and buy a product to supplement what they already have. We (unfortunately!) don’t have any inside information here, but it’s tough to argue that the major players need help when we see inference costs come down by 20-200% every quarter.

This aligns very strongly with our belief that value will accrue at the ends of the spectrum: to foundation models on one end and application companies on the other. The major players already have foundation models themselves or are partnered with leading labs, which means that the obvious blindspot is in the applications themselves — or in the case of Databricks/Snowflake/IBM, the infrastructure to support applications.

The news this year — especially this past week! — has kept us on our toes, and we’re pretty sure that the second half of the year will have plenty more drama to keep all of us scrolling through X. Every enterprise believes that there’s so much at stake that it makes sense to throw large sums of many at any problem that’s sufficiently important. It seems to almost be a venture-style model — if you take some big swings, you don’t need all of them to hit as long as one or two of them do. If that mindset is right, throwing a few billion dollars at a problem prevents tens or hundreds of billions of dollars of market cap being wiped out when your business is disrupted. It’s an entirely rational bet, and it’s going to make for some interesting outcomes.