The rumors of OpenAI potentially acquiring Windsurf last week triggered a whole round of new takes on the internet: concerns about moving up into the application layer, fears of widespread consolidation, and our perennial favorite, concerns about the demise of all AI application startups when OpenAI inevitably creates super-intelligence. We’ve talked about many of these topics in the last year — where innovation will come from, when and whether acquisitions will happen, and the ongoing claims that, this time, the incumbents actually will win the arms race.
While we typically try not to spend too much time reacting to the news, this is right up our alley so we couldn’t resist. As we wrote this post, we found ourselves referring back to many of the things we’ve written about over the last year, so we’ll link back to the relevant deep dives as we go.
At a high level, we generally agree that — if this acquisition does in fact materialize — this is going to trigger a wave of activity and anxiety for all of the large players in the space. That’s likely going to lead to a round of early acquisitions, so we thought it would be a good time to revisit our thoughts on consolidation, differentiation, and innovation. Let’s dive in!
At this point, it’s pretty clear that there’s a stark difference between innovating at the model layer and the application layer. We won’t make the case for the umpteenth time about Cursor being a UX innovation. You don’t need one for the other, and it’s likely that the teams that do model innovation (which requires deeply technical research skills) look very different from the teams that do application innovation (which requires you to ship fast and iterate efficiently). Most products will fall somewhere on this spectrum; however, it’s pretty clear that if you take a snapshot of the landscape today, the model innovation side of the world has a few, large concentrated players while the application innovation side has many small, narrowly focused companies. Even companies like Mistral that are relatively well-capitalized haven’t been able to keep up with the pace of innovation from the largest labs.
In response to these rumors, we saw plenty of takes flying around that with LLMs commodifying, model builders are realizing that they have to be in the application layer (or else!). We first wrote a year ago that LLMs were commodifying, and that trend has only accelerated since. (All model releases are starting to sound the same.) That doesn’t meant that foundation models are going to be worthless — as we wrote about a couple weeks ago, the market for applications in aggregate will be larger than foundation models, but foundation models will be larger than any one application. It’s very likely that the OpenAIs and Anthropics of the world — who are clearly trying to build generational companies — will move into application areas as they mature, but that doesn’t mean there isn’t value at the foundation model layer, and these companies are probably already in something of an Innovator’s Dilemma — they want to make sure the most value accretes to the model assets they’ve invested hundreds of milions of dollars into.
In that world, acquisitions are inevitable. We actually think it’s more likely that this potential acquisition causes a stir amongst the non-AI big tech incumbents. One of our predictions for this year was that a large tech company would spend ≥ $1B on a strategic acquisition — this would certainly qualify. And while we see why OpenAI’s going after this kind of application, if we had to guess 4 months ago, we’d have said that a more traditional incumbent (e.g., Microsoft) would have been the natural acquirer for this kind of product. In our own day-to-day, we generally feel that the SaaS incumbents are way further behind cutting-edge AI applications, and regardless of the outcomes of this particular acquisition, we expect to see more acquisitions come from the traditional players.
That said, just because an acquisition happens it doesn’t always work out (AOL-Time Warner, anyone?). Not all acquisitions fail, and not all acquisitions succeed. What gives us pause about a foundation model lab like OpenAI acquiring an application company is that OpenAI and foundation model labs very clearly aren’t application companies. If they were, we wouldn’t all be stuck in model selector hell or be trying to figure out what Claude 3.5 Sonnet (New) meant. The folks at each of these labs are obviously incredibly smart and are capable of figuring things out — but cultural norms run deep, and the culture is clearly focused on AI innovation at the expense of product clarity. If this acquisition weren’t to pan out successfully, we’d probably look back and say that the reason it didn’t work is the same reason that Apple’s software services have historically struggled — the shift in cultural norms is just too large.
At the end of the day, we’re still confident that the rumors of startups’ deaths have been greatly exaggerated. It’s not that foundation model labs don’t know that applications aren’t important, and it’s not that they’re not capable of building good applications if they tried. But what it takes to build good applications is very different from what it takes to build good foundation models, and that’s not going to change in the near future. Good applications are still going to be more than a single model call, and the value is going to be in applying models thoughtfully to specific use cases while integrating deeply into customers’ workflows. No matter how many acquisitions OpenAI tries to do, they’re not going to be able to build every application.
There are likely other implications of this acquisition. It’s possible that there’s a chain reaction of other large players getting a sense of FOMO. Cursor likely isn’t going anywhere (and presumably already turned down OpenAI?), but if OpenAI christens this as the first major application area, everyone’s going to want a horse in the race — maybe even GitHub. This also might trigger antitrust concerns as we start to see more interest in potential consolidation.
In the end, the winners won’t be those who build the most impressive models or make the splashiest acquisitions — they’ll be those who deliver real, tangible value directly to customers. OpenAI’s aware that they can’t just be the system internals or that they’ll become a commodity. (Llama’s strange licensing model already is reminiscent of the “Intel Inside” branding.) What matters is owning the customer experiences and the insights. Ultimately, businesses won’t rearrange themselves around AI — the AI systems will have to meet businesses where they are.