The bull case for AI startups
Why enterprises aren't immune from AI disruption
Over the last year, there’s been a lot of hand-wringing about how AI doesn’t present the same opportunities for startup innovation that previous waves of technology innovation (e.g., mobile apps, cloud infrastructure) did. Correspondingly, there’s been a lot of complaining about how AI startups are overhyped and all doomed to fail. Any hype train will of course have its hangers-on, but generally, we think this is wrong — there’s a ton of room for innovation and opportunities for success amongst startups.
At first glance, this criticism might seem reasonable. Unlike Microsoft’s efforts to build a mobile OS or IBM and Oracle’s decade-long quest to catch up to the major cloud providers, every tech company has been laser-focused on AI. There’s been an incredible amount of time and money invested, so with all those resources, how could there still be opportunity for startups?
The simple answer is that everything that was true of the difference between enterprises and startups 2 (or 5 or 10) years ago is just as true today. Of course, that doesn’t mean that every startup will succeed or even that every enterprise will be disrupted. Even in the examples we use below, incumbents have many advantages and might well succeed.
It’s also worth noting that we aren’t talking specifically about foundation model companies but about startups using AI in any capacity. In all likelihood, foundation models — which require massive capital outlays — are the areas in which the incumbents are more likely to succeed. In doing so, they also will enable the the success of startups that build on those models.
The rest of this blog post examines both why any startup can innovate and succeed and also where incumbents have clear advantages. Where applicable, we also point out the positive steps incumbents are taking to defend themselves — but on balance, we believe there’s clear opportunities for startups.
The innovator’s dilemma still exists. Let’s start with the simplest example: search. Startups like Perplexity and You.com have come flying into the search space over the last year with AI-powered summaries in response to search queries. To be honest, we personally haven’t found these tools super useful — search and chat still have separate places in our workflows. But the rate of growth for these tools has forced Google to respond, and many Google searches now have AI-generated answers. It turns out they are pretty bad! But Google will probably figure it out over time, as they did with Gemini.
This is a bold move from Google: If AI answers start diverting ad clicks, it could dramatically eat into their search ads business, which is one of the foundations of their business. The fact that Google is willing to make this move shows they are taking the threat more seriously than an enterprise might have in the past. The seriousness is a good sign but also has the feel of an enterprise playing defense against a startup rather than innovating on its own terms.
Meanwhile, companies like the Browser Company (which makes Arc) are taking a totally different approach, bypassing search altogether and summarizing from multiple sources to generate custom content. (This is what we’d personally really like to see succeed!) Chrome hasn’t headed in this direction yet, so Google’s response is still unclear. Google obviously could build this, but this would just further divert ad clicks — though perhaps ads could be embedded in these pages? Regardless of how they proceed, there are hard decisions with incredibly uncertain outcomes to make.
Enterprises won’t do weird things that might fail. This is more an observation about consumer technology, but startups have the leeway to do unexpected or even wacky things. Some of those wacky things — like Rabbit or Humane — might not always work out at first. But the reputation risk for a startup is low, so those products will naturally get built there first. Meanwhile, Apple would never ship something so half-baked. A consumer, AI-first device will go through years of testing and iteration before it ever sees the light of day.
While Rabbit and Humane might be negative examples, there are positive examples too: Character AI is the most obvious one that larger tech companies might struggle to recreate because of reputation concerns.
In the B2B side, the analog is breaking the boundaries that traditional job functions have. We wrote about this just last week, so we won’t repeat the full argument. Briefly, LLMs will likely be good at the basics of many jobs and bad at the expertise of most — that means good enterprise products will likely not limit themselves to one job title.
Building innovative products at scale is hard. One advantage that every startup has is that it can (should?) identify a niche and sell narrowly before it starts to expand into other markets. Large enterprises often can’t design for smaller markets because their existing customers are spread across many markets.
Microsoft’s suite of Copilots are a great example. They are simultaneously trying to compete against tools that do content validation (e.g., Grammarly), against tools that auto-generate content (e.g., Jasper, Gamma), against general-purpose enterprise search (e.g., Glean), and against tools that offer vertically-customized knowledge assistants. While researching for this paragraph, I discovered that there’s also a live call transcription service, so they’re also competing against the millions of AI-generated call summary tools. The result has been a product that’s received mixed reviews at best and seems to do everything in a sort-of-okay fashion.
Microsoft’s inherent distribution advantage with Office 365 means Copilot certainly will get traction, but each one of the areas listed above is likely a reasonably large market on its own. For customers who care about customization and quality, the limited functionality and negative press will create ample opportunities for startups in each of the markets. In other words, focused startups can still very much out-execute incumbents trying to do it all.
Lock-in and lock-out are both real. As we’ve discussed extensively, AI is all about data. That presents a distinct advantage and disadvantage for incumbents, but we think the disadvantages are bigger. The advantage is that if you’re Google, Microsoft, etc., you already have your customers’ data. The established trust means that your ability to get something useful in customers hands — especially using internal APIs — is significantly accelerated.
The disadvantage is that you have a limited view of the world. A startup doesn’t care where the data is coming from, and many startups end up building large suites of data connectors to their customers’ systems. Microsoft systems aren’t going to pull data from Google Docs or Atlassian or GitLab, but a startup absolutely will.
That type of competitor lockout is where startups have a huge advantage. A customer may well be worried about a startup’s security posture, but those are solvable problems, especially as AI startups mature over the next couple years. On the other hand, getting an incomplete or incorrect answer because the data simply wasn’t (and never will be) available in the tool you’re using is a much more difficult roadblock to navigate. Customers are still in AI checkbox mode today, but that’ll change as buying processes mature. As that happens, products that are focused on quality will need to have all the data they can get their hands on, and startups have a better path forward than enterprises.
Again, these things are all true about non-AI startups, and these things have all been true for decades. The point is that the fact that enterprises are aware of the threat from AI startups has made them more responsive — but this awareness doesn’t erase the advantages that startups have.
It’s hard (read: impossible) to say where startups will win or where incumbents will outlast them. (If we could, we could be making a lot more money than we do now playing the market!) But we’re certain that AI startups will succeed. Some of that success will come from novel uses of the technology that we haven’t yet foreseen, but you can be sure plenty of startups will succeed by out-executing incumbents.