Will OpenAI eat our lunch?

Startups, incentives, and innovation

A common question B2B startups used to get asked in the 2010s was why AWS (or Google or Salesforce) simply wouldn’t build their product. Contrary to what the skeptics might have thought 15 years ago, there have in fact been many successful companies that came out of the cloud and SaaS era, despite the existence of large and motivated incumbents. The 2025 version of that question is why OpenAI or Anthropic or Google (again!) won’t build your AI application — why won’t they eat your lunch?

Source: GPT-4o image generation.

To us, that question seems even more ridiculous than the 2010s version. Despite the fact that OpenAI occasionally puts out press releases saying that they’re building $20k per-month agents, we’re not very concerned that OpenAI’s going to build an AI Support Engineer that suddenly puts RunLLM out of business.

Before we dive into why, it’s worth saying that this isn’t a question of capability. Do the foundation model labs have the technical wherewithal to build an AI Support Engineer? We’re absolutely sure they do, and if OpenAI said tomorrow that the path to AGI is to build an AI Support Engineer… well, we’d check our calendars to see if April 1st came a little late, and then we’d be worried. (But we’d still be convinced it was a joke.) The other minor caveat is that we’re not accounting for some hypothetical future where ASI (artificial superintelligence) solves all of humanity’s problems at once.

With all of that out of the way, let’s dive into why AI applications have plenty of defensibility. The reason why you might think that foundation model providers will go after AI applications is because it’s a large and attractive market. In fact, we’ve made the argument a couple times around here that AI applications in the aggregate will be more valuable than the foundation model companies will. This is partially because foundation models are currently in a race to the bottom, and it’s also because the application space will bridge the gap between what LLMs are capable of and how businesses actually benefit for them. For what it’s worth, we asked Deep Research to estimate the market size of AI applications, and it’s guess was that the market would be $175B by 2030. That seems reasonable — and potentially like an underestimate if popular AI applications truly enable businesses to scale more efficiently.

While the numbers are attractive, the issue comes down to execution. The foundation models labs are all taking massive swings at building better models (and eventually building some kind of ASI). That looks very different than what it takes to build a great AI application — here’s why.

It’s a distraction from their core business. While Google’s done a great job of over the years of building a suite of easily accessible productivity tools, they’ve never fully made it into the depth of B2B SaaS products like CRMs and ticketing systems because most of Google’s SaaS applications are optimized for being consumer-facing. Building B2B products is a different muscle — it requires a focus on integrations, managing workflows, and measuring very specific ROI. The foundation model labs are even further from being able to do these kinds of things.

As we’ve seen with other incumbents in the past (e.g., Microsoft in mobile, whatever is going on with Apple Intelligence) making these types of mindset shifts is incredibly difficult even for the most successful of companies, and when you invest resources into these kinds of shifts, you inevitably distract yourself from the core, successful parts of your business. Even if the foundation model labs were to try to get into AI applications, they likely wouldn’t succeed, and it’s far too early for them to risk taking their eye off of the model arms race.

Great model companies aren’t necessarily great product companies. While we wouldn’t take anything away from the successes these companies have had in building great models, it’s pretty clear that their product experiences leave much to be desired. Even Google, which has historically been a good product company has struggled to build great AI experiences — at a dinner last week, we spent 20 minutes talking about all the ways in which Gemini could be integrated into GSuite but was simply failing at today. OpenAI and Anthropic have clearly focused on the quality of their models while settling for relatively generic experiences. The true innovation we’ve seen in end-user AI experiences has come from applications like Cursor and Granola which are comparatively light on AI innovation but end-used obsessed.

All that is to say that both parts of the equation are important, but success in one kind of thinking doesn’t necessarily imply success in the other. Again, even if the foundation model labs were to try to get into AI applications, they likely wouldn’t succeed.

The data matters! We wrote about this extensively last week, so we’ll lean on that post as the primary source:

from “AI is (still) all about data”

• First, generic LLMs won’t get better at highly specialized tasks. the data that would help them with this kind of improvement simply won’t be available to the large model providers. When we say ‘highly specialized tasks,’ we don’t mean things like programming (which LLMs have obviously gotten incredibly good at) — instead, we’re thinking about things that require general expertise and domain knowledge, like writing sales emails for complex products or doing highly complex technical support.

• Second, AI applications will specialize over time. The availability of the same data that generic model providers are missing will enable specialized applications to get better. In turn, that means better results, deeper insights, and more value for customers. As the first generation of AI application companies get entrenched, it’s going to become harder and harder for skeptics to argue that a generic LLM could do that task — that simply won’t be the case.

Bonus: The B2B SaaS incumbents aren’t threats. Finally, you might be tempted to tell us that even if OpenAI and Anthropic don’t eat our lunch, the likes of Salesforce, Atlassian, and Zendesk will. In many ways, we’re even less worried about these incumbents than the foundation model labs, and the proof is in the pudding. Despite large marketing budgets and lofty Dreamforce keynotes, we’re seeing relatively little actual AI impact (or AI software for that matter) being built by these organizations. That could always change, but we’re already seeing mass adoption in lots of different AI verticals, so it’s unlikely this changes in the near future.

---

Nothing’s ever written in stone, certainly not in the incredibly fast-paced AI market. That said, we’re seeing that (good) AI applications are much more than just a GPT wrapper and already gaining traction, fast. It’s not an insurmountable lead yet, but the value of the data and the depth of the applications that are being built will only continue to grow. We feel pretty confident that OpenAI’s not going to swoop in and eat any good application’s lunch anytime soon, and the longer that continues, the harder it will be for that to happen.