Why I Disagree With Sequoia's Thesis on AI

The App Layer will not be the best place to make money

Happy Sunday and welcome to Investing in AI. Today’s post is very contrarian to the thinking of mainstream VC. I’ve heard almost every partner I know at every VC firm I know tell me “the infra space for AI is overplayed, the best opportunities are at the app layer.” In fact, during Sequoia’s recent AI day, they said as much during their overview presentation. Today I want to give the opposite perspective - that the app layer is a difficult place to make money and that infrastructure is the best place to invest right now. I want to explain why as a builder and an angel investor I’m focused primarily on infra.

Setting aside the idea that whatever is hottest in VC in a given year tends to turn out badly as a category, what are the reasons this belief in the app layer is so prevalent, and why don’t I believe in it?

First let me qualify a few things. One is that, there is almost always opportunity at every layer in the stack, the question is about how much opportunity and what is a better bet. Second, “app” could start to mean many different things in an AI world so, the idea I am criticizing here is the idea that most existing end user software/SaaS applications will be replaced by AI versions. Second, I don’t consider deeply verticalized AI companies where the app is combined with hardware, sensors, and/or human-in-the-loop expertise to be traditional “apps,” but I do think those companies will be successful. It’s just that the app is a small part of the overall value prop.

And at the end I will give you a different point of view that may contradict my core thesis here because it re-imagines the app layer as more of an infrastructure layer.

The criticism of the infrastructure layer is that it has been played out and the winners have been selected: NVIDIA, the hyperscalers, LangChain, MCP, etc. I get why people think this. Almost every startup you see is building with the same stack. But you know what? In the early to middle years of the web - the LAMP stack dominated.

But just a few years after this was the dominant stack and everyone assumed it was set, the world shifted. NoSQL became part of the stack. Postgres overtook MySQL, and PHP faded as the dominant programming language for web apps. Even in the AI space, Tensorflow grew rapidly after its introduction and seemed to be overtaking PyTorch, only for Pytorch to make a major comeback. In my view, none of this is remotely close to being over and decided.

The reason I’m so interested in AI infrastructure is because the underlying technology of AI is shifting rapidly. Use cases are changing. And I still expect a few technological discontinuities to come into play - like the rise of a non-transformer based model class. Some of these previous shifts in AI infrastructure leadership happened because the tools were optimized for technology like RNNs or CNNs and then LLMs became the driver of application development and new entrants were better positioned for that world. I believe that is happening now, and will continue to happen for the next decade as AI development settles into a market equilibrium. There will be massive opportunities for AI infrastructure startups.

THE APP LAYER

Now let’s move on to what I dislike about the app layer. Yes, I believe everything will become agentic and AI apps as a category will be massive. But my hesitancy to invest there rests on two beliefs I have:

1. The app layer will be extremely fragmented, making it more competitive, pushing down margins, and making it harder to make money.

2. The winners will largely be existing companies, not startups. Obviously there will be a few exceptions but, I don’t expect AI companies to beat SaaS companies the way SaaS companies beat software incumbents.

To expand on point number 1, you have tools like Loveable that make it easier for non-programmers to build apps. And if you know how to code, tools like Cursor can speed it up dramatically. When you lower the barrier to software creation, you get more software created. That means companies will shift their build vs buy decisions and build more software, so they can get exactly what they want, and buy less software. It also means that entrepreneurs can re-segment a market easier. Build a targeted app, accept a smaller TAM, and still build a nice tech lifestyle business. The ability to build software will matter less as a core competency, and the ability to understand what to build, or understand a key market segment will matter more.

Some categories will be less fragmented. Most people don’t even know the requirements to build say, a payroll system or software to track international sales taxes. But for common categories that everyone uses, I expect those categories to be fragmented, competitive, and have worse margins. Common application categories will become a race to the bottom in pricing.

The second point is - incumbents have too many advantages. AI is often an accelerant more than a disruptor, and large companies benefit from existing customer relationships, strong brands, and lots of data. Many of the app layer winners will be the existing firms converting to AI or adding agentic capabilities, over the startups trying to unseat them. Plus tech companies today understand the innovators dilemma and defend against it better than companies in previous eras.

One big market impact of AI is that it will take human workflows and push them down to compute workloads. So the compute layer and the infrastructure layer to manage it are going to have larger TAMs than most people estimate today, while the application layer is going to grow, but less so, and have more fragmentation.

One caveat here - I think there is a chance to build apps as infra, and that could be big. What do I mean? Well, when people dislike an app, it’s usually not the app’s logic they dislike. It’s usually the UI and the workflow. You could build an “app” company that had everything you needed for the core app but made it easy for users to control UI and UX and customize it to their own needs, and that could be useful.

Imagine a CRM tool that provided all the infrastructure you needed to easily add your own UI on top that was customized to your company far beyond anything you can do today when configuring those tools. If you consider that an “app” instead of infrastructure, then, call me bullish again on the app layer.

That’s my take. Feel free to let me know if you have a different perspective. And as always, thanks for reading.