My view on AI Startup Landscape today

I currently look at the SaaS Landscape as an onion: AI frontier lab in the core, and industry incumbents (e.g. finance, coding, medical) at the outer layer. And AI startups are positioning themselves in between.

Since the kickstart of the AI race a few years ago, there have been some remarkable AI Startup Unicorns: Harvey, Perplexity, Cursor. At the time, their thesis and product market fit was amazing.

Since then, Frontier AI labs have not only caught up but rendered their products obsolete overnight.

Perplexity: ChatGPT, Claude, and Gemini all cite web sources when answering your question.

Harvey: Harvey tried to post-train their own model, but Opus 4.7 (and Mythos) and ChatGPT5.5 have become good where they beat it on their own BigLaw Benchmark. Then is the product differentiator mainly the UI? I say no; everyone has connectors at this point (under the hood MCPs, or skills), and with Claude CoWork or Codex, you just let them control your computer to use any software (e.g. Powerpoint, Word, etc.). If the models are good enough to wield these tools and skills, the moat becomes really weak.

Cursor: This is an interesting one. I thought Cursor was the best AI native startup. When it first came out, I was blown away. Essentially, they forked Microsoft VisualCode IDE, and added auto-tab completion of code (fast!) and Sidebar chat to implement the code change right there. Github Co-pilot (which Microsoft owns) was first to the market, but the quality was poor and it was really slow. But the auto-tab completion was a banger; it felt like AI was reading my mind and knowing what I wanted to type next even.

Then personally, I moved off Cursor because it broke when Microsoft cut off support for the Python LSP in Cursor. In other words, when I was working on Python code, everything was greyed out (normally, LSP allows Python syntax highlights so it’s easier for the developers to read through the code; without it, reading the code takes a much longer time, almost impossible). It’s reasonable; Cursor took their product and ran away with it, and Microsoft wasn’t seeing a single dime from Cursor’s success. So Cursor needed to build LSP for every coding language (which is not easy), and there was zero customer support from Cursor. I remember having to dig through their forum and reddit to triage and troubleshoot my problem (even tried some hack my coworker suggested that violated their terms of service).

Then around this time, my peers told me about Claude Code. At first, I didn’t quite understand it. You had to open a new terminal window in an already crowded screen real estate, and it wasn't clear how I was supposed to interact with it. Should I be looking at the file on my IDE that I asked Claude Code to reference? What should I ask it to do? So far, most of my code has been written by me, and Cursor stepped in when I had questions about syntax or filling out boiler plate or repetitive code.

Then, I understood why Claude Code is so good and got addicted. I deleted Cursor the next day. It changes the paradigm of writing code. Claude Code became an AI senior software engineer where I tell it what to do and it does everything end-to-end. Most importantly, when I get stuck on a bug, it’s super tedious to run and verify; Claude Code does this automatically! So at this point, why do I need an IDE (e.g. Visual Code)?

This single handedly destroyed Cursor in my opinion. Not because they have a negative profit margin (aka burning cash). I don’t know any developer that still uses Cursor. All the rage is Claude Code, Codex, and open source coding harness that uses open-source models (e.g. OpenCode, pi). And I think Cursor is shifting in this direction too, where there is a UI for Agent Manager, and you chat with an AI Coding Agent to write the code. And at this point, you wonder: why would I use Cursor over Claude Code?

Last week, Cursor got into an agreement with SpaceX where they either get bought out at $60BB or get paid $10BB if they help SpaceX train a new coding model. For SpaceX, they didn’t have success coming up with a good coding model, and they are sitting on a lot of idle compute and they can spend their overpriced SpaceX options to pay Cursor. For Cursor, I don’t know how much coding data they were legally able to get, but I think the most valuable data is user feedback on its code suggestions. However, I feel like this signals the end. Either Cursor comes up with the coding model that blows Opus/GPT out of the water by the end of year, or they go into irrelevance.

And other startups in this field but don’t have time to mention: Lovable, Pika

So given this, I’ve been looking at startups that got a lot of hype (especially 3 years ago) and looked at where they are at now. And I don’t think any of them have made it. And I’ve been thinking about what’s been happening.

I had a thesis 3 years ago: Sam Altman promoted young people to go build AI start ups, and now when AI labs need to reach profitability, he’s building stuff built by AI startup unicorns that have proved product market fit. When there was a stable enough LLM, I think Altman did 2 things right: productionalizing into ChatGPT, and making this available via API. Use cases were not clear yet. But that was ok: as the previous president of YC, he knew start ups were the best way to stress test different use cases for AI most efficiently. So around this time, I saw 2 things: tightening of junior engineer positions in big tech, and push towards start up funding (e.g. YC went from 2 funding cycles per year, to 4). Young, jobless, hungry developers were throwing themselves at AI, and were building ChatGPT for X, Cursor for Y, etc.

Some failed, but some succeeded very well (unicorn status). And they were all built on top of OpenAI/Anthropic’s models.

So, given this relationship, if you are a PM at a frontier lab, your job is easy: copy what successful AI startups are doing (if worth going after), and copy other labs’ features.

And I think that’s why Claude and ChatGPT have very similar features, and why OpenAI is now going all-in on Codex after seeing the success of Claude Code.

I think there are 2 important things the AI labs are facing: the performance of LLM is plateauing (yes even Mythos), and they don’t have a good picture of what’s next after the LLM. When LLM was productionalized as a chat interface, there was nothing like it before. It opened up so many possibilities. The product was… unfamiliar. But now, almost everyone has used some of it so far. And the difference between the performance of the latest model and the previous one… is unnoticeable (based on trust-me-bro benchmarks released by the labs). LLM at core is very simple: every turn of chat involves copy-pasting the entire previous conversation since LLM at heart doesn’t have memory. And frontier AI labs are all-in on LLMs right now: Anthropic doesn’t even have a multi-modal model (e.g. image generation), and they are not in robotics. So they are mainly relying on clever addendums to LLMs: tools, skills, MCP, structured output, coding harness, browser-use.

So I see the AI frontier labs turning inwards. And the industry incumbents are more and more adopting AI. Thus, the AI startups are getting squeezed from both sides.

So I think from startup perspectives, thin light LLM wrapper is no longer a viable business strategy. Also, SaaS that can be built using ClaudeCode over a weekend is also no longer viable since it can be replicated quickly.

To me, successful AI startup needs to have 4 things:

1. Expertise/Domain Knowledge that others may not have. This depends on your work experience or expertise. I don’t know anything about medical for instance, but I know how Autonomous Vehicles (and Uber systems) work.
2. Flywheel: to gain speed after initial launch, your startup needs to have a flywheel effect where the data collected as part of your product workflow improves the product and/or the company. Also, if the product can be used to further improve your product/company, that’s a bonus.
3. Will Frontier AI lab make a similar feature? In other words, will a feature/product from the AI lab make your company obsolete in a day? This is evident if you are a thin AI wrapper. I think an easy way to avoid this is to go into areas where it’s not lucrative for the AI labs to go into. This may involve a lot of forward deployed, and/or boring work. Or go into a niche area (similar to point 1): even a frontier AI lab can do so much.
4. Do you actually enjoy the work? I saw 22 year old founders claiming they are deeply passionate about compliance or audit work; give me a break. If you are gonna spend 60-80 hours per week on this, it better be in the area you are interested in. What does your Youtube/Reddit algorithm look like for example?