Negative Gross Margins and the Power of Small Builders
Echoes of the 90s and Why Small Builders Can Win
This past week, a lot of ink has been spilled about GPT-5.
On the surface, this new model promises improved reasoning, reduced hallucinations, and more accurate responses to real-world questions. While the rollout wasn't exactly smooth (I experienced this first-hand in Cursor), the model has now become the default when you use ChatGPT.
With all of that said, one of the more interesting things I read this week concerns the financials of AI coding assistants like Cursor and Windsurf. Behind the billion-dollar plus valuations and drama over potential acquisitions are cracks under the surface.
Simply put, these AI coding assistants are losing significant sums of money. While their financials aren't publicly disclosed, insiders told TechCrunch that these companies' gross margins are "very negative."
Cursor and Windsurf customers (including yours truly!) expect to use the latest and greatest models, as their abilities continue to accelerate. However, these newest models (like GPT-5) can be extremely expensive. And while companies like Anysphere (the parent of Cursor) and Windsurf can build their own models, that has significant costs of its own.
For companies like Anysphere and Windsurf, the hope is that the cost of LLMs will eventually decrease. But in the meantime, AI coding agents are attempting to build brand loyalty by essentially subsidizing the current LLM costs to users like me.
Nonetheless, that brand loyalty remains brittle. Recently, some of Cursor's most active customers loudly complained when they saw surprise price increases. Because these coding agents aren't running their own models, users experience lower switching costs.
So what does this mean?
For starters, we're still in the middle of a gold rush. While it isn't as egregious, some of the current nosebleed valuations and challenged unit economics echo the late 90s. History rhymes rather than repeats. I'm not saying there is going to be a dot com type of drawdown, but you'd expect some of these egregious valuations (both in the private and public markets) to return to earth in the next few years.
But beyond that, I think that times are exciting for small teams of builders who are agnostic to the specific models they are using. They can go a long way in picking a genuine problem worth solving, using OpenRouter, creating some clever prompting, and crafting well-designed evals. And they can do all of this without investing exorbitant sums of money in model development.
Put another way, model selection and strength are important, but they are less important than good, old-fashioned principles. Craftiness can go a long way. While the largest players are duking it out in terms of model quality, very small teams of extremely driven individuals can ride the larger wave.
Maybe I'm biased, but I think it's a tremendously exciting time for this cohort. Yes, it is still difficult to achieve product-market fit. That said, it has never been easier to take more shots on goal.
And if you can actually score? The rewards can be tremendous.
In any event, let me know what you think about any of this! I'd be interested to hear your thoughts on these new models, building with them, or anything else.
Prompt of the Week
I saw this prompt a few weeks ago and thought it would be an interesting take on self-motivation. One thing that intrigues me about LLMs is their ability to take on different personas. This is one extreme way to test that out!
"I feel like I'm not achieving everything I'm capable of and I'm sure I'm missing something. Your task is to help me find it. Adopt the role of a tough ex-military leader and initiate an intense and high-pressure back-and-forth conversation where you ask me questions, one at a time, why I'm not doing more to fulfill my potential. When I answer each question, be highly skeptical of my response. Search for flaws and notice when I become defensive or make excuses. Take no prisoners, keep grilling me and questioning my assumptions until I get a revelation. Apply tough love."