Jeff Smith

Today @Ghost crossed $10M ARR, as a bootstrapped non-profit foundation building open source software. Indie publisher revenue earned with Ghost now ~$130M, and accelerating. The world of technology is shifting rapidly, and so is the world of media, creators and journalism. It's hard to keep up with, and even harder to predict. My strong belief, though, is that open software that you own and control is going to be even more important and relevant in the future than it is now. So we're going to keep building it.

:taps the sign: x.com/jasoncwarner/s… This isn't about Cursor, so forget the name used. This is about what is happening in the world. Cursor, as I understand it, is finetuning chinese models so at least they realize what I'm about to say. Let's walk through this so we fully understand it. In the '90s a bunch of tech companies built out the internet. Those tech companies became critical infrastructure and were massively rewarded for it and became new tech giants. That infrastructure allowed a whole new type of company to exist, companies like Amazon and Google and eventually Facebook. And for companies like Microsoft to make a transition if they could see the future (folks like IBM and Oracle didn't see the future). In the '00s, those same companies built out new critical infrastructure called hyperscaler clouds which enabled a whole new generation of company to exist. Those hyperscalers became the most valuable companies in the world bc they controlled the most valuable commoditized asset on the planet at the time. And the new companies they enabled, the likes of AirBnB, Uber, GitHub, Shopify etc, became great companies in their own right...but nothing like the scale of AWS, Azure, GCP etc. Now it's happening again. Intelligence is becoming the new critical infrastructure upon which every company on the planet will build. It is enabling new types of companies to exist that couldn't before. And like the previous transitions, companies that see this transition can create and capture value in bold new ways. Right now it's down to a handful of companies. Google is the only full stack player: They have dirt, datacenters campuses, TPUs, GCP, and Gemini. By default, they lap the Amazon and Microsoft (don't get me started on Microsoft's continued fumbles here). They are 100% fully vertically integrated. When they got good at building Gemini, it all fell together for them for the next two decades. OpenAI, Anthropic (best independent lab on the planet and only investable frontier lab IMO), and perhaps xAI round out the frontier capable players. Poolside, Chinese labs (with Poolside, the most efficient labs in the world...see notes at bottom), and (perhaps) SSI will eventually become frontier capable. Those round out the list. Three things matter for the next decade: 1. Energy Infrastructure 2. Compute Infrastructure 3. Intelligence Infrastructure Most things after those are rounding errors. 1+2 = powered datacenter campuses 3 = frontier model providers or domain specific model providers 1+2+3 = full vertical integration At Poolside we have a saying which is "Everything collapses into the model" which means that eventually all things we think of as valuable become part of the models (and agents*) that get produced. *Note: first party agents from model providers will *always* be better than third party agents that come from non-model providers. This is simply a reality that modern agents and models are trained together. This really really means a large company producing an agent without backing of it's own model is especially vulnerable. Another way to think of it is that the surface area available to applications is exactly equal to the current capability gaps of this generation models. Each time models become more capable, they eat up more surface area available to third party applications *that are simply arbitraging the current model generation capability gap*. This doesn't mean third party apps aren't valuable, but it does mean if the only value third party apps have is the current capability gaps of models, those third party apps have diminishing value. This is not debatable in 2026. It arguably wasn't debatable when I wrote the post a full two years ago either, except people are really bad at living in an exponential. You know who isn't bad at living in an exponential? Jensen. Nvidia gets exactly what is happening while wall street futzes around with "capex buildout costs", Jensen knows exactly where every gigawatt in the world is going because he knows the future. Do you think you are smarter than Jensen? I know I'm not. How should non-model companies (and companies who are incapable of building models) behave in this moment? It's simple but not easy: Get good at building smaller but still capable models that push *your value prop*. Do not cede your ground to third party model providers hoping you can hold on and survive, you won't be able to with the fullness of time. The tech is just too powerful. If you are a company with a mid double digit billions market cap and above and you are not making models, you are default long-tail dying right now. It is your job to figure out how you don't die. "But Jason, what about data as a moat?" Data is valuable but if your *only* valuable asset is data, you are fighting a multi-pronged war. And if this is true it is even more paramount that you get good at training your own models. My god, this makes the whole thing more existential for you! "But Jason, it costs so much to build models! How can one compete with how much money the model providers raised?" Here's real dirty little secret of the AI industry. That's not real. epoch.ai/data-insights/… OpenAI 2024 compute spend above ^^. It cost them roughly $500m to build frontier model in 2024. But they spend $4.5b on RnD, which means, clusters for their researchers. The dirty little secret of the AI industry is there is no such thing as a gigawatt training cluster, there is no such thing as networking a million GPUs to train a model. The cost to train frontier AI is people with knowledge, a system capable of experimenting to find and iterate on model recipes, and $500m to train the final recipe. The first two are the hard part. The third one is just cost of doing business. And we wrote extensively about Poolside's way of building models which allows Poolside to do things that frontier labs do but at a fraction of the cost and fraction of the time. We call it the Model Factory and it's part of our secret weapons (along with our proprietary RL research): poolside.ai/blog/introduci… Every single company worth double digit billions not getting good at training their own model is the modern equivalent of saying "I can't possibly run a database as good as Oracle therefore I shouldn't try and just pay Oracle to do it" or "It costs too much for me to have engineers build and maintain our software, I'll just pay Accenture and Microsoft to do it". How many of us would build our forever homes (our companies) on two year leased back land (api call to model provider where we pass all our data to them)? I prefer to own the ground my home is built on. Get good at building your company specific models or look back in 10 years and realize you IBMed yourself.

i've fully converted to @conductor_build — feels like going from typing with two fingers to having eight arms.

✅ First day at @NotionHQ

Two tools that I’ve come to love: @usemonologue. More whimsy, less UI, and great dictation. @conductor_build — the team is cooking! Tried it months ago, came back, and barely recognized its thoughtful git worktree integration and opinionated UI. Paired with @linear, each ticket becomes context for a fresh agent in a new worktree. Suggestions on how to set up for Ralph, lmk

I'm calling it Dotly. Give it a try! TestFlight here: testflight.apple.com/join/Dn5YsQYY