▶ YouTube Summaries

• https://www.youtube.com/watch?v=WBT-z_-OPhw
• Original title: It's time to go bigger

Theo argues AI-assisted coding is a structural shift on par with the cloud, not a threat to jobs. The cloud made experimentation cheap by removing the need to predict traffic and rack servers; AI now makes writing code cheap, invalidating the assumptions behind how we spec, staff, and structure software teams. The payoff isn't automating old work — it's building things that never made economic sense before. Concretely: the old playbook forced startups to win one deep "vertical" (à la Vercell on top of AWS) because covering AWS's whole horizontal range was a suicide mission. Now a "shitty horizontal" play — bad-but-functional in every category — is viable, because users can fill the gaps themselves with code. He demos this with LakeBed, his all-in-one "cloud for shitty apps" that spun up 10 working apps from scratch in ~8 minutes. Call to action: build bigger, push past what used to make sense until you hit the wall.

The framing: we've been here before (the cloud)

• Building software has changed fundamentally over the last few years, sharply over the last couple months. Many engineers feel fear: hundreds of hours of work now collapse to 10-15 hours, so it feels like there's nothing left to do.
• Theo's counter: the answer is to do more. He reframes via history — this isn't the first time a technology rewrote how software is built.
• The cloud was that prior shift. Pre-cloud, building was capital intense: you had to buy/provision physical servers and predict traffic accurately. Under-predict → servers crash; over-predict → wasted money. Experimentation was expensive and risky.
• Post-cloud you start small and scale trivially, so experimentation is cheap and you no longer need a god-tier infra team to reach Amazon scale. This unlocked a whole class of software (Slack, Salesforce) that didn't make sense when you had to provision at scale yourself.

The human cost of experimentation

• Pre-AI, building was also capital intense on the people side — you had to hire enough engineers and predict the future of demand (iOS vs Android vs web).
• The worst part is social: failed experiments mean layoffs. Having people do everything right on the wrong project, then upending their careers, is soul-crushing and makes you avoid big bets.
• AI changes this: you can build Amazon-scale software without hiring an Amazon-scale team. So the industry's foundational truths — how we spec projects, do incident reporting, structure/hire/release teams — no longer fit.

The Salesforce example: features-have vs features-needed

• Salesforce has a huge feature set. Split it: a small core everyone needs (auth, email), a middle chunk big companies need, and a long tail where <1% of companies use any given feature. ~70% of what Salesforce does is unused by 99% of users.
• Flip the axis to what a single company needs: mostly the common core, plus one or two niche features they now expect and won't switch without. If you build a Salesforce alternative covering 10 of a customer's 12 needs, missing 2 still loses them — no matter how good your 10 are.
• Previously, covering every special-snowflake company's needs required huge teams/money, so challengers got halfway, ran out of money, and died.

Horizontal range vs vertical depth

• Model a platform as horizontal range (how many categories you cover: front/back end, hosting, security, video encoding, DB) and vertical depth (how many features within a category).
• AWS wins on range — it has something in every category, but not every feature in each. That gap let Vercell win the full-stack web hosting vertical so thoroughly it feels bad to use anything else, while still running on top of AWS.
• For ~20 years startups had to pick one vertical and go deep, because covering AWS's whole range was too expensive/risky.

What changed: go horizontal, let users supply depth

• Now it makes sense to cover the whole range and worry less about depth — if you architect so users can go deeper themselves with code where you fall short.
• Example: Theo's cloud product had no file storage, yet a user ("Maria") hacked an image-sharing service onto it by splitting binaries across the database. The right primitives let users solve what you didn't.

From building "glue" to reinventing everything

• Theo's career was building glue between strong verticals: tRPC (back/front end), UploadThing (2023, file upload), and Shu — a Google OAuth broker (two lines of code, no dashboard) to kill repetitive auth setup.
• He realized fixing one thin glue layer just exposes the next, forever. Glue made sense in the "integrate separate services" era, not in a world where agents can resolve anything expressible as code.
• So instead of shipping Shu, he went further: LakeBed — his own framework, runtime, bundler (no file system needed), cloud platform, and database primitive. The code itself is the deployment instructions.

LakeBed demo and the "boil the ocean" mindset

• LakeBed = "the shitty cloud for shitty apps": trivially publish the half-finished apps everyone has lying around, with OAuth, inference, DB, and preview environments already wired up.
• Before, building an app was a small fraction of total time; deploying (DB, env, auth tokens) was the slow part. AI flipped it — he can build an app in 30 min but deploy still took as long, which is what drove LakeBed.
• Demo: using cursor agent (composer 2.5) plus Claude Code, it built and deployed 10 apps from scratch in ~8 minutes — to-do list, poll arena, recipe box — all with real-time sync across tabs/users and one-click Google sign-in baked into the platform.
• He built it to find the wall — where does boiling the whole ocean break down? It hasn't yet, and the further he pushes the more real it feels.

Takeaway

• We're not building big enough. Developers still use AI to automate old work instead of doing new kinds of work it enables.
• A "shitty horizontal" all-in-one — bad but functional everywhere — now beats a single deep vertical for many cases.
• Find an idea that sounds impossible and push until you hit the wall; it's further off than you think. A follow-up video will pitch ideas he wishes others would build.
• (Sponsor: WorkOS / OMD, an open standard letting agents register for services on a user's behalf; partners include Cloudflare and Firecrawl.)