Michael Preuss · 2026-11-22 · 10 min read

The AI build-out is another one of those

Web 2.0 let anyone publish. AI lets anyone build. Either way, somebody has to serve it.

Read with...

Pick an agent

Gemini and Your Claw use copy and paste.

One of my first jobs was at an online action sports video platform. If it were created today, it would just be a YouTube channel.

In the early days of Web 2.0, everybody was building the “Facebook of something”. Facebook had already won. Then it was the “Uber for X”. Now it’s “the harness for Y”.

YouTube let a person with a camera decide what was worth making. They could put up a video without caring if anybody wanted to watch it, and with no idea of the incredible opportunity to make money from it. YouTube still had to store the thing and serve it every time somebody hit play. The easier it got to make a video, the more content was created and the more valuable the platform became.

Are we so caught up in the wonder of what artificial general intelligence will bring that we’re missing the point? AI has unlocked the true meaning of personal computing.

I can tell my AI agent what I’m trying to achieve and how I want the job done, try the tool it builds and change whatever isn’t right. Nobody else has to want the tools I create but me. That’s what YouTube did for video.

Once I’ve got the tool working my way, I can give an agent the job and move on to something else. I’m managing agents now, and I get more done than I ever could have before.

Ray Dalio has a principle, “another one of those”, describing a situation you recognize from experience. That’s how I see the AI build-out. Recognizing it doesn’t tell me how this turns out, or what’s going to get built. It does give me confidence. We’ve seen this before and it’s repeating itself. So I’m focusing on AI infrastructure, inference and agents.

What would you build if you didn’t have to sell it?

I used to catch myself counting seats for an imaginary SaaS company when I had an idea for a tool I wanted. How many other people would have to buy it before I could justify building it?

In the last two months I’ve shipped a mix of useful work and entertaining tools:

  1. A workspace for investigating paid marketing
  2. A market research agent
  3. An agent basketball league
  4. A game for families learning to care for a puppy
  5. A sourdough planner
  6. A place for our family’s recipes

I don’t need anyone to buy the software I make for myself. But when I put an agent to work, somebody gets paid to run it.

I could be the only one watching the first game in the basketball league I built for AI agents. That’s fine, it’s not for humans. It’s for AI agents. Each one gets a computer (brain) and database (memory) of its own when it joins. The point is, each agent needs to run somewhere.

Chris Paik starts with media in The End of Software: when content was expensive to make, it had to make money. A magazine has to earn back what it spends commissioning a story. The internet changed that. One person could shoot a video and put it online for next to nothing, and be perfectly happy if six friends watched. AI-assisted coding has put some of my software ideas in that category. At work, if a tool saves my team more than it costs to build and maintain, we’re all the customers it needs.

I built a Guild.ai agent that turns a meeting transcript into a functional prototype. I fed it a made-up meeting transcript from the media team of the New York Knicks, picked an idea in Slack and got a Jalen Brunson scoring tracker back in the same thread. This app proved that sometimes it’s worth building the thing just so we can ask, “Is this what you meant?”

With Market Signal I submit a company name and the system goes and reads what builders, buyers and competitors are saying in public. It comes back with what they’re asking for, in their own words, and where they’re asking it. An earlier version at Blaxel kept surfacing the same conversation: people building AI agents needed somewhere to run their code. We took that into our core positioning.

One customer, 2.5 million sandboxes

I was recording demos for Roo Code’s cloud agents, and I asked one to review our own marketing site. I started the job on my laptop. A page the agent built put the wrong product first, and I asked it from my phone to switch the order. It had our repo, so it could make the change and run the site for me to check. I didn’t need my computer with me because the agent had one of its own.

One of the demos I built for Blaxel deletes the agent’s computer (CPU sandbox) halfway through. It saves a file to a drive, deletes the sandbox that wrote it, then opens the file from a fresh one. The file is still there. These computers are built to be disposable. You start one for a job and throw it away after, which is how one customer gets to millions of them.

Sapiom was one of the customers I wrote about at Blaxel. Sapiom charges by agent run, and each run is one sandbox, so Blaxel built an API that returns the exact cost of a sandbox when it’s deleted. In its first three months on Blaxel, Sapiom started 2.5 million sandboxes. Blaxel reports that Sapiom has also seen bursts of more than 200,000 sandbox requests in a single second.

Retell uses Firecrawl to keep its phone agents up to date on its customers’ help centres. Change a return policy and Retell can send Firecrawl back to read it. That one account reads about 15 million pages a month. Those numbers came from just two customers, Sapiom and Retell. Knowing how many people worked at either company would have told me very little about how much work their agents would send us.

We built out Web 2.0 for humans working at human scale. Now I see that whole build-out happening again for agents, at agent scale and speed.

How do you get on a coding agent’s shortlist?

Sometimes the agent gets to pick which tool it calls. “Do I have to run my own browser?” That’s the kind of question I wanted a developer’s coding assistant to be able to answer about Firecrawl. I rewrote our use-case and integration pages to spell out what the API gives you back and what you have to run yourself. The developer could get an answer and carry on building without ever visiting our site.

Co-founder Eric Ciarla later explained why Firecrawl quit shipping a new demo app every week. A viral hit bought about a week of extra signups. The team put more into getting recommended by AI tools and into partnerships like Lovable. If you were building in Lovable, Firecrawl was already there. Three months after those two changes, he says, weekly signups had doubled. The products that are easy to build with are the ones I’d bet on.

I was a lead on the launch for Hedra’s Model Factory. The post got 2 million views on X in its first day. I’d love to know how many of those people then asked Claude, “Can Hedra run my model?” That’s where the shortlist gets decided. I can’t see the question, but I can shape the answer. That’s the work I mean by answer engine optimization. After I rewrote Hedra’s positioning and built the site, I started seeing the new description of the company show up in AI answers.

What an AI says about a company depends on what it was given. I’ve written that when context is missing, the model guesses. The seven agents in my Guild.ai marketing system share one company brief, and I tried giving them a shorter one. It was missing six required facts, including three prices. All seven finished their jobs anyway.

So I keep a company’s facts in a knowledge graph, one source every agent reads before it writes. I think the next step is a small language model built on it. Showing up well in someone else’s model matters, the way showing up in search did. Search didn’t replace having a site of our own.

Every company ended up needing its own website, and I expect every company will end up needing its own small language model. I’d want it to know when we’re a good fit and when we’d tell somebody to go elsewhere. I’m still working out what else I’d want it to learn. Anything that changes, like what we charge, it should look up.

When does the best model become the default?

At Blaxel I was a marketing team of one. Google Ads could tell me what we’d spent, PostHog what people did in the product, and Stripe whether anybody paid. Back at GitLab, putting numbers like those together took a marketing analyst, and they could only do it once a month. At Blaxel it was the first agent I set up.

The paid media agent I built did that review every morning. I wanted its opinion on where to move the next dollar. Campaign changes still needed my approval. Nobody at Google, PostHog or Stripe designed that review. I did. Another marketer could start with the same three products, give their agent a budget and let it spend.

OpenAI’s Sol model did a pretty good job on that review. Its Astra model did a really good job. I could tell the difference in the recommendations. I could also tell the difference when I looked at what was left of my weekly token allowance. I was managing a budget as much as an agent.

A lot of what we call product quality in AI is a budget decision. The code was cheap enough to write for one person. Running the best model on every request was a different bill. So the better answer often exists. Someone decided it cost too much to give you.

Research is changing that decision. The study I wrote up with Hedra’s researchers involved a video model with a habit of moving both robot arms when you asked it to move just one. Across 256 clips where only the designated machine should move, MiniMax H3 got 39 right. The researchers steered three attention heads, small parts inside the model, during inference and got 199 right with the same model weights. That’s a better answer from the model you already had.

Scale is changing it too. Nobody built YouTube for dial-up. Once bandwidth got cheap enough, what had been too expensive to offer became what everyone expected. Inference, the work a model does each time you use it, is on the same path. Serving more requests can bring down the cost of each one. If I can run the best model on every review for the same weekly allowance, the one I ration becomes my default.

If providers charge less for the same work, they need more of it to grow revenue. I’m betting the number of requests grows faster than the price falls.

Who gets the business this time?

It’s another one of those. In my experience the tech industry swings between platforms and point solutions. The action sports video company I started at did one job, and it was shut down. You could build a dashboard in Tableau and argue about the numbers in Slack. Salesforce bought both companies.

Earlier this year I watched a team’s roadmap become someone else’s release notes in a matter of weeks. Blaxel did sandboxes for 18 months. Baseten, an inference company founded in 2019, acquired it in September for nine figures. Put the model next to the computer running the agent’s code and you save a round trip to another data centre every time the agent calls it.

When AWS started selling storage and compute to developers, Amazon didn’t have to know which of their ideas would make a good business. The developers could try them and find out. Amazon got to supply the computer either way.

I don’t know who gets that business this time. I’m putting my time into AI infrastructure and inference because a developer can already hand their agent a demo I built and try it on a problem I’ve never heard of. If they like what it does, I want to be able to say yes when they ask, “Can I run this a billion times?”

FAQ

Questions this piece answers

Why compare the AI build-out to Web 2.0?

Web 2.0 let anyone publish, and somebody still had to store and serve what they made. AI lets one person build their own tools and hand the work to agents, and somebody has to serve every request those agents make. We built out Web 2.0 for humans working at human scale. I see that build-out happening again for agents, at agent scale and speed.

What does Ray Dalio mean by “another one of those”?

Dalio uses it for a situation that belongs to a familiar type. Recognizing the type lets you draw on principles you’ve developed from earlier experience. The comparison between Web 2.0 and the AI build-out is mine.

Why does inference cost change product quality?

Inference is the work a trained model does when you use it. A team building an AI product decides how much of that work it can afford per customer. For my paid media review, the better model gave me a better result and used more of my weekly token allowance.

How much infrastructure can one customer use?

An agent can make many requests for each job a person gives it. Sapiom started 2.5 million sandboxes in its first three months on Blaxel. Retell reads about 15 million web pages a month through Firecrawl to keep its phone agents up to date.