I run TNT, a community and accelerator for AI and deep tech founders from MIT and Harvard. More people want to meet our founders than ever, and they want to meet them earlier, sometimes before there's even a company. It isn't just investors either. Banks, law firms, cloud companies, developer tools and software vendors keep asking, and more of them keep coming. I get texts and emails every day from people who want to partner with us or sponsor something. Part of that is because TNT has grown, but I think a lot of it is the market. Communities have essentially become marketplaces where founders and the people who want to work with them find each other early. That only works if it works for founders, so a big part of our job is making sure the people who show up are worth their time.
All of them want the best teams, but they can't always tell which ones those are, so they cast a wide net. That's the power law at work. In most businesses, results cluster around an average, but with startups a handful of companies end up worth more than all the others combined. In Horsley Bridge's data on hundreds of venture funds, about 6% of investments produced about 60% of the returns. VCs have always built their business around that. You can't know in advance which company will be the one, so you meet a lot of them early and make sure you're there when one takes off. In the AI era, a lot more people care about the power law than they used to, and they're showing up much earlier.
None of this is entirely new. Silicon Valley Bank was started in 1983 to bank tech startups, and by 2010 law firms were offering startups deferred fees and introductions to investors. Banks and law firms aren't trying to become VCs, and a bank still makes money from plenty of smaller companies. But they're betting on the same power law with their time instead of their money. They spend a little time with a lot of promising founders, so they're already the firm a company works with when it gets big.
What's new is how early that now happens, and the reason is speed. Stripe found that the top 100 AI companies on its platform reached $1 million in annualized revenue in a median of 11.5 months. That's about four months ahead of the fastest-growing SaaS companies at the height of the subscription boom. When a company can go from nothing to real revenue in a year, the window to meet it before it's obvious gets very short.
You can see the same logic in the startup programs that are everywhere now, like the cloud credits AWS has offered through Activate since 2013. You can also see it in companies that sell to startups first. It's very hard to get a customer to rip out HubSpot or Salesforce, so newer CRMs go after AI-native startups before they've picked anything. SaaStr wrote that if you poll current YC startups, the vast majority are using neither.
To see what an early relationship is worth to each of these players, I worked with Claude to model one startup's first six years from seed. We compared a fast-growing AI-native company with a classic SaaS company. Then we tracked what a seed VC, its cloud and AI API providers, a bank and a law firm would each earn.
Figure 1. Illustrative model built with Claude from published benchmarks. Inputs and sources are in the appendix.
Almost everyone's curve has the same shape, and the fast company pays off years sooner. The cloud and AI API providers' curve looks a lot like the VC's, because they hand out credits early and then earn on everything the company runs. The law firm is the exception, since closing a bigger round costs more but not proportionally more.
The gap could get wider from here. A lot of people expect AI to eat software, and AI-native services to take over work companies now pay people and outside firms to do. If that's right, many of the biggest customers ten years from now barely exist yet. It's hard to see why you wouldn't try to get in with them early.
The same math that sends a bank to meet a hundred founders is now running inside the startups themselves. A firm can't tell in advance which founder will matter, and a growth team can't tell which video will take off. In both cases the answer is to make a lot of cheap bets early and then back the few that work. What's changed is that agents and AI tools have made each attempt almost free, so AI-native companies can run far more of them than anyone could before.
Growth is the clearest example. The playbook is to farm content out to a lot of creators, see which format actually goes viral, and then double down on it. Cal AI's co-founder Jake Castillo described it simply: once they found a format that worked, "we could just copy and paste it across the millions of fitness influencers out there."
Cluely ran the same play at a huge scale. It reportedly paid more than 700 clippers to repost its content, and according to its founder's own posts, 1% of them drove 99% of the views. That's the power law in a single number. When its numbered video series drew about four times the views of its other posts, it kept that series going for 35 episodes.
Other companies are building the same machine. Wispr Flow has a roster of 50 creators making 500 to 600 paid ads a month. It pays them a cut of the ad spend, so the money follows whichever ads win. Clipping has even become its own market, where marketers pay people to post short clips of longer content, typically $1 to $5 per 1,000 views. It has quietly become a staple for some advertisers.
Outbound and engineering work the same way. Instantly lets a sales team test up to 26 versions of an email in a single step and put their volume behind the ones that get replies. Cursor lets an engineer run up to eight agents on the same prompt and keep the best result.
A/B testing sits underneath all of this, and the pattern was there before AI. In 2012, a Bing engineer tested a change to how ad headlines were displayed. The idea had sat in the backlog for more than six months, and it raised revenue 12%, worth more than $100 million a year in the US. Researchers who later studied Bing's experiments found the gains are fat-tailed, with outlier ideas accounting for a large share of them. Now that teams can test so many more things, the results start to look like a portfolio.
Figure 2. Illustrative, not data.
Evals bring the same idea into AI products, which is probably why they've gotten so much attention. An eval is a test that scores what an AI system produces against a written definition of a good answer. The guide to OpenAI's open-source evals framework calls creating good evals "one of the most impactful things you can do" when building with LLMs. OpenAI's president Greg Brockman put it more bluntly, posting that evals are "surprisingly often all you need." Lenny's Newsletter, summarizing a conversation with OpenAI's product chief Kevin Weil, said writing effective evals is becoming a core skill for product managers. Lenny later devoted a full episode to why evals are "the hottest new skill for product builders." His two guests have trained more than 2,000 PMs and engineers, including many teams at OpenAI and Anthropic.
The reason ties back to the power law. When an agent can produce a hundred versions of something, the hard part shifts from making attempts to knowing which one is good. An eval is how a team writes that judgment down, so it can score every attempt instead of the few a person has time to read. That's why some people now call evals the new product requirements. Before you can test anything, you have to decide what success looks like. Models can run the tests, but defining success is a judgment call, and getting it right takes real taste.
When it costs almost nothing to try something, the scarce skill is recognizing which attempt is worth backing. That's true for a VC reading a pitch, and it's just as true for a growth team looking at a hundred clips or an engineer comparing what eight agents produced. That judgment is what people are pointing at when they talk about taste.
It also doesn't work everywhere. The research on Bing's experiments is careful to say that running more, smaller tests is the better strategy only when outcomes are very fat-tailed. It works where attempts are cheap and the upside has no ceiling. Anywhere else, a lot of small bets just spreads you thin.
If you're a founder, people are betting their time on you because of what you might become. That puts you in a stronger position than you might think, so it's worth knowing what each credit, intro and early account is really buying. For everyone else, including communities like TNT, showing up early is becoming the default. What still sets you apart is whether founders find the time with you worth it.
I worked with Claude, Anthropic's AI model, to research the benchmarks in this piece, build the model behind Figure 1 and check the sources. Claude checked every source against the exact claim made here and recomputed the model's numbers independently. The figures are illustrative. They show the shape of the economics, not any real company's results, and the inputs below that are assumptions are labeled as such.
Figure 1 inputs
Figure 2 is illustrative and not based on data.
Sources