One week ago I announced Grazier Ventures and Aqen.ai. Since then I’ve been asked a version of the same question about a dozen times: how are you actually going to build more than one company at once?
The question is fair. It’s also built on an assumption that’s five years out of date.
Most people’s mental model for starting a company still treats intelligence the same way 2015 did — as a scarce input, bundled with a human, billed by the hour, and deployed one problem at a time. Under that model, founding four things in parallel is either arrogance or delusion. Pick one.
But intelligence isn’t priced that way anymore. And once you internalize what it is priced at, the math of how many companies one builder can credibly start in a month shifts so far that the old intuition becomes actively misleading.
This post is about that shift. Specifically, it’s about two different economic curves that most people are collapsing into one — the marginal return of intelligence and the marginal return to intelligence — and why getting them confused is the reason so many smart operators are still underestimating what a single builder can do right now.
Two prepositions, two different curves
In economics, the marginal return of an input is what you get from adding one more unit of it to a production function. The marginal return of labor. The marginal return of capital. Each additional unit buys you some additional output, usually less than the last one did. That’s the diminishing-returns curve you drew in Econ 101.
The marginal return to an input is different. It’s the economic rent that accrues to whoever owns that input. Returns to labor flow to workers. Returns to capital flow to owners. The question isn’t “how much output does the next unit produce” — it’s “who captures the value that input creates.”
For most of human economic history, these two curves, as applied to intelligence, were functionally the same. Intelligence was bundled with humans. You couldn’t buy marginal cognitive work separately from a person’s time. So the returns ofintelligence and the returns to intelligence accrued to the same place — the skilled human — and nobody needed to distinguish between them.
That bundle just came apart.
The marginal return of intelligence is collapsing
This is the part that sounds bearish.
The cost of the next unit of inference is falling on a curve that looks like bandwidth did in 1998. Per-token prices have dropped another order of magnitude since I started writing about this. The gap between the fifth-best model and the fiftieth-best is closing. Any task where the input is well-defined and the output is checkable is getting cheaper to solve every quarter.
The follow-on effect is exactly what you’d predict: applying more intelligence to a bounded problem has fast-diminishing returns. The tenth agent reading the same PDF tells you what the first one already said. Running three parallel research loops on the same prompt doesn’t triple the insight. I wrote about the symptom of this a few weeks ago when I called out tokenmaxxing as lines-of-code thinking for the agentic era — companies ranking engineers by token consumption are measuring the input on the side of the curve that’s flattening.
If you zoom out, the marginal return of intelligence, as a raw commodity input, is trending toward the same place every other infrastructure commodity eventually lands. Cheap, abundant, undifferentiated, and strategically uninteresting on its own.
So far, so bearish.
Enjoying this? I write about AI implementation and engineering leadership every week.
The marginal return to intelligence is exploding
Here’s what most of the discourse is missing. The returns to intelligence — the value captured by whoever is positioned to wield it — are doing the exact opposite.
They’re compounding. Fast. And they aren’t flowing to the intelligence itself.
They’re flowing to whoever owns the bottleneck around the intelligence. The judgment about which problem to point it at. The taste to recognize which output is actually good. The context that lets an agent succeed instead of producing plausible garbage. The coordination across everything a company has to do that isn’t the product — what I described last month as the non-product surface area.
This is the same pattern you see every time a previously scarce input gets commoditized. When bandwidth got cheap, the returns didn’t flow to the pipes — they flowed to the companies that knew what to put through them. When compute got cheap, the returns didn’t flow to the servers — they flowed to the people who knew which applications to build on top. Intelligence is next. The returns won’t flow to the tokens. They’ll flow to the judgment wrapped around the tokens.
And that judgment is still almost entirely human, still almost entirely scarce, and — critically — still almost entirely unbundled from the intelligence itself. For the first time, you can actually ask the question: given a builder with good judgment, how much intelligence can they direct?
The honest answer, as of May 2026, is: quite a lot more than one company’s worth.
Four ventures, one builder, same week
This is where I stop making the argument in the abstract and show you what it actually looks like under the new math.
One week into Grazier Ventures, the portfolio has four things in it. Here’s what each of them is, and more importantly, what each of them teaches about where the returns are flowing now.
Aqen.ai is the AI cofounder for solo founders. You bring the idea; Aqen coordinates everything else — formation, finance, marketing, branding, and product — against a shared Company Brain that gives every agent founder-level context. Aqen is the substrate. Every other Grazier venture gets built on it. Aqen’s whole job is to absorb the non-product surface area that used to require a team, so a single builder’s judgment goes further. It’s the part of the thesis that says: if the returns flow to judgment, then the highest-leverage thing you can build is the system that lets more judgment get deployed per unit of human time.
Tidehelm is the advisor founders, investors, and CTOs call before the hardest decisions — CTO mentoring, fractional technical leadership, AI strategy, and board advisory. Tidehelm is the operating venture. It’s where fifteen years of scaling engineering organizations through IPOs, acquisitions, and a unicorn goes to keep earning. Tidehelm is interesting on this post’s thesis because it’s the one venture where the human is the product — the judgment isn’t wrapped around the intelligence, the judgment is the intelligence. Tidehelm is what happens when you separate the part of your work that’s actually irreplaceable from the part AI can now absorb.
Highball Platform is an AI-native data platform for real-time trains. Infrastructure, APIs, and tooling for the next generation of rail and transit applications. Built and operated by one. Highball Platform is the infrastructure bet — a specialized data platform in a legacy industry that, five years ago, would have required a team of six and eighteen months of capital runway to even prototype credibly. Now it’s one builder. The work didn’t get smaller. The leverage got bigger.
Highball.app is real-time train tracking for families and rail enthusiasts. A consumer iOS app built on top of Highball Platform, where AI turns raw transit feeds into a delightful, personal experience. Highball.app is the consumer surface — and the proof that the platform is real. Platforms that don’t have a product running on top of them are just architecture diagrams. Highball.app is what happens when the infrastructure bet has to actually serve a human being who just wants to know where the train is.
Four ventures. One builder. Same week. Different categories, different business models, different end customers, and — critically — a different answer in each case to the question “where does the leverage come from?”
Aqen makes the judgment of any builder go further. Tidehelm is judgment as the product. Highball Platform is judgment aimed at a legacy industry. Highball.app is judgment aimed at delight. The common thread isn’t AI. The common thread is that every one of them is a structure for directing cheap intelligence with scarce judgment — and none of them would have been possible for a single person to start in 2020.
Why this inverts the old startup math
For the last twenty years, the scarce input for a new company kept moving.
First it was capital — you needed millions to buy the servers. Then it was talent — you needed a team that knew how to stand up a modern stack. Then it was distribution — you needed to outspend someone on ads, or get lucky on a platform. At every stage, the answer to “what do you need to start a company” included a list of humans who had to be hired, paid, managed, and retained to carry the parts of the business that weren’t the product.
That’s what a seed round was really for. Not the product — founders usually built that themselves. The round was for the non-product surface area. The COO, the CFO, the first head of marketing, the agency retainer, the lawyers, the recruiter. The company that had to be built around the product.
The new math is strange in a specific way. The product part hasn’t gotten much cheaper — building something real still takes focused attention. What’s collapsed is everything around the product. And that’s the part seed money used to pay for.
If the returns of intelligence are going to zero and the returns to intelligence are flowing to judgment, then the rational corporate structure for the era is one that maximizes the ratio of judgment deployed to overhead carried. That’s a holding company with a shared operating substrate. That’s why Grazier Ventures is a holding company and not a single startup. The structure isn’t a flex. It’s the shape of the math.
The question you should actually be asking
If you’re sitting on an idea right now and the thing stopping you from building it is an implicit calculation that you’d need a team, or a round, or six months of runway before you could even start — that calculation is running on 2020 assumptions.
The right question isn’t how much intelligence can I afford to throw at this? The marginal return of intelligence is collapsing; you can afford to throw quite a lot at it and it won’t be the thing that saves you.
The right question is do I have the judgment to direct intelligence at a problem worth solving? Because the returns tointelligence — the value that actually compounds — flow to that, and not to the tokens.
I’ve been writing for a year about what AI does to the capacity of a team. This is the version of the same argument aimed at a single person with an idea. The intelligence is cheap now. Your judgment isn’t. Spend accordingly.
I’m testing the thesis with four ventures. If you’re sitting on one of your own, the floor is lower than you think.
Let’s build.
This post connects to ideas from The Non-Product Surface Area, “Tokenmaxxing” Is Lines-of-Code Thinking for the Agentic Era, Minutes Added to Workforce, and The Builder’s Bet.
Project content
Personal & Professional Branding
Created by you