The Operator's Supercycle

What I’m seeing in the work, and why I think the next decade rewards a different kind of founder.

The hardest parts of building Leverage AI took eight years and a real team and they still do. The customer relationships, the regulated workflows, the data, the institutional trust. None of that got cheaper.

What changed is everything around them.

The build cost has collapsed. Things that used to take ten engineers a year now take two engineers a couple of months. The reach you can extend out from a real moat got an order of magnitude cheaper and yet the moat itself did not.

In my opinion, that asymmetry is the actual structure of this AI cycle and it changes who is positioned to win.

The pair to the moats argument

I wrote earlier this month about where defensible value is moving in the AI era. The TLDR: as code gets cheap, defensibility shifts toward money movement, compliance, and infrastructure. The new moats are the things that do not compress with better code generation.

The previous post answered the structural question. This post answers the operator question that follows.

If durable value lives in regulated workflows, institutional trust, and infrastructure embed, then the people best positioned to capture it are not the ones who can build the smartest model. They are the ones who can credibly deploy into real industries, who already understand the moat layers, and who can now extend their reach 10x faster than they ever could before.

That is a different kind of founder than the last cycle rewarded. And it is a once-per-career window.

What twenty years taught me to look for

I have spent the last twenty years cycling between building, operating, and investing across aerospace, defense, fintech, consumer, and enterprise software. From the Pentagon, to a White House innovation program, to founding companies, to Bank of America as an EIR, to running a national COVID response, to building Leverage, to Inauguration Capital. The throughline was never one industry. It was always being a generalist who shipped.

Most of the time, the honest answer to “is this a step change” is no. Most cycles are incremental. The internet was a step change. Mobile was a step change. Cloud was a real shift but more gradual than its proponents claimed. Crypto, in retrospect, mostly was not. AI in 2018, 2020, 2022 was not yet, despite the noise.

What changed from my perspective in the last twelve months was not a new model release, but instead watching what small teams with the right expertise can rapidly ship into legacy industries.

What the build cost actually looks like now

The collapse is uneven, which is the punchline.

The periphery, which is to say everything around the moat, is dramatically cheaper to build than it was 18 months ago. Internal tools. Workflow automation. Customer-facing UI. Integration glue. Reporting layers. Data pipelines for non-regulated data. The work that used to require a full delivery team can now be executed by a team of two and strong AI tools.

The moat layers did not get cheaper. The licenses, audits, customer relationships, regulated data, institutional trust, and bilateral integrations that gate real industries still take years and AI does not shorten any of them.

The result is a strange shape. The companies that already have the moat can now extend their reach at one-tenth the cost. The companies that do not have the moat can build a beautiful periphery in a weekend and still cannot enter the market. That gap is widening, not closing.

Build cost compressed unevenly. The cheapness shows up where it does not gate the market. The expense remains where it does.

The forward-deployed shift

The other thing I am watching change is how engineering itself is structured around real customer problems.

For most of the SaaS era, the product was generic. You built a feature set, and customers configured it for their workflow. The smaller the customer, the more they accepted out-of-the-box. The larger the customer, the more they hired implementation partners.

That model is breaking, and what is replacing it looks more like precision medicine than precision software. Engineers are being deployed forward, embedded inside the customer’s actual workflow, and shipping software that fits one institution’s specific shape rather than a generic one. The unlock is that AI tools collapsed the cost of customization to the point where bespoke is now economic for the seller, not just the buyer.

This matters because forward-deployed engineering compounds completely differently than seat-based SaaS. You are not selling a license, but instead embedding into an operating reality. Every deployment makes the next deployment faster, because the patterns transfer. Every customer strengthens the moat.

The operators who already understand a regulated industry, who already have customer relationships, and who can now staff a forward-deployed engineering team for the cost of one senior hire are positioned to do something that was not economic five years ago.

Why generalist operators get the asymmetric trade

The conventional read is that AI rewards depth. The deeper your domain expertise, the better your prompts, the better your fine-tuning, the better your product. There is some truth to this.

The less obvious read is that AI also rewards breadth, and rewards it more than the previous cycle did.

The reason is execution speed. When the build cost on the periphery collapses, the constraint shifts to who can credibly identify which industries are exposed, who has the relationships to land the first customer, who can navigate the regulatory layer, and who can then ship into it before the incumbent absorbs AI into their own stack. Those are not deep-specialist skills. They are operator skills, accumulated across multiple domains.

A generalist operator with years across regulated industries, public-sector deployment, financial services, and frontier hardware now has an asymmetric advantage that did not previously exist. The cost of acting on a pattern they recognize just dropped tenfold.

The arbitrage is real, and the window for it is open right now.

Where I could be wrong

Two scenarios would weaken this thesis materially.

First, if foundation model providers absorb the forward-deployed layer faster than expected. If the next generation of models can credibly customize themselves to a specific institution’s workflow without bespoke engineering, the operator advantage compresses. I do not think that is the base case in the next three years, because customization in regulated industries is bottlenecked by audit, governance, and bilateral trust rather than by model capability. But the probability is not zero.

Second, if the moat layers themselves erode faster than the new reach can compound. If regulators move faster, if institutional trust transfers more easily across providers, or if compliance becomes a commodity layer in the next three to five years, the asymmetry I am betting on weakens. I hold this thesis with awareness of the conditions that would break it.

The diagnostic

The moats post asked what a great engineer with Cursor and a weekend cannot replicate. This post further asks:

What can a great operator with the right expertise build in 12 months that the incumbent thinks is 5 years out?

If the answer is nothing, the industry is not ripe yet. If the answer is something, that is where the asymmetric returns of this cycle are going to be made.

Most of the interesting work in the next decade is figuring out which industries fall into the second bucket. That is the work that excites me.