The closer you get to the AI frontier, the more you may start wondering if there are any moats left. The pace and marvel of progress may make you believe that the top AI labs will soon eat software. And after software, AI labs will eat the world, as most work can be expressed as software. Or at least some believe it can.
But you will also quickly realize that competition from open-weight models is not trailing far behind, and has a good chance of democratizing the bulk of token volume, even if not the top of the market. While there will always be demand for SOTA, it might be fleeting enough and concentrated in domains where the premium truly gives you a massive advantage over everyone else. Think domains where slightly better intelligence translates quickly into wins, from cyber to patentable matter. Luckily, many fields still encounter friction with the real world, and that friction may be more significant, and have more economic leverage, than a few months’ AI lead.
If you believe closed SOTA and open-weights will co-exist, you’ll inevitably start thinking about routing as the obvious aggregation and value capture layer, only to realize that barriers to entry in routing are modest at best, and that every vertically specialized app, from coding to finance, will bring its own. And if all you are left selling is cheaper inference, you can’t ignore the competition from the leading labs: all they need to do is release an extremely low-cost model that matches the open alternatives to make routing not worth it (at least if all your customers care about is costs).
Once you move beyond routing, you’re inevitably narrowing on the app layer. And focusing on this level of the stack may even give you a temporary reprieve from the moatpocalypse whiplash. Plenty of historical examples support the idea that the last mile is where all the friction usually hides, and therefore a great place to plan for value capture once things stabilize. This was true for other general purpose technologies such as crypto, despite endless discussions about fat versus thin protocols capturing the rents from open networks. The irony in that case was that while the technology started as subversive, value accrued to the most boring intermediaries: the ones with a banking charter.
But the app layer isn’t perfectly shielded from AI either. Often, what’s standing between your customers switching en masse to a series of prompts or a new tool is just more and better digital traces feeding the labs’ training. A model that is ok at design today, given the right data, can get exceptionally good at solving most of your customers’ needs. It’s the fruit of the Bitter Lesson, eating one industry vertical at a time. From design to legal, finance and accounting, it is unquestionable that each major model release further erodes the need for specialized tools for a meaningful share of users. Jevons’ paradox may come to your rescue, or not.
With more capable models, startups and incumbents alike can do more with less, and SaaS tools that used to be procured are now cloned in-house at a fraction of the cost. Of course, better models cut both ways, as the best SaaS providers can expand their offering and out-iterate internal teams by observing evolving needs across a wider customer base. They can also use this historical opportunity to switch from selling seats to selling the end-to-end work—assuming they are willing to underwrite the risks and become liability-as-a-service (LaaS) providers.
In many cases, you’ll have a mix of SaaS contracts being cancelled when the internal version recreates the 10% of specific features a company actually needs, and larger contracts being signed where SaaS has become robust LaaS. Just look at how rapidly Stripe and Ramp have turned from fintechs to foundation labs focused on finance and payments: there will be no SaaSpocalypse for the firms that leverage the tools to capture more of the value chain.
Ultimately, we have to accept that turbulent phases of technological change are both extremely exciting and confusing. Whenever you feel like you’ve found something to anchor your beliefs on, you are quickly humbled and reminded by the latest breakthroughs of just how shaky our current assumptions are.
Ironically, AI is exacerbating consensus and muting contrarians because it is accelerating idea recombination: everyone is (inadvertently) distilling everyone else in real time. In 1890, Alfred Marshall defined industrial clusters as places where “the mysteries of the trade become no mysteries; but are as it were in the air”. For AI, that place today is San Francisco. As robotics advances, it may well be one of China’s river deltas.
Regardless, much of the debate around the surviving moats is inevitably online, on X. Some of it is obvious and priced in, and some of it is plainly wrong. So rather than giving you a list of moats, this piece will propose a few economic first principles. By running them through your own accumulated experience, you’ll be able to map them to what is and isn’t a moat in your specific domain.
Economic First Principles for a Messy AI World
About a year ago, with Jane Wu and Kevin Zhang, we wrote a piece that delivered a very simple insight: anything that can be measured will be automated.
As more and better data comes online and is fed to models, they will be able to replicate more of human thinking and behavior. Because better AI also leads to better measurement, the dividing line between what stays human and what companies can defend is a fast-moving frontier.
Of course, some domains and problems are harder to codify than others: this is the land of unknown unknowns where assigning probabilities to events is extremely difficult, if not impossible. Financial markets, frontier R&D, and domains where the right answer doesn’t matter because what we care about is that others converge with us on the same result: think “status games”, and any good or service where social consensus is the product, from art to cryptocurrency.
If you buy our thesis, the resulting first principle is basic: you can either compete on advancing measurement, or you can double down on domains that escape measurement to begin with.
The first is one of the most thriving areas of work right now, as everyone has realized that you cannot train more capable models without capturing more of what shapes the actual workflows that matter in the economy.
The second is the land of the “meaning makers”, and is one where human connection may matter. Although often people overestimate just how much a lower-cost service may become a substitute for a high human-touch endeavor if the results are comparable, from mental health support to forms of care.
In a follow-on paper with Jane Wu and Xiang Hui on the economics of AGI, we went one step further. If you believe measurement is the key to understanding where this is all going, then you quickly realize that agentic labor is useless whenever the agents are missing key measurement that we, as humans, are using every day in performing the job. We bring that missing measurement in through our experience whenever we perform verification, which is essentially deciding if a model output is ready to be used in production.
Verification is the ultimate bottleneck to AI scaling. It is the human act of bringing in tacit knowledge that the machines haven’t seen yet. It is also something fundamentally different than having AI verify AI. It is, by definition, the residual of that process. The part of the code review that you cannot automate away without running into issues later. The part of the software factory that still needs to ping a human.
Which leads to a second principle: if execution becomes free and abundant for anything that can be measured, value accrues to the verification last mile.
How do you build a moat in the verification last mile? Fundamentally, there are only two ways: you either collect unique, proprietary data following the first principle, or you hire and retain some of the world’s best experts in the relevant core domains. Of course, building better verification tooling can also be very profitable in the short term—and is the reason why Cursor was worth so much to SpaceX—but ultimately tooling has to evolve into a data play to be sustainable.
How do you acquire better data? A range of options here, from sensors, to digitizing additional workflows or interactions with the real world, to looping in and capturing the insights from talented humans. Crucially, the “weights” in the world’s top expert brains are some of the most realistic simulators of a particular slice of reality, and the market will price that accordingly.
Once you view the evolution of AI through the lenses of what’s measurable vs. not, and what needs verification vs. not, you cannot unsee that many network effects that have made digital platforms so resilient over the last decades are weaker than they appear.
Yes, distribution is a moat, especially in domains where you may need a license to operate in the first place, but not all distribution is equal. In many cases, AI agents will be able to help consumers and businesses multi-home, switch and migrate to new providers, and constantly hunt for better deals. They will also be able to aggregate demand across marketplaces, undoing some of the market liquidity benefits of the leading venues.
While most network effects lose their value as models become more capable, there is one type that, if anything, becomes stronger. We called it a verification-grade network effect, and it has a very particular shape: it is a network effect that gets stronger as you process more transactions and are able to observe more out-of-distribution events. Why? Because you can feed those back to your experts, have them perform verification, and then use the resulting digital trails to automate things further.
Summarized as a third principle: scaling verification faster than your competitors establishes verification-grade network effects. These allow you to build better and more capable models, widening your lead. We’ve seen early glimpses of this with Tesla’s Full Self-Driving, Stripe’s AI Payment Model, and Palantir’s AIP, as well as through the advancement of tools like Claude Code, Codex, Cursor, and Devin.
“Any problem in computer science can be solved with another level of indirection.”
How do you survive the moatpocalypse? It’s simple.
You obsess about what is currently measured vs. not. You use what is measured to automate your execution, and what is not measured as an opportunity to build new, unique, proprietary data in the domains that matter to your workflows. To do so, you inevitably need to hire the best global talent to perform verification, and set up systems that can feed back their decisions, evals and experience into the next training run. That loop is how you establish and defend the only type of network effect left, the verification-grade one.
Of course, compute and the physical infrastructure running all of this are also extremely valuable, and are obvious bottlenecks as our insatiable appetite for inference scales. Same with regulatory friction and licenses in heavily regulated industries. But if you zoom out, the main reason why those two are bottlenecks to begin with relates to how you make progress in those domains. You make it, respectively, through friction with atoms and physics, or with regulators and policymakers.
So in the end, if you commit the common sin of adding a layer of indirection, the only principle you should care about is maximizing your path towards meaningful friction with the real world. This is the actual reason why many are obsessing about “agency”.
In a world where intelligence is cheap, agency is a proxy for individuals and companies that are willing to drive through the grunt work of facing reality and collecting better measurement as fast as possible.
Great piece. Two quick thoughts:
First, Veeva's tacit knowledge stops being tacit once digitized — other models will distill it.
Second, you tell everyone how to survive individually, but what happens when everyone follows the same playbook?