Three years ago, an internal memo from a senior Google engineer leaked to the public. Its title: “We have no moat, and neither does OpenAI.” The argument was half right. The old moats are eroding faster than most leadership teams have priced in. New ones are forming, and they're harder to copy than the ones they replaced.
Companies are still investing millions in moats that are structurally weakening while ignoring the ones that are compounding. This isn't a prediction, it's already visible in the numbers.
Two-person teams now ship what used to take thirty engineers and a year of runway. The marginal cost of building software is collapsing toward zero, and with it the cost floor that protected incumbents for two decades. The gap between AI-native operators and companies still bolting AI onto 2015-era org charts widens every quarter.
What follows is a framework for distinguishing real moats from expensive head starts. A way to tell which is which, which moats are eroding, which are compounding, and which advantages most companies are mistaking for protection when they're really just expensive head starts.
What is an AI Moat?
An AI moat isn't a different category of moat. It's the same definition tested against a faster environment.
A moat is a structural advantage that:
- Compounds over time (not just a head start that erodes)
- Is hard to replicate even with full knowledge of how it works
- Increases in value as the competitive environment intensifies
A moat is not a feature, a tool, or a funding round. It's the reason the same feature, tool, or funding doesn't work as well for your competitor. What changed isn't the definition, it's the pace at which old advantages are failing the test.
The Old Playbook: Pre-AI Moats
Eleven advantages, scored on a 1, 10 scale before and after AI. Three tiers of erosion.
You needed massive server farms, global sales teams, and years of engineering to compete. Cloud + AI means a two-person team ships what used to require 30 engineers and $5M.
The cost floor that justified massive headcount has collapsed. Small AI-native teams now ship what previously required engineering organizations measured in dozens, with infrastructure that runs on commodity cloud.
The pain of leaving was the product. Migrating a CRM was a 12-month nightmare. AI makes extract-transform-rebuild near-trivial. Data isn't trapped anymore.
AI-assisted data extraction and rebuilding has compressed CRM and ERP migrations from twelve-month projects into multi-week ones. The pain that was the product is no longer the product.
Software patents bought 3 to 5 years of breathing room. Pace of innovation outstrips the patent cycle. Business method patents functionally dead post-Alice v. CLS Bank (2014).
In Alice Corp v. CLS Bank (June 2014), a unanimous Supreme Court ruled that abstract ideas don't become patentable just by being implemented on a computer. In the months that followed, business method software patents fell by roughly 60%, and Section 101 rejections at the USPTO more than doubled. The patent moat for software methodology stopped being credible.
Hiring all the best engineers was a real strategy. Google, Meta, Apple ran this for a decade. AI amplifies individual capability so dramatically that a small sharp team outperforms a bloated one. Headcount became a liability.
Meta's “Year of Efficiency,” announced by Mark Zuckerberg in early 2023, cut roughly 21,000 roles across two rounds. Operating margin moved from 25% to 35%; net income jumped 69% to $39.1 billion. The market read it as confirmation that headcount had become a liability rather than a moat, and Big Tech kept cutting in 2025 and 2026, this time explicitly to redirect spending toward AI.
Trust plus a built-in capability signal: only a big company could build this. Still creates trust and shortens sales cycles. The capability signal is gone.
Brand still buys you a meeting. It no longer buys you the category. For almost any enterprise software product, buyers now have a credible alternative built by a team they'd never heard of six months ago.
Owning the channel was how incumbents killed startups. Still matters, but distribution without differentiation is a treadmill.
When the product on the other end of the pipeline isn't meaningfully better, distribution stops compounding. It just feeds more leads into a process that loses on merit.
Having the biggest dataset was the ultimate unfair advantage. Foundation models now reason powerfully over smaller, well-structured datasets.
Adobe is the canonical case of a data-and-distribution moat being repriced in real time. Through 2025 and into 2026, multiple firms (Oppenheimer, Jefferies, BMO, Goldman Sachs, Citi, UBS, TD Cowen) cut their ratings, with consensus dropping to its most bearish level since 2013. The cited reason: AI-native creative tools are eroding pricing power on workflows Adobe owned for two decades.
Raising $100M used to mean you could outbuild and outspend. The cost floor for building competitive software collapsed. Now $100M mostly buys you a faster burn rate.
The math of large rounds inverted. When the marginal cost of building competitive software was high, $100M bought you years of runway and a real lead. When it's collapsing, a lean AI-native competitor with $2M can ship in parallel.
Hardest classic moat to crack because it's structural, not technical. AI-powered cold-start solutions are beginning to chip at the bootstrap problem, but slowly.
Airbnb, LinkedIn, and most marketplaces still hold. The structural problem of bootstrapping the other side of a two-sided market hasn't gone away. AI helps close cold-start problems on the demand side, but it doesn't manufacture supply.
Lobbying power, compliance regimes, audit cycles. AI doesn't change lobbying power. Possibly strengthening as AI-specific regulation increases.
Banking and healthcare compliance regimes still take years to navigate. AI-native challengers face the same incumbents, the same audits, and the same lobbyists. As AI-specific regulation hardens, this moat may even widen.
Factories, supply chains, logistics networks. AI optimizes operations but can't conjure a chip fab.
TSMC's fabs and Amazon's logistics network don't get conjured by generative anything. AI optimizes the operations on top, it doesn't build the plant.
The New Playbook: Post-AI Moats
The moat didn't disappear, it shifted. There are arguably more defensible positions now. They just moved from things you can buy to things you have to earn. Three tiers of defensibility.
AI lets you encode decision-making judgment into software that acts, not just documents that describe. The moat isn't the tool, it's the opinionated workflow. A competitor can copy your UI. They can't copy the 200 decisions baked into why step 3 comes before step 4.
Nvidia is the textbook case at scale. CUDA, launched in 2006, isn't just a programming interface for GPUs, it's nearly two decades of encoded decisions about how parallel computing should work, embedded in roughly 5 million developers, hundreds of optimized libraries, and 40,000+ companies. Competitors with comparable hardware lose to the methodology that runs on top of it.
Every interaction teaches the system something, AI reasons over it, and the next interaction improves. Not “we have more data,” but “our data makes us smarter with every use.” Competitors start at zero. The gap widens with usage, not with funding.
The pre-AI version was Spotify's Discover Weekly: usage data improving recommendations. The post-AI version is sharper. A vertical SaaS that watches how its customers actually run their operations builds a model of that domain that no funded competitor can replicate from a cold start.
The specific way you chain AI tools, human decision points, and feedback loops into a repeatable system. Invisible to competitors. Not legally protectable (Alice killed business method patents) but practically obscure. By the time someone reverse-engineers V12, you're on V20.
The way a team orchestrates AI across proposal generation, client onboarding, and ongoing reporting isn't a tool, it's compressed experience. A competitor could be handed a flowchart of every step and still not understand why step 4 had to come after step 7. Process intelligence is the moat that exists because it can't fit in a screenshot.
AI makes execution nearly free. Knowing what to build, what to cut, and when to ship is the actual bottleneck. Tools are available to everyone, the workflows aren't.
Apple's sustained design moat is the pre-AI version. The post-AI version is more democratic and more brutal: two teams now have access to identical models, identical APIs, and roughly identical execution capability. The output diverges entirely on judgment.
Not speed of building V1, everyone's fast at V1 now. Speed of learning from real usage. The team shipping V4 while competitors finish V1 has three versions of user feedback competitors haven't seen.
V1 is no one's moat anymore. The compounding advantage lives in V2, V3, V4, the architectural decisions only a team with real usage data can make. By the time a competitor ships their V1, you're three versions in.
AI gives everyone generalist capability overnight. Deep domain knowledge is what you can't prompt your way into. Generalist capability is now a commodity. Specialist judgment isn't.
The difference between a generic AI tool that demos well and one that actually works for healthcare scheduling, legal discovery, or agency operations comes down to domain knowledge that took years to accumulate, and you can't prompt your way into it.
The tiebreaker when two competitors can build the same product in a week. On a level playing field, the human layer appreciates rather than depreciates.
When two vendors can ship the same product in a week, trust becomes the tiebreaker. Reputation, relationships, the fact that someone returns your call, those things stop being nice-to-haves and start being load-bearing.
Being woven into a client's daily workflow. Not just stickiness, the system accumulates context about that specific client. A replacement starts dumber. The moat isn't “it hurts to leave,” it's “the replacement starts dumber.”
This is different from old-school switching costs, which were about pain. Embedded operations are about context, a system that has watched a client run their business for six months knows things a replacement would have to relearn from scratch.
Plugin builders, integration developers, power users who extend your platform. AI accelerates ecosystem contributions, but only for the platforms that earn the contribution in the first place.
Figma's plugin community, Notion's template ecosystem, WordPress's plugin economy. AI lowers the cost for outside developers to extend a platform, which means the ecosystems compound faster.
In a world drowning in AI-generated content, the ability to filter signal from noise becomes a product. If they recommend it, it's worth my time, in a world where everyone's search results are SEO-optimized noise.
Wirecutter at its peak was the model. In a world where generative content is functionally infinite, vertical curators who earn that trust again, in a specific domain, with skin in the game, will own attention that AI can't commodify.
Not whether you use AI, but whether your org's decision-making, hiring, and incentive structures assume AI as a given. Most companies are bolting AI onto 2015-era org charts. The throughput gap is structural.
There's a difference between a company that “uses AI tools” and one that has rewritten its hiring rubric, its meeting cadence, and its definition of “done” around AI as a given. The second is building from a different starting assumption, and it shows in throughput.
How fast you go from raw signal to an informed decision. Different from iteration speed. AI compresses the analysis layer, the bottleneck moves to organizational willingness to act.
AI compresses the analysis layer to near-zero. The companies that read the same signal as everyone else but act on day one rather than day thirty are converting AI capability into actual advantage.
How To Audit Your Own Moat
Three questions worth running through honestly:
What are you currently defending? Look at where your budget goes. If it's going toward scale, headcount, data accumulation, or switching costs, you're investing in eroding moats.
What would break if a competitor with AI built your product in 30 days? If the answer is “nothing, we'd still win because of X,” X is your real moat. If the answer is “we'd be in trouble,” you don't have one.
Where is your knowledge that competitors can't see? The encoded decisions, the process intelligence, the domain expertise baked into how you operate. If it's all in people's heads and not in your systems, it's a vulnerability, not a moat.
Where's your moat?
8 questions. Two minutes. Tells you what you're defending and what you're missing.
The Bridge: Cultural Arbitrage
The companies still investing in pre-AI moats are spending money to defend positions that are structurally weakening. The companies building post-AI moats are investing in things that compound. That gap between what's available and what's adopted, the cultural arbitrage, is where the competitive advantage lives right now.
The moats shifted. Most companies will spend the next year defending the wrong castle. The ones that won't are the ones encoding their methodology into systems that compound, not adding more headcount, not adding more features, not raising another round.
Most companies won't know which castle they're defending until it's too late to stop.
Want us to map your moat?
One conversation. No deck. We'll tell you which positions are still earning their cost and where the build hasn't started.