The five moats AI can't reach

AI is making information, content and the ability to sound smart almost free. So the moats that used to protect a stock are being repriced. This framework asks one question of every company we cover: is its advantage something AI can learn, bypass or buy? If it is, it is no longer a moat.

What is no longer a moat

Information asymmetry. Brokers, media and experts who lived off information gaps are losing their edge. AI has made information effectively free.
A pure software or algorithm lead. A smart team's edge in code can now be copied within months. 'We have great tech' is not a moat on its own.
Content production. Text, images and video can be generated at scale. The ability to produce content is no longer worth anything by itself.

Rule: if AI can learn it, it is no longer a moat.

The five layers that have become scarcer

Layer 1: a proprietary data flywheel

AI has driven down the value of public data. What is scarce is data only you can keep collecting, and that compounds with use: user behaviour generates data, data improves the service, and a better service attracts more users. The crucial test: as AI gets stronger, this moat should become more valuable, not less, because whoever owns the unique data can extract more from it with AI. Example: Hims & Hers (multi-product cross-selling and a closed loop of health data, with high switching costs).

Layer 2: regulation and licences

AI can copy capability. It cannot copy a banking charter, an FDA approval or payment-rail compliance. In the AI era, regulatory moats gain value. Examples: Nubank (a US OCC charter), DLocal (an emerging-market payments licensing network), Circle (stablecoin compliance after the GENIUS Act).

Layer 3: physical assets and networks

AI lives in bits. It cannot move atoms. Energy, the grid, data centres, rare minerals and logistics nodes become bottlenecks in the race for AI compute. Example: data-centre landlords collecting picks-and-shovels rent from the AI wave.

Layer 4: trust, brand and habit

When anything can be faked, a brand people trust without a second thought becomes a luxury. Coca-Cola and Costco hold decades of trust and habit that AI cannot steal. This is the classic Lynch and Buffett moat, and the most underrated one in an AI hype cycle.

Layer 5: owner-operator capital allocation

The scarcest layer of all: a founder who genuinely thinks long term, allocates capital at an elite level and is deeply aligned with shareholders. AI can write code. It cannot make high-stakes, bet-the-company decisions. Examples: David Vélez at Nubank (75% voting power), Elon Musk, Andrew Dudum at Hims & Hers.

The verdict

Scarcity has shifted from capability to credentials, assets and trust.

Anything AI can learn is no longer a moat. Only what AI cannot learn, cannot bypass and cannot buy is a moat.

The companies to look for combine all three:

a closed loop of private data, which AI helps monetise rather than disrupt;
a regulatory, licence or physical-asset barrier AI cannot prise open;
a top-tier owner-operator.

The deepest moats of the AI era sit in fintech and healthcare platforms that hit all three at once.

Applied to three holdings (assessment as of 18 June 2026)

Tesla: strong on physical assets and the owner-operator, but the data flywheel is contested

Physical assets and network: strong. The Supercharger network, global factories, the battery supply chain.
Owner-operator: strong, and double-edged. Musk is a top-tier capital allocator and also the biggest single uncertainty (divided attention, political risk).
Data flywheel: disputed. FSD's real-world driving data is a unique closed loop in theory, but monetisation has been pushed back repeatedly. A data moat that does not turn into profit remains a story, not a verified moat.
Brand: moderate. Strong, but being eroded by politicisation and Chinese competition.
Regulation: mostly a headwind, not a moat.

Hims & Hers: strong on data and the owner-operator, with licences and assets stronger than first assessed

Data flywheel: core and real. Multi-product cross-selling and closed-loop health data. This is the foundation of the long thesis.
Owner-operator: moderate to strong. Dudum has shown he can execute and expand into new categories quickly.
Regulation: mixed, tilting towards a moat. On GLP-1s, regulation is more threat than shield. But the company owns hard-to-replicate facilities regulated by state pharmacy boards: a 352,012 sq ft campus in New Albany, Ohio (R&D, manufacturing and dispensing), peptide and compounding capacity in California, and GLP-1 fill capacity.
Physical assets: meaningful. Beyond the above, Hims Labs is a diagnostic lab the company owns. Hims & Hers is not a pure asset-light platform. By integrating vertically, it is converting GLP-1 regulatory risk into regulated capacity under its own control.
Brand: moderate. Strong direct-to-consumer awareness, but not yet Coca-Cola's depth of trust.

SoFi: a moderate moat, not a buy-and-forget wide moat

Banking licence: the hardest layer. The 2022 US bank charter lets SoFi take deposits (lower cost of funds) and hold loans on its balance sheet (earning net interest margin rather than one-off fees).
Technology stack (Galileo and Technisys): moderate. Payment and card APIs, plus a core banking system sold to other fintechs. A genuine second growth curve, though customer churn is real.
Cross-sell flywheel: building. The one-stop financial app is working, but not yet self-reinforcing.

The reality check: financial products are commoditised and switching costs are low (users will move for 50 basis points), so SoFi's moat is efficiency and scale, not monopoly. If growth slows, it thins fast. AI also lowers the bar for building a good financial app, so any edge in user experience will be matched. The defensible layers are licence, cost of funds and scale. Watch loan default rates, member growth and net interest margin.

SoFi and Nubank are the same species in different markets: SoFi fights for share in a saturated US market, while Nubank expands in Latin America, where many people were never well served by banks. That is the single biggest reason their moats differ in depth.

The discipline that keeps this honest

The biggest failure mode is scoring a holding well because you already own it. The framework exists to find holes in a thesis, not to validate it.

For each layer, ask first: where is this claim inflated?
If a moat depends on something that has not yet turned into profit, call it a story moat, not a verified moat.
If a fact behind a score turns out to be wrong, rescore openly, up or down.
The question for new capital is never 'which holding do I add to?' It is 'which scarcity am I missing?'

The scarcest thing of all

When AI makes information and the ability to sound smart infinitely cheap, the rarest thing left is the refusal to let the tool think for you: patience with slow, hard work (reading a full 10-K, holding for years), first-hand judgement from primary data, willingness to hear that you are wrong, and real human trust. If forced to pick one: the willingness to pay the price for long, hard, slow things when everything else is instant. Value investing pays precisely because most people lack it.