How does an AI business stay defensible when every team calls the same frontier model and the three most circulated moat substitutes — data, domain, and distribution — each crack under production scrutiny?
A reportedly ten million dollars spent training a finance-specific model. Beaten, without specialized training, by a general frontier model. That is the Bloomberg anecdote circulating across AI strategy decks, and it is the wrong thing to focus on. The harder question it opens: if the model is now a commodity, what exactly survives?
The model is no longer the moat. Eighteen months ago that was a contrarian take; today it opens every AI strategy deck. The replacement framework has not kept pace. Data loops, encoded domain expertise, distribution: these get recited as surviving pillars the way mortgage interest gets listed on a closing disclosure. Present, but rarely examined. The source laid out the most-circulated version in a recent Forbes Tech Council contribution. It is a useful frame. It is also undertested in production. Each leg has a stress fracture that wire coverage does not catch.
Most claimed data moats are lakes. A static data lake is just accumulated history; it does not compound from new work. The argument is that a human-in-the-loop review step (a clinician correcting a draft, a claims adjuster redlining a summary, a customer tagging a wrong answer) produces labeled data on every cycle competitors cannot replicate without replicating the product surface. The fracture: most enterprises build the loop and cannot staff it. A loop with three corrections a month is not a moat; it is a slow leak. Founders' decks claiming "proprietary feedback data" should be asked how many labeled corrections per day the product actually produces, not how many rows sit in the warehouse.
Most "we understand this industry" claims are starting lines, not leads. Frontier large language models ate the engineering half of vertical software, so domain expertise stopped being a moat and became a starting condition. The advantage now has to live in edge cases, compliance rules and outcome definitions the model cannot infer from general training. That part is right. The fracture: most compliance encoding is narrower than it sounds. The example cited (a 17-year-old cannot legally opt in to marketing under the Telephone Consumer Protection Act) is a structural rule a general agent reads from the statute. Harder cases exist (multi-state licensing, clinical decision support, jurisdictional product classification), but they are narrow. Enterprise buyers should audit the specific list of edge cases the product handles that a careful prompt cannot reproduce. If the list cannot be produced, the moat is the slide deck.
API access is not distribution. Distribution in the agent era means placement where the work already happens: inside the customer-relationship-management system, the clinical workflow, the underwriting queue. Owning the seat is owning the route to the work; that part is correct. The fracture: placement without a closed loop just means the company is annotating someone else's model. If the workflow improvement does not return signal to the product, the distributor becomes a free data source for the next model jump. Founders selling distribution as a moat need a named mechanism, contract, architecture or trust posture, that prevents feedback extraction. Without it, they are selling real estate, not defensibility.
Andreessen reads the same shift differently. He has argued that investors are still buying infrastructure and orchestration as the model layer commoditizes. Benedict Evans, in conversation, frames it as investors buying growth rather than moats while the model layer races. Both views need reconciling with the three legs; neither cancels the question the piece forces.
AltaDX positions itself as a productized AI outsourcer for fitness, wellness and beauty enterprises. By the framework, those clients would have to staff the human-in-the-loop step themselves to capture the data moat the vendor is promising them. That tension, between a framework a vendor sells and the operating reality of its customers, is the part of the consensus that has not been examined.
The replacement for a model moat is not a list of three virtues. It is a tighter, falsifiable claim. A loop that produces labeled corrections at production volume, beating a competitor with a bigger static dataset. A domain encoding that names the specific edge cases the model cannot learn from the statute book. A distribution channel with a named mechanism that prevents feedback extraction. Each is auditable. Each has a counterexample waiting. The audit question for any founder or buyer operating on this framework: can you prove each of these against a well-resourced competitor's next model jump?