Why the Base Model Isn’t the Moat.
PH—007
Every competitor can buy the same foundation model tomorrow. The moat is what sits on top. This is the four-layer architecture we install, and why it compounds in calendar time rather than engineering time.
I The Wrong Mental Model
The prevailing assumption, still, is that the base model is the product. It is not. The base model is a raw material. Anthropic sells access to it. OpenAI sells access to it. Mistral and Google and a dozen others sell access to it. The price is dropping every quarter. The capability is rising every quarter. It is the least defensible part of any AI product.
The moat has to be somewhere else. Four somewhere-elses, specifically.
II Layer One · The 2026 Corpus
The first layer is training data that did not exist two years ago. The 2026 corpus is assembled from enterprise AI shipping patterns. What actually made it to production in 2025 and 2026. Not blog posts about what might. Not benchmarks about synthetic problems. The signal that matters is which architectures survived their first quarter in a real environment, and which did not.
A competitor could assemble this. It would take them a year and they would have a 2025 corpus, not a 2026 one. By the time they were done, the 2027 corpus would exist and they would be behind again. The corpus is not a dataset. It is a research cadence.
III Layer Two · Quarterly Refresh
The second layer is the refresh cadence itself. The corpus retrains every ninety days. The model that ships to customers this quarter knows things the model from last quarter did not. The delta between quarters is public in the quarterly brain update.
A competitor catches up by refreshing faster. This is a calendar-time problem. You cannot compress calendar time. Every quarter we ship a refresh, the window they have to catch up in closes by the length of a quarter. If they ship at our cadence, they stay where they were. If they ship slower, they fall behind.
IV Layer Three · The Compounding Install Loop
The third layer is the one most visible in the public case studies. Every install teaches the methodology. We had no ontology for home services at install one. We had a rough draft at install three. At install eleven, the filter was catching recommendations a year-one installer would not have thought to check.
A competitor starts at zero installs. They cannot skip ahead. Consulting methodologies get better with casebooks. Casebooks are written after the fact. The compounding loop writes itself during the work.
V Layer Four · The Per-Industry Ontology
The fourth layer is the one we argue about the most, internally. It is the most expensive to maintain. It is also the most defensible. Each of seven per-industry ontologies encodes regulatory, economic, and vocabulary constraints that a generalist base model does not know. The ontology is a translator layer between what the model can do and what the industry allows.
A competitor can hire subject-matter experts. They cannot skip the step where those experts watch the model fail a thousand times and decide which failures to encode. That is what coding an ontology means.
VI What This Means for Buyers
The implication is uncomfortable for the buyer who thinks they are choosing between AI vendors. They are not. They are choosing between vendors who build on a commodity base (some well, some badly). The base model is not the distinguishing feature. What the vendor has put on top is.
Ask a vendor which base model they run on. If the answer is a long defense of a single model, they have no moat above it. Ask them what happens when the base model gets swapped. If the answer is “the product gets better on its own,” they have a real moat. That is the test.
VII Why the Layers Compound
None of the four layers is original. Corpora exist. Refresh cadences exist. Install methodologies exist. Ontologies exist. The pattern is that they compound in parallel. The corpus feeds the refresh. The refresh informs the install. The install teaches the ontology. The ontology improves the corpus. Four feedback loops, each on a different time horizon.
This is what is meant by a moat that widens with calendar time. Not one advantage. Four advantages whose outputs are each other’s inputs.
The next vendor through the market does not race us on any one of them. They race us on all four, simultaneously, starting from zero. That is a different kind of race.