The Model Is Not the Moat — Context Is

Something quietly finished happening this month. Three frontier labs shipped new models within a fortnight of each other. The benchmarks landed inside a rounding error. The prices landed inside a rounding error. The demos, if you squint, landed inside a rounding error. For the first time in this cycle, the honest answer to "which model should we use?" is "any of them, and probably a different one next quarter."

That is a strange sentence to write in a market that has spent three years treating model choice as strategy. But look at where the conversations have moved. Nobody at a serious enterprise is arguing about parameter counts anymore. They are arguing about whether their semantic layer is honest, whether their evals catch regressions before customers do, and whether their agents can be trusted with the credit card. The model is table stakes. The layer around it is the whole game.

The uncomfortable read for most AI strategies written last year is this: if your plan was "pick the best model," your plan has already been overtaken by the market. The best model is now a commodity you rent by the token, and your competitor rents the same one.

A moat is something a competitor cannot copy in a weekend. "We use GPT-6" is not a moat, because so does everyone else with a credit card. "We use a fine-tuned model" is a shallower moat than it looks, because the base model underneath will be surpassed within six months and your fine-tune will have to be redone against a new checkpoint.

The leapfrog cycle is now short enough that any advantage tied to a specific model has a half-life measured in quarters. That is not a bug in the market. It is what a healthy foundation-model market looks like — the same way nobody's cloud strategy in 2015 was "we use faster EC2 instances than our competitor."

https://bricksnotes.com/context-advantage/book/chapter-24 is the chapter on Choice — the discipline of keeping model portability as an explicit design goal, so that when the frontier moves, your product moves with it instead of being anchored to yesterday's provider.

If the model is not the moat, what is? The honest list is short, and every item on it has one thing in common: it takes years to build and cannot be bought off the shelf.

Your semantic layer — the single, canonical set of business definitions the model resolves against. What "active customer" means at your company is not on the internet. It lives in the heads of your finance team, your legal team, and a Slack thread from 2023. Encoding that once, in one place, so every agent inherits it, is the argument of https://bricksnotes.com/context-advantage/book/chapter-6.

Your evals — the running record of what "good" looks like on your traffic, not on a public benchmark. Every regression caught before a customer sees it is a compounding advantage that a competitor cannot replicate by switching providers. https://bricksnotes.com/context-advantage/book/chapter-14 and https://bricksnotes.com/context-advantage/book/chapter-15 are the eval playbook, from golden sets to shadow evals on production.

Your retrieval — the parts of your business that the model can actually see, and how honestly they are represented. A model with perfect reasoning and no access to your ontology is a very expensive intern. https://bricksnotes.com/context-advantage/book/chapter-12 and https://bricksnotes.com/context-advantage/book/chapter-13 are about grounding retrieval in meaning, not just similarity.

Your guardrails — the harness that decides what an agent is allowed to touch, when, and with whose approval. Every high-blast-radius action wired through an explicit gate is a piece of trust you have designed in, not begged for. That is the whole subject of https://bricksnotes.com/context-advantage/book/chapter-18.

Your feedback loops — the mechanism by which the humans in the system correct the machine, and the machine remembers. Every correction that flows back into your evals, your resolvers, and your prompts is a data asset a competitor does not have. It is a small, boring, unglamorous flywheel, and it is the one that wins.

The book organises this argument around four pillars — Context, Control, Cost, Choice — because the moat is not one thing, it is a stack.

Context is meaning: the semantic layer, the ontology, the retrieval that reflects the business rather than the shape of your documents. Without it, every agent is guessing.

Control is trust: the guardrails, the approvals, the evals, the audit trail. Without it, capability is a liability.

Cost is durability: the discipline of measuring tokens, latency, and cache hits as first-class product metrics. Without it, your economics collapse the day the loop scales.

Choice is optionality: the architectural decisions — model abstraction, portable prompts, provider-agnostic tools — that let you swap the engine without rebuilding the car. Without it, you are one price hike away from a rewrite.

https://bricksnotes.com/context-advantage/book/chapter-4 is the short version of the 4 C's lens, and https://bricksnotes.com/context-advantage/book/chapter-3 is the definition of context we keep coming back to. Between them, they are the shortest path to a mental model for the next five years.

Consider two teams we have watched build customer-facing agents this year. Both started with the same base model. Both had roughly the same engineering budget. Twelve months in, one has a product that customers renew and refer, and the other has a product that customers screenshot when it breaks.

The difference had almost nothing to do with the model. Team A spent their first quarter building a resolver for their five core business entities, wiring an eval that ran on real (anonymised) customer traffic, and getting their guardrails to a place where the on-call engineer trusted the agent to run overnight. Team B spent their first quarter A/B-testing three model providers.

When the next frontier model shipped, Team A pointed their stack at it and their evals told them, within a day, that it was a net win — and where it regressed. Team B pointed their stack at it and had no idea. One team compounded, the other churned. The model was identical.

https://bricksnotes.com/context-advantage/book/chapter-27 is the reference architecture that Team A converged on, mostly by accident, and that Team B is now trying to retrofit under deadline pressure.

If we were advising a team that had accepted this argument and wanted to start on Monday, the shortlist would be five things — and none of them are "try a new model."

One: pick the five business entities your agents talk about most and encode them in one resolver, one meaning, one place. This is the smallest possible semantic layer, and it will pay back in a month. https://bricksnotes.com/context-advantage/book/chapter-6.

Two: stand up an eval harness that runs against a shadow of production. A hundred rows is fine to start. Grow it every week. https://bricksnotes.com/context-advantage/book/chapter-14.

Three: put every high-blast-radius action behind an explicit gate. Writes, sends, spends. Nothing else changes; you just now have a place to say no. https://bricksnotes.com/context-advantage/book/chapter-18.

Four: log every agent trace, and route the tail of the distribution — the strangest, longest, most expensive traces — to a human review queue once a week. That queue is your future dataset. https://bricksnotes.com/context-advantage/book/chapter-13.

Five: put a thin abstraction between your product and your model provider, so that switching providers is an afternoon, not a quarter. https://bricksnotes.com/context-advantage/book/chapter-24.

None of these require a bigger model. All of them compound.

Most "AI strategies" we have read this year are, on close inspection, model-shopping documents. They compare providers, they pick a winner, they set a budget, and they call it a strategy. That is a procurement plan, not a strategy.

A strategy is a bet on what will still be scarce in three years. Intelligence, in the raw form of a frontier model call, will not be. It will be cheaper, faster, and more accessible than it is today, and every one of your competitors will have it. What will still be scarce is a company that has spent those three years quietly building the layer around the model — the meanings, the evals, the guardrails, the feedback loops — that make intelligence useful in its specific business.

That layer is what The Context Advantage is a book about. If you have not started yet, the free chapters at https://bricksnotes.com/context-advantage/blog are a fair afternoon of reading. The full playbook, with the four moats, the reference architecture, and the eval and guardrail chapters, is at https://bricksnotes.com/context-advantage/buy.