Why the model can't be the advantage

Every company now says it runs on AI. The claim costs nothing to make, which is why everyone makes it. If you sit on a board, allocate capital, or choose partners, you need a way to tell the firms that built something real from the ones that added a chatbot and updated the website.

Start with what the claim cannot mean. It cannot mean the model, because almost nobody owns one. The models are rented from a handful of labs, they improve on the labs' schedule, and competing firms can often rent the same one on the same afternoon. Prompts can be copied; so can interfaces. Whatever a firm's durable advantage in AI turns out to be, it has to survive the next model release — and a moat you replace every twelve months is not a moat. A company whose edge disappears when a model provider changes its pricing, terms, or technology has built on rented ground.

The model-swap test

Satya Nadella gave the cleanest version of this argument in a note published in June 2026. A firm that uses AI seriously, he argued, is building a learning loop on top of the models: its own accumulating record of decisions, corrections and outcomes, feeding back into how the work gets done. In his words: "This loop becomes the new IP of the firm. I think of it as a hill climbing machine. And unlike most assets, it compounds."

A sharp test follows. Imagine swapping the model underneath the business for a rival's. Does the firm keep its accumulated expertise, the way it would keep a twenty-year company veteran? The model is not the memory of the firm; it is an engine applied to that memory. A firm should be able to install a better engine without forgetting the road it has travelled. If the intelligence lives in the model, the firm never really owned it; if it lives in the loop, the engine changes and the memory remains.

That is the model-swap test. It deserves two companions.

Two companion questions

The second question comes from the venture firm CRV, in its founder's guide to AI-native companies: remove the AI entirely. Does the business merely degrade, or does it stop working? A bolted-on feature degrades gracefully, because the real work never depended on it. A firm rebuilt around the technology cannot go back, any more than a logistics company could go back to paper maps.

The third is about measurement. MBI Deep Dives' "Owning the Hill" observes that firms with private evaluations, tests drawn from their own work, can beat frontier models on their own narrow domain, and that without such evaluations a firm cannot tell whether the compute it pays for bought it anything. So ask: who keeps the scoreboard, and with whose questions? A firm grading itself on a public benchmark is measuring the lab's progress, not its own.

What the loop actually is

What is this loop, concretely? Nothing exotic. Work enters the system. People and machines make a decision. The firm records the decision and its reasoning. Later evidence tests whether the decision was sound, and the answer changes how the next case is handled. A folder full of documents is not a loop. Neither is a stream of model outputs.

The overlooked half is refusal. A firm that evaluates ten opportunities and proceeds with one has learned something from all ten. The nine rejections never appear in any public record. They are the part of the dataset no competitor can see, buy, or reconstruct, and they often reveal a firm's judgment more clearly than its successes do. The research firm Alpha Matica calls this the shift from data moats to closed-loop moats: every logged judgment is a turn of a flywheel a rival cannot purchase, and "after 12–24 months of compounding, the gap often becomes uncrossable."

Where the loop compounds fastest

The loop does not compound equally everywhere. It compounds fastest in narrow, difficult fields: repeated decisions, high stakes, regulated access, proprietary cases that build a record nobody can scrape. Menlo Ventures draws a useful distinction between defensive moats (regulatory certification, compliance infrastructure) and generative moats, the compounding data that widens the gap over time. A Stanford Law white paper on vertical AI reaches the same conclusion from the legal side: in regulated industries the compliance burden becomes a barrier, and years of proprietary domain records become an asset no new entrant can recreate on demand. The strongest firms hold both: licenses alone create a protected position, not a learning company; data alone creates storage, not a loop. The value lies in joining the two: regulation buys time, and the learning loop makes productive use of it. Which tells you where to look: narrow domain, real stakes, private record. A horizontal chatbot has none of the three.

Running the test on one firm

The nearest place to watch the test run is the firm publishing this essay. Bamboo DCM works in Brazil's corporate and structured credit market. It takes a mid-market company's need for growth capital and turns it into a transaction institutional investors can fund: structuring, due diligence, distribution and post-deal monitoring, through a regulated operating model built around that loop.

Hold it against the checklist. The defensive moat is the licensed floor: two CVM licenses, the primary-offering coordinator license (Resolution 161) and the securitization license (Resolution 60), plus adherence to ANBIMA rules. They took years to earn and they apply to every deal. The generative moat is the record. Since 2022 the firm has structured over R$900 million across 25+ transactions, roughly 60% of them with first-time institutional issuers — not names the market already banks. The figures are modest, but their shape matters: the firm is repeatedly judging whether companies new to institutional credit can become investable, and under what terms. Behind the numbers sits the refusal data: roughly nine in ten opportunities screened are turned down, and the reasons for each pass are logged with the same care as the deals that close. Post-issuance monitoring keeps the loop running past signing, feeding evidence about how transactions behave back into the next decision.

Now run the questions. Swap the model, and the record of what was seen and decided stays put. Remove the intelligence layer, and screening, structuring and monitoring stop working at their current shape and cost, because the work was rebuilt around it rather than decorated with it. Check the scoreboard: not a leaderboard, but the firm's own history: was this risk read correctly, did this structure hold up as the analysis said.

What owning looks like

The test travels. Ask the three questions of any company with an AI slide in its deck. The firms that pass will mostly look unglamorous up close: records kept, reasons written down, rejections logged, evaluations run against their own past decisions. That is what owning looks like.

The model is rented. The loop is owned. When a company tells you it runs on AI, ask which one it is betting on.

References

Satya Nadella, "A frontier without an ecosystem is not stable (June 2026)" https://snscratchpad.com/posts/frontier-ecosystem/

CRV, "What is an AI-native company?" https://www.crv.com/content/what-is-ai-native

MBI Deep Dives, "Owning the Hill" https://www.mbi-deepdives.com/owning-the-hill/

Alpha Matica, "The Data Moat 2.0: Escaping LLM Commoditisation by Owning the Endless Feedback Loop" https://www.alpha-matica.com/post/the-data-moat-2-0-escaping-llm-commoditisation-by-owning-the-endless-feedback-loop

Menlo Ventures, "Software Finally Gets to Work: The Opportunity in Vertical AI" https://menlovc.com/perspective/software-finally-gets-to-work-the-opportunity-in-vertical-ai/

Stanford Law, "Defensible Moats for Vertical AI Application Companies in a New Competitive Landscape (June 2026)" https://law.stanford.edu/wp-content/uploads/2026/06/Defensible-Moats-for-Vertical-AI-Application-Companies-in-a-New-Competitive-Landscape.pdf

About Bamboo

Bamboo DCM is an independent structurer and distributor of corporate and structured credit in Brazil. We turn mid-market companies' growth-capital needs into transactions institutional investors can fund. Since 2022 we have brought over R$900 million to market across 25+ transactions, ~60% of them with first-time institutional issuers. We hold CVM coordinator (Resolution 161) and securitization (Resolution 60) licenses, and we are ANBIMA-adherent.

Regulated activities are conducted by Bamboo Securitizadora S.A. (CNPJ 48.343.871/0001-34), which acts as coordinator of public offerings under its CVM Resolution 161 coordinator license and issues and services CRI, CRA and debentures under CVM Resolution 60. This content is informational and is not an offer, recommendation or promise of returns.