Where the Moat Lives Once the Model Is a Commodity

Foundation models stopped being a moat the moment inference started falling 10x a year. When the core capability is a utility every competitor can rent from multiple vendors, defensibility has to live somewhere the price curve cannot reach. That somewhere is data network effects, proprietary data, and process power in vertical AI.

This is a playbook on where durable advantage actually sits once the model commoditizes. The short version. The model is the rented engine. The moat is what the engine is bolted to. At Avante Ventures we build for the second thing, because the first is no longer ownable.

The model is rented, so the moat moves

The model commoditized because inference got cheap faster than almost any technology in history. According to [a16z](https://a16z.com/llmflation-llm-inference-cost/), for an LLM of equivalent performance the cost is dropping by 10x every year, a factor of 1,000 in three years. GPT-3-level quality went from $60 per million tokens in late 2021 to about $0.06. Independent trackers at [Epoch AI](https://epoch.ai/data-insights/llm-inference-price-trends) confirm the direction, with some tasks falling 40x annually.

A 10x annual price drop, available to everyone, is the definition of a utility. So the strategic question is no longer which model you use. It is what compounds around the model that a competitor with the same model and more money cannot replicate by next quarter. Three mechanisms pass that test. Read them in order of durability, and see how they connect in [/why-avante](/why-avante).

Proprietary data: a stock that decays

A one-time proprietary dataset is a stock, and stocks decay. The sharpest correction to data-moat hype is still a16z's [The Empty Promise of Data Moats](https://a16z.com/the-empty-promise-of-data-moats/), which argues there generally is no inherent network effect from merely having more data. Worse, data hits diminishing returns. In their support-chatbot example, past roughly 40% query coverage the cost of adding unique data goes up while the value of each new record goes down.

Data is only defensible under tight conditions. Access has to be genuinely exclusive, or accuracy in a high-stakes domain has to drive a usage-to-feedback loop. A dataset a rival can also buy or scrape is a cost line, not a moat. Owning a pile of data is a head start. The moat is whatever keeps the pile growing faster than anyone can copy it.

Data network effects: the flow that compounds

A data network effect is the durable kind because it is a flow, not a stock. Each customer's usage improves the product for the next customer, so the asset refills faster than it decays. NFX, in its [Network Effects Manual](https://www.nfx.com/post/network-effects-manual), treats network effects as the core mechanism of defensibility and credits them with the majority of value created by technology companies.

The qualifier matters, and it is where most founders fool themselves. The loop only counts when more usage measurably improves the product on a dimension the customer cares about, and when a competitor starting cold cannot match that improvement. Scale alone does not do this. A feedback loop does.

• Scale, not a moat. A new entrant with the same model and more capital can replicate your data position within a quarter.
• A real flow. Your product is structurally better the longer it runs, because usage feeds an asset rivals cannot buy.
• The test. If a well-funded competitor cannot catch your data position by next quarter, you have a network effect, not just scale.

Process power and workflow lock-in

Process power is the moat that compounds inside regulated workflows. Hamilton Helmer's [7 Powers](https://blas.com/7-powers/) defines Process Power as operational excellence plus hysteresis, the resistance that keeps a process hard to copy even when its outputs are visible. It pairs with Switching Costs, which arise when a customer values compatibility across repeated purchases from one firm over time.

When an AI product becomes the system of record for a regulated process, leaving means re-validating a compliance trail, retraining staff, and re-integrating adjacent systems. The cost of leaving is the moat. This is why vertical AI beats horizontal AI. A general assistant has no workflow to anchor. A vertical product that lives inside a licensed, audited, regulation-dense process anchors deep and stays.

Mapping it onto 7 Powers

Helmer's framework is useful here precisely because it forces honesty about which power you actually hold. The model is none of them. It is an input every rival shares. The durable AI moats map onto three of his seven, and the 2025 evidence backs it.

According to [Insignia Ventures](https://review.insignia.vc/2025/04/15/moats-ai/), AI made building easier and defending exponentially harder, with software reaching $1 million ARR faster than ever. Their emerging-market case studies land on the same three powers. A used-car platform compounds a data flywheel from 160-plus data points per vehicle. A lender pairs a proprietary ERP with financing and holds 3% NPL versus an industry 20 to 30% through a downturn. A digital bank turns a scarce regulatory license into distribution no rival can match.

• Network Economies. The data network effect, where each user makes the product better for the next.
• Process Power. Operational excellence plus hysteresis, the workflow lock-in of a system of record.
• Cornered Resource. A scarce license or exclusive data feed a competitor cannot rent or scrape.
• Switching Costs. The re-validation, retraining, and re-integration cost of leaving a regulated workflow.

The anti-moats: wrappers and buyable data

Knowing what is not a moat is half the playbook. A thin wrapper on a public API rents the capability every competitor can rent, adds a prompt, and owns no compounding data. None of Helmer's powers apply to it. A proprietary dataset a rival can buy or scrape is a cost line wearing a moat's clothing.

The other failure modes are quieter. Data scale gets mistaken for a data network effect, when more rows without a usage-driven loop is just a decaying stock. Model dependency masquerades as strategy, until a 10x annual price drop and multi-vendor availability erase it. And a clean dataset with no channel to deploy it loses to a worse dataset already living inside a workflow customers use every day.

How Avante engineers the moat

Avante Ventures is a venture studio building AI-native companies in Brazil and Latin America. The method is a copilot to data to fund flywheel. Build an AI copilot to do real work inside a vertical, capture the proprietary data the work generates, then use that data asset to raise and deploy capital. The copilot is the wedge. The data network effect is the moat. The capital is the compounding.

The studio launches 3-4 ventures per year through a six-stage system. Research, Partner, Build, Traction, Revenue, Compound. Capital per venture is $500K-1.5M across pre-seed, with co-founder economics retained. The model is benchmarked to GSSN data showing studio IRR of ~50% versus ~19% for traditional VC, roughly 2.5x, a benchmark for the studio model rather than a claim on Avante's own return.

The reason this works in LATAM is structural. Services account for roughly 70% of Brazilian GDP with low software penetration, the exact surface where a vertical product can become the system of record. Pair that with domain operators carrying 10+ years of Brazilian-market scar tissue, and the data compounds in places competitors cannot reach. Judicial-asset workflows and insurance risk scoring are flows, not stocks. The model will keep getting cheaper. The moat is everything you build so that no longer matters. See how we operate in [/principles](/principles).