Seven moats for AI startups: applying the Seven Powers framework

Executive overview

Most college students and early founders worry about moats before they have anything worth defending. The real question is not whether a moat exists on day one — it does not — but whether the problem is painful enough to solve first.

Speed is the only moat that matters at zero. Once you have traction, seven structural moats become available. The Seven Powers framework (Hamilton Helmer, 2016) maps every durable competitive advantage a business can hold. These categories have not changed; only the specific AI-era manifestations are new.

Speed: the moat that precedes all others

• Speed is not in Helmer's book but is the dominant early-stage moat.
• Large companies cannot ship at startup pace — cursor ran one-day sprint cycles in 2023–24.
• Incumbents face coordination overhead: PRDs, PMs, approval chains. Startups do not.
• Speed buys time to discover which verticals are genuinely valuable before defending them.

Process power: the complexity moat

• Process power means building something so operationally complex that replication is prohibitively costly.
• The hackathon version of any AI agent takes a weekend. The production version takes years.
• Mission-critical reliability (e.g. KYC for banks, loan origination) requires handling edge cases that require deep domain knowledge just to know they exist.
• Plaid's moat: thousands of financial institutions, each requiring custom crawlers and integration logic — surface area too wide for any fast follower to replicate.
• Existing SaaS companies (Stripe, Rippling, Gusto) hold this moat too — years of backend logic that cannot be cloned from a landing page.
• The last 10–20% of reliability is painstaking drudgery; large-lab teams optimising for AGI will not prioritise it.

Cornered resources: proprietary access others cannot buy

• Cornered resources are assets that cannot be arbitraged — patents, regulatory relationships, unique data.
• Government AI contracts (Palantir, Scale AI) require years of embedding, cleared facilities, and trust — once held, they are written into procurement doctrine.
• For most startups the relevant form is data and workflows acquired through forward-deployed engineering: sitting with a customer, mapping every step of a process, translating it into prompts, evals, and fine-tuned models.
• Each workflow captured is a dataset no competitor can buy.
• Own model trained on proprietary data is the strongest form — but not the only one; context engineering alone covers most needs for the first two years.

Switching costs: the cost of leaving

• Switching costs make it expensive to migrate even when a better alternative exists.
• Classic form: data locked in legacy systems (Oracle databases, Salesforce CRM) — migration costs a year of productivity.
• AI-native form: lengthy onboarding and deep workflow customisation per customer (6–12 month pilots at Happy Robot, Salient).
• Once a 7-figure enterprise contract is live, no buyer will rerun a bake-off; the switching cost is in the customised logic, not just the data.
• Consumer memory (e.g. ChatGPT accumulating preferences over time) is an emerging personal switching cost.
• Counterpoint: LLMs may dramatically reduce classic data-migration switching costs by automating schema translation and browser-based export — a lever startups can use against incumbents.

Counter-positioning: doing what incumbents cannot copy

• Counter-positioning means taking a position that would cannibalise the incumbent if they tried to match it.
• SaaS incumbents charge per seat. AI agents that work will reduce headcount — incumbent success shrinks their own revenue.
• Founder-controlled companies can navigate this self-disruption; non-founder-controlled ones rarely do.
• AI-native startups price on tasks completed, which aligns incentives with actually delivering work.
• Incumbent engineering cultures often cannot ship AI-native products even when they try — per-seat pricing and inability to do the work reinforce each other.
• Avoka (HVAC software): captures 4–10% of customer spend on customer support vs. the 1% wallet share typical in vertical SaaS — unlocking spend categories that were never software budgets.
• Second-mover counter-positioning: Legora entered legal AI after Harvey, competed on application-layer quality rather than fine-tuning. Speak competes against Duolingo on actual language acquisition, not gamification.

Brand: recognition that persists past product parity

• Brand becomes a moat when consumers choose a product even against equivalent alternatives.
• ChatGPT acquired the consumer AI brand against Google despite Google owning the world's largest consumer internet brand — a near-unprecedented brand reversal.
• Google's ad business made it structurally unable to ship an aggressive AI product; OpenAI had no such constraint.
• For early-stage startups, brand takes too long to build to be a primary strategy — but counter-positioning can establish brand positioning faster.

Network economies: value that compounds with users

• Network economies: product value increases as more users join.
• Classic form: social networks (Facebook), payment rails (Visa).
• AI-era form: data flywheels — more usage generates more training signal, which improves the model, which improves the product.
• Cursor's free tier collects keystroke-level data across millions of developers; this feeds autocomplete models that improve with scale.
• Enterprise AI agents accumulate private workflow data per customer; combined with evals on failures, this creates a compounding improvement loop competitors cannot replicate without the same customer base.
• Evals are the operational mechanism that turns usage into moat — they are not a one-time exercise but a continuous flywheel.

Scale economies: infrastructure that gets cheaper with volume

• Scale economies: large fixed-cost infrastructure amortised across growing volume.
• Applies most clearly at the model layer — training frontier LLMs requires capital only a handful of companies can deploy.
• DeepSeek's announcement challenged the assumption that scale was unassailable at the lab level, but it still required expensive foundation models as a base.
• At the application layer, scale economies are rare — the main example is web-crawl infrastructure.
• Exa built a large proprietary web crawl before the agent wave; that fixed investment now serves many API customers at low marginal cost. Channel3 and Orange Slice (YC) are replicating this model.

How to use this framework

• Do not use moats to decide between startup ideas — you cannot forecast which moat will matter in five years.
• A moat is defensive. If you have nothing to defend, the framework is irrelevant.
• Find a person with a genuinely painful problem, solve it, and ship fast. The moats surface as you work with customers and build the product.
• Speed is where every startup begins. Process power and cornered resources are where most winners end up.