Founders are obsessed with moats right now—and for good reason. In a world of near-infinite competition, margins trend to zero unless you can defend something real. But here’s the uncomfortable truth: early on, the only moat you actually have is speed.

Not “we ship fast-ish.” I mean Cursor-level speed—one-day sprints in 2023–2024—while big companies take weeks, months, sometimes years to push features through PRDs and committee. In greenfield markets where nobody knows which products matter yet, the team that cycles daily and learns fastest wins the right to worry about moats later.

Speed is missing from Hamilton Helmer’s Seven Powers, but it shouldn’t be. It’s the gateway power. Ship relentlessly; make something people truly want; then stack the classic moats as you scale. That’s the actual sequence. If you’ve got nothing valuable yet, your “moat” is just a puddle.

Once you have traction, process power shows up first. Think of what banks demand from AI agents handling KYC or loan origination. A hackathon demo gets you 80% of the way with 20% of the effort; production-grade reliability on tens of thousands of decisions per day requires the last 1–5% to work almost all the time—and that last mile takes 10–100× the effort. That drudgery is a moat. Plaid-style surface area across thousands of financial endpoints, CI/CD that never breaks, evals that catch edge cases—this is why Stripe, Rippling, and Gusto are hard to copy. Better engineering, done repeatedly, compounds.

Cornered resources come next. Sure, in pharma that’s patents. In modern AI, it’s privileged access: regulated buyers, DoD environments, or proprietary customer workflows and data you collect by being a forward-deployed engineering team. That proprietary data lets you tune models and prompts so your unit economics improve—Character-AI-style 10× serving cost reductions are the blueprint. Having your own best-in-class model helps, but it’s not mandatory on day one; careful context engineering will get you 80–90% of what customers need for the first two years.

Switching costs are evolving, too. The old world was Oracle or Salesforce: migrating schemas and retraining a sales org could cost a year of productivity. LLMs will lower those data-migration costs, but AI startups are creating a new lock-in: months-long onboarding that encodes custom logic and compliance into agents. Six- to twelve-month pilots that convert to seven-figure contracts make a second bake-off irrational. On the consumer side, memory is becoming sticky—tools that actually remember you raise the pain of leaving.

Counter-positioning is quietly lethal. Incumbent SaaS sells per seat; good agents reduce seats. The better their AI, the more revenue they cannibalize. Startups price on work delivered or tasks completed—and then they actually deliver. That culture shift is nontrivial for late-stage incumbents. Second movers who out-execute often win: legal AI teams focusing on application quality over fine-tuning aesthetics; customer support agents like Giga ML that “just work” faster in onboarding. Agents also have superhuman edges: instantly handle 200 languages, infinite patience on bad connections. In vertical SaaS, this flips wallet share: from ~1% “software” take to 4–10% when you absorb operations (AOKA’s HVAC support example). That’s not a feature; that’s a business model moat.

Network effects in AI look like data flywheels and eval pipelines, not just “more friends = more fun.” The more usage you have, the more ground-truth failures you capture, the better your prompts, tools, and models get. Cursor’s telemetry—every keystroke improving autocomplete—compounds quality. Brand still matters (ask Google how it feels to chase ChatGPT), but the durable edge is usage → data → better product → more usage.

Finally, scale economies mostly live at the foundation layer. Training frontier models and crawling large slices of the web (think EXA’s “search for agents”) are capital-intensive, with low marginal costs at scale. Even with DeepSeek-style RL efficiencies, the base models remain expensive—another reason application-layer speed matters early.

So here’s the playbook. Find existential pain—work that’s so broken someone’s promotion or business is on the line. Ship daily until you own that pain. Use the speed moat to earn time, users, and cash. Then layer in process power, cornered resources, switching costs, counter-positioning, network/data effects, and—when relevant—scale. Think five years out, sure, but execute like you only have five days. Because in the beginning, you do.