Video summary

The 7 Most Powerful Moats For AI Startups

Main summary

Key takeaways

Business

Core idea: why “moats” matter for AI startups

  • The discussion frames moats (defensibility against infinite competition) as existential for startups: without something to defend, competitors can drive margins → ~0.
  • AI/agent businesses can have deep moats, but they’re often non-obvious early.
  • Early-stage guidance: don’t delay building while trying to forecast a moat 5 years out. Instead:
    • find a painful, high-value problem
    • prove product value first

Business framework: “The Seven Powers” / “Seven Moats”

Presented as Hamilton Helmer’s Seven Powers (reinterpreted as Seven Moats), updated for AI startups (2025 context). The practical message: different moat “types” can apply depending on the company and vertical.

1) Speed (added / emphasized beyond the book)

  • Only early moat is execution speed (e.g., ship faster than incumbents).
  • Concrete example: Cursor
    • Reported feature sprint cycles ~1 day in the early 2023–2024 period.
    • Strategic point: larger labs/incumbents have heavier process (PRDs/specs/ops), so startups can outrun them early.

Actionable takeaway

  • Optimize for: short release loops, rapid iteration, “ship daily/weekly,” and customer-driven learning.

2) Process Power (systems that are hard to replicate)

  • Moat comes from building complex, hardened execution systems that only look like a simple product at the surface.
  • AI translation: a “hackathon version” is not defensible.
    • Defensibility comes from reliability under real-world conditions, including edge cases and operational hardening.

Concrete examples

  • Mission-critical AI agents for banking
    • Greenlight (KYC for banks), Casca (loan origination).
    • Claim: a weekend demo is useless; production systems protect against millions of dollars in bank losses.
  • Plaid as process power via enormous integration surface area:
    • many institutions
    • CI/CD complexity
    • rapid onboarding capability

Operating insight

  • “Last 10%” engineering reliability can require 10x–100x more effort when you need ~99% accuracy for real workflows.

3) Cornered Resources (hard-to-replicate assets / access)

Resources are scarce because they’re:

  • Regulated (patents + FDA-approval-like constraints)
  • Access-gated (government procurement + specialized facilities)
  • Embedded expertise (people/“brain space” inside government/customer orgs)
  • Exclusive data/workflows gained via deep customer integration

Concrete examples

  • Scale AI / Palantir
    • government access
    • DC/agency presence
    • special data centers
    • embedded relationships
  • Forward-deployed engineer model (FDE)
    • Startups embed with customers to extract tailored workflows:
      • email request → enrichment → sometimes call-center touches
      • translating process into prompts/evals/data
  • Character.AI cost moat
    • Fine-tuned models to reduce inference cost ~10x (as described)

Strategic point

  • “Best cornered resource” may be a specialized model that performs work better at lower cost—but it’s not the only moat.

4) Switching Costs

  • Classic definition: customer is “trapped” because switching is painful/expensive in time, money, and operations.
  • AI-era twist: switching costs increasingly come from:
    • onboarding
    • deep customizations of agent logic
    • not just data migration

Concrete examples

  • Enterprise AI agent pilots lasting ~6 months to 1 year
    • converting to seven-figure contracts
  • Happy Robot
    • deep workflow integration (e.g., DHL logistics operations)
  • Salient
    • integrates with banks’ distinct workflows:
      • loan consolidation
      • debt recovery
      • fraud monitoring
      • compliance

Counterpoint / leverage

  • AI may also reduce switching costs via LLM/codegen that transforms data schemas and can automate migration.
  • Two switching-cost “flavors”:
    • Old SaaS: painful data migration between systems of record
    • New AI: lengthy onboarding that leads to deep workflow/logic customization
  • Consumer analog: memory/personalization increases switching pain.

5) Counter-Positioning (beat incumbents by doing something they can’t copy)

  • Definition: a move that is difficult for the incumbent to imitate without cannibalizing their own business.
  • Observed AI SaaS dynamic:
    • incumbents building agents often use per-seat pricing
    • if AI reduces needed headcount, revenue can shrink as automation replaces staff

Concrete pricing strategy shift

  • Many AI startups price around:
    • work delivered / tasks completed
    • rather than seats
  • This aligns incentives with outcomes (but requires real ability to complete the work reliably).

Operational/organizational problem (leadership insight)

  • Late-stage companies struggle to:
    • reset engineering culture to be AI-native
    • embrace context/prompt engineering
    • ship products that actually “do the work”
  • Without that, they can’t justify outcome-based pricing.

Vertical AI SaaS example (wallet share expansion)

  • Avoca (HVAC customer support software)
    • starts with a small budget surface area (customer support spend)
    • grows share of wallet over time:
      • from ~1% wallet share (example comparator: ServiceTitan’s situation)
      • to ~4%–10% as it ties into additional spend categories
  • Workforce displacement framing:
    • AI improves support quality; humans often leave anyway due to high attrition (reported 50–80% annual attrition)
    • people shift from managing call centers to managing AI agents and handling edge cases

6) Innovation in the application layer (second-mover advantage)

  • In verticals, an early winner may emerge, but second movers can win by counter-positioning through better product focus.

Concrete examples

  • Legora vs Harvey (legal AI)
    • Harvey differentiated early via fine-tuning
    • Legora’s counter: focus on the application layer/product, not only model tuning
    • reportedly resonating in branding and sales
  • Giga ML vs incumbents
    • claim: product works better out of the box
    • enables faster sales/onboarding than “well-known” customer support companies
  • Duolingo vs Speak (consumer language learning)
    • Duolingo criticized as “game app vs language learning”
    • Speak differentiates by learning through speaking using voice/LLM; avoids competing on gamification

Underlying principle

  • Counter-positioning overlaps with branding moats: consumers choose the “place for the job,” not just features.

7) Network Effects (AI-era form: data + eval flywheels)

  • Classic network effects: value increases with more users (Facebook, Visa).
  • AI translation: network effects increasingly become:
    • more data → better models → better product → more usage
    • plus an eval flywheel:
      • more usage generates workflow outcomes
      • improves evals and iteration quality

Concrete examples

  • ChatGPT → GPT future training loop (described as chat data feeding subsequent model training/reward signals)
  • Cursor
    • reported use of user keystroke/tab completion data for training
    • more developers → better autocomplete
  • Enterprise angle
    • working with large companies grants private data + workflows
    • improves model/evals

(Additional) Scale Economies (mostly discussed as model layer)

  • Classic definition: lower unit costs due to scale.
  • In AI, the scale-economy moat is argued to mostly live at the model layer:
    • training frontier LLMs is capital intensive
    • inference becomes cheaper once the model exists

Concrete example

  • DeepSeek
    • framed as significant due to publicized cheaper RL unlocks
    • however, it still builds on expensive foundation models
    • so the scale moat is not eliminated

Business-level execution example: Exa

  • Exa crawls large web portions with significant fixed investment.
  • Moat: once crawled, the dataset can be reused for many customers.
  • Mentioned: Channel 3 and Orange Slice running similar “crawl once, serve many agents” patterns.

Actionable playbook for early-stage AI startup founders

  • Start with existential pain, not moat hunting
    • find a customer outcome that’s “I might get fired / we could take over everything next year”-level urgent
  • Go from 0→1 by solving the painful workflow first
    • moats will “stumble upon themselves” as you learn
  • Plan your moat as you scale
    • early: speed
    • after product-market fit / scaling begins: process power, switching costs, data/evals network effects, and cornered resources
  • Don’t forecast 5 years out to choose ideas
    • moats are defensive and depend on real assets/customers

Key metrics / targets mentioned

  • Reliability target: ~99% accuracy for certain mission-critical agent workflows
  • Engineering effort nonlinearity:
    • achieving the “final 5%–10%” may require ~10x to 100x effort
  • Cursor speed example:
    • ~1-day sprint cycles for shipping features (early period)
  • Enterprise pilot duration:
    • ~6 months to 1 year
  • Enterprise contract size:
    • seven-figure contracts (exact range not specified)
  • Inference cost reduction example (Character.AI):
    • ~10x
  • Avoca wallet share example:
    • ~4%–10% (vs earlier ~1% comparator framing)
  • Workforce stat (HVAC customer support attrition):
    • 50%–80% annual attrition
  • Switching-cost dynamic:
    • LLMs may reduce data migration costs (no explicit KPI; qualitative only)

Presenters / sources

Presenters (speakers): Jared (host/guest), Gary, Diana, Harj, Dan, Tatiana

  • Plus references to guests such as Sam Altman, Michael Truel, Bob McGrew, and Varun (Winstorf)

Framework/source book: Hamilton Helmer, The Seven Powers: The Foundations of Business Strategy (2016)

Referenced companies (examples): Cursor, OpenAI, Anthropic, ChatGPT, Claude Code, Case Text, Greenlight, Casca, Plaid, Stripe, Rippling, Gusto, Oracle, Salesforce, Happy Robot, Salient, DHL, Zendesk, Intercom, Front, Avoca, ServiceTitan, Legora, Harvey, Giga ML, Duolingo, Speak, Character.AI, Scale AI, Palantir, Visa, Exa, Channel 3, Orange Slice.

Original video