Y CombinatorHow Hamilton Helmer's 7 Powers Apply to AI Startups
What happens when you map Hamilton Helmer's 7 Powers to AI startups: counter-positioning and switching costs win; speed alone is not a moat.
CHAPTERS
- 0:00 – 1:40
Why AI founders feel the “moat problem” so acutely
The hosts frame why moats suddenly loom larger for AI startups: the fear that anything can be cloned as a “ChatGPT wrapper,” or that model labs will subsume the product. They emphasize moats as the defense against competition driving margins to zero.
- •AI intensified founder anxiety about defensibility and copyability
- •“Wrapper” meme vs. building an enduring business
- •Moats as protection against infinite competition and margin compression
- •Founders often over-index on moats before proving value
- 1:40 – 4:20
The 7 Powers framework (and why it’s basically “seven moats”)
They introduce Hamilton Helmer’s book The 7 Powers and position it as a timeless categorization of defensibility, even if the examples are dated. The episode’s goal is to translate the framework to modern AI startups.
- •The 7 Powers = categories of durable competitive advantage
- •Framework is stable even as AI changes implementations
- •Rebooting 2000s-era examples for 2025 AI companies
- •Use the framework as a lens, not a blocker to starting
- 4:20 – 5:42
When to think about moats: after you’ve found something worth defending
The group warns founders not to reject ideas because the moat isn’t obvious upfront. The right sequence is: find a painful problem, build something people want, then discover moats through customer work and product iteration.
- •Early-stage priority: solve a real pain point (zero-to-one)
- •Moats tend to emerge while building with customers
- •Don’t use moats to choose between ideas via 5-year forecasts
- •A moat is defensive—no point without something valuable to defend
- 5:42 – 8:11
Speed as the “missing” moat—and why startups can outship incumbents
They argue speed is often the only early moat, especially against large labs and incumbents. Examples like Cursor show how radically short iteration cycles can create a durable lead while big companies move slower due to process overhead.
- •Speed is a real early moat (even if not in the book)
- •Big-company “cruft” slows shipping; startups can iterate daily
- •Cursor’s one-day sprint cycle as a case study
- •Relentless execution as an edge versus labs and big tech
- 8:11 – 10:17
Forward-deployed engineering: discovering data + workflows in the field
They describe a common AI startup pattern: embed with customers to learn messy real workflows and collect the ingredients for defensibility. Early on, startups explore what’s valuable; later they defend once the “gold vein” is found.
- •Startups act like exploratory teams in greenfield AI markets
- •Embed with customers to uncover valuable verticals/workflows
- •Defensibility increases once the value is proven and scaled
- •Moats become relevant as competition arrives after success
- 10:17 – 14:35
Process Power: the hard-to-replicate “last 10%” of reliable agents
Process power is framed as the accumulated complexity and engineering required to make an agent work reliably in production. Hackathon demos are easy; mission-critical bank or legal agents require years of edge-case handling, evals, and operations.
- •Process power = operational/engineering complexity that’s hard to copy
- •Demo agents vs. production-grade reliability gap
- •Examples: Casetext; agents for KYC and bank loan origination
- •Schlep blindness: painstaking edge cases become the moat
- 14:35 – 19:28
Cornered Resources: regulatory access, proprietary data, and custom models
They define cornered resources as scarce assets that confer durable advantage—patents, regulatory approvals, privileged distribution, or proprietary datasets/models. In AI, this often comes from deep customer access (data/workflows) or owning a specialized model.
- •Cornered resources can be patents/regulation or privileged access
- •Gov/DoD examples: Scale AI, Palantir; hard-to-enter channels
- •Startup-scale version: customer-embedded data/workflow capture
- •Owning a specialized model can be a resource, but not the only moat
- 19:28 – 24:56
Switching costs in AI: from data migration to memory + deep customization
They explain classic switching costs (systems of record like Oracle/Salesforce) and how AI can reduce migration friction. They also highlight a newer AI-native switching cost: long onboarding that bakes custom agent logic into a company’s operations, plus consumer “memory” personalization.
- •Classic switching costs: data/workflow migration pain in SaaS
- •AI can lower migration cost via schema-mapping and automation
- •AI-native switching cost: deep custom agent onboarding (long pilots)
- •Consumer switching costs emerging via agent memory/personalization
- 24:56 – 29:49
Counter-positioning: incumbents trapped by seat pricing and culture
Counter-positioning is winning with a model incumbents can’t copy without self-harm. The hosts argue per-seat SaaS pricing conflicts with AI automation (fewer seats), and many incumbents struggle culturally to become AI-native—opening room for startups priced on outcomes.
- •Counter-positioning = incumbent can’t copy without cannibalization
- •Per-seat pricing becomes an Achilles’ heel when AI reduces headcount
- •Startups shift to pricing by work delivered/tasks completed
- •Incumbents often struggle to reset engineering culture for AI
- 29:49 – 31:28
Workforce displacement reality: automation as job transformation (Avoka example)
They address fears about AI replacing workers by highlighting high-attrition roles like customer support in trades. In cases like Avoka, AI can improve economics and shift humans to higher-leverage work (managing agents, handling edge cases) rather than eliminate stable jobs.
- •Many support jobs are unpleasant and already high-attrition
- •AI agents reduce toil and improve service quality
- •Humans move to oversight and complex exception handling
- •AI can increase wallet share by automating labor, not just software
- 31:28 – 35:12
Second-mover counter-positioning: beating early AI winners with better strategy
They discuss how later entrants can win by choosing a different bet than the early leader. Examples include Legora vs. Harvey (product/application focus vs. fine-tuning differentiation) and GigaML vs. established support-agent players (faster onboarding, out-of-box performance).
- •Second movers can win by learning from first movers’ wrong turns
- •Legora vs. Harvey: emphasize product layer over fine-tuning narrative
- •GigaML: faster onboarding and better default performance as wedge
- •AI agents can be superhuman (languages, patience, consistency)
- 35:12 – 37:26
Brand and speed: how ChatGPT outran Google (and why that’s strategic)
They argue brand can be a moat but takes time—yet AI is a rare case where OpenAI built a consumer brand faster than Google could respond. They link this to counter-positioning: Google’s ad-driven model and organizational constraints slowed disruptive action, while OpenAI shipped quickly with a small team.
- •Brand becomes a moat once consumers default to you
- •ChatGPT achieved greater daily usage than Gemini despite Google’s reach
- •Google constrained by ads/cannibalization and organizational inertia
- •Origin story underscores speed and focus as decisive advantages
- 37:26 – 41:03
Network economies in AI: usage-driven data + eval flywheels
They reinterpret network effects for AI as data and evaluation loops that improve models and product quality with scale. Examples include ChatGPT learning from interaction history and Cursor improving from fine-grained telemetry; enterprise agents improve through private workflow data and evals.
- •AI network effects often manifest as data/model improvement loops
- •Chat history and feedback can feed training and iteration
- •Cursor-style telemetry: clicks/keystrokes improve autocomplete
- •Enterprise agents: private data + evals create compounding advantage
- 41:03 – 43:57
Scale economies: model-layer defensibility and application-layer infrastructure bets
They describe economies of scale as large fixed investments that lower unit costs, most visible in frontier model training. They also highlight emerging application-layer scale plays like Exa (web crawl infrastructure) that can be amortized across many customers and become a moat for agent ecosystems.
- •Training frontier LLMs is capital-intensive (scale moat at model layer)
- •DeepSeek reframed perceived training costs and shook assumptions
- •Application-layer example: Exa’s expensive web crawl reused across customers
- •Expect more crawl/infra-backed agent companies as web agents improve
- 43:57 – 45:05
Final advice: don’t start with moats—start with painful problems and ship fast
They close by re-emphasizing that founders should prioritize discovering acute pain and delivering value quickly. Moats matter later, but the foundational advantage early is speed and execution toward a problem customers truly feel.
- •Lead with “make something people want,” not moat theorizing
- •Look for existential, high-stakes customer pain
- •Go zero-to-one first; moats emerge from real usage and iteration
- •Speed remains the dominant early advantage
