Gary argues founders should find a real customer problem and solve it before trying to forecast a long-term moat. The panel says speed is an early advantage: Cursor shipped features in daily cycles, while larger companies can take weeks or months. Think about defensibility once customers value what you have built.
The 7 Most Powerful Moats For AI Startups
Speakers
Y Combinator hosts and guests Jared, Diana, Gary, Harj, Dan, and Harsh discuss seven durable business moats for AI startups on The Light Cone, drawing on examples from companies including Cursor, Plaid, and Exa.
Process power comes from building complex systems that are hard to reproduce. Examples include AI agents for banking and legal work, where a demo may be quick to build but reliable, mission-critical performance takes deep engineering; Plaid’s integrations with many financial institutions are another example. Cornered resources can include regulatory access, customer workflows and data, or specialized models.
Switching costs can come from migrating data, retraining teams, or replacing deeply customized AI workflows. Happy Robot’s logistics integrations and Salient’s banking workflows illustrate how long enterprise pilots can lead to hard-to-replace systems. Counter-positioning can favor AI-native products when incumbents struggle to adopt task-based pricing or risk cannibalizing per-seat revenue; the panel also points to Speak’s focus on spoken language practice and Legora’s application-layer approach.
The panel describes brand as a reason customers choose a product even when alternatives are comparable, and notes ChatGPT’s lead over Gemini as an unexpected consumer AI example. AI network effects can come from usage data and evaluations that improve models and workflows over time. Scale economies matter most at the model layer, while Exa illustrates an application company investing in costly web crawling that can serve many customers.
The closing advice is to prioritize speed and find a specific, severe customer problem rather than choosing an idea based on a predicted moat. The panel distinguishes mild convenience from urgent pain, such as a risk of being fired or a business failing. Build something that relieves that pain, then develop defenses as the business grows.
Takeaways
- Start by finding a customer with an urgent problem; moat analysis comes after you have something valuable to defend.
- Use startup speed to discover valuable AI workflows before larger companies commit to them.
- For enterprise agents, reliability and integration depth can matter far more than a convincing demo.
- Treat customer workflow data and evaluations as assets that can improve the product with usage.
- Consider whether incumbent pricing and processes make it hard for them to adopt an AI-native product.
Mentioned in this talk
Windsurf
Anthropic
Cursor
Google
OpenAICaseTextGreenlightCasca
Plaid
Stripe
Scale AI
Palantir
Character.AI
SalesforceSalientHappy Robot
DHLAvocaServiceTitanGigaMLLoraHarvey
DuolingoExaThe Light Cone
ChatGPTThe Seven Powers: The Foundations of Business Strategy
Claude Code
Gemini
Evaluations
Strands speak toolAI agentslarge language modelsspeedforward-deployed engineeringprocess powercornered resourcescontext engineeringFine-tuningswitching costscounter-positioningpersonal brandnetwork effectsscale economies
Sam AltmanHamilton HelmerVarunSearch inside this talk
What was claimed 20
- builtAuthor
“The book was written by Hamilton Helmer”
YouTube Unverified1 source· Talk
- featureCursor: Product development and sprint cycles for shipping features were one day
“his product development and sprint cycles for shipping features were one day”
YouTube Unverified1 source· Talk
- featureCasca: Provides loan origination support to banks
“Casca does loan origination”
YouTube Unverified1 source· Talk
- featureGreenlight: Provides KYC processing for banks
“Greenlight, which worked with Tom, does KYC for banks”
YouTube Unverified1 source· Talk
- featurePlaid: Supports a very large number of financial institutions and their websites
“The number of financial institutions they have to support is enormous”
YouTube Unverified1 source· Talk
- featureScale AI: Does extensive work with the Department of Defense
“Scale AI, which is doing a ton of work with the DoD”
YouTube Unverified1 source· Talk
- featurePalantir: Works with government and builds AI capabilities for warfare-related uses
“you’ve got to go through Palantir or Scale”
YouTube Unverified1 source· Talk
- featureCharacter.AI: Fine-tuned models to reduce serving costs by 10x
“fine-tuned them to bring down the cost of serving those models by 10x”
YouTube Unverified1 source· Talk
- featureHappy Robot: Provides AI agents integrated into customized logistics workflows
“They start with workflows that are highly customized for each company”
YouTube Unverified1 source· Talk
- featureSalient: Builds AI voice agents for financial workflows
“Salient is building an AI voice agent for the financial industry”
YouTube Unverified1 source· Talk
- integrates withHappy Robot integrates withDHL
“Happy Robot has customers like DHL, where they integrated deeply into many workflows”
YouTube Unverified1 source· Talk
- featureAvoca: Builds customer support software for HVAC companies
“Avoca, for instance, is building customer support software like ServiceTitan, but for HVAC”
YouTube Unverified1 source· Talk
- metricAvoca: Can capture 4% to 10% of an HVAC company’s spending, compared with about 1% for ServiceTitan
“They can get HVAC companies to pay for customer support ... 4% to 10%”
YouTube Unverified1 source· Talk
- featureGigaML: Offers a customer support product that works out of the box with faster sales and onboarding
“the product fundamentally works better out of the box”
YouTube Unverified1 source· Talk
- featureLora: Focuses on the application layer and product quality in legal AI
“Legora has focused on that”
YouTube Unverified1 source· Talk
- featureStrands speak tool: Uses an LLM and voice for language practiceUses an LLM and voice for language practice
“Speak, which uses an LLM and voice to help you practice”
YouTube Unverified1 source· Talk
- metricChatGPT: Has more daily consumers than Google Gemini, according to the speaker
“ChatGPT has more consumers using it per day than Google’s Gemini”
YouTube Unverified1 source· Talk
- uses
“they’ll use the data to train it”
YouTube Unverified1 source· Talk
- featureExa: Crawls a large portion of the web and reuses the crawl for multiple customers
“Once you've crawled a big chunk of the web, you can reuse that same crawl for many customers”
YouTube Unverified1 source· Talk
- featureExa: Provides an API for AI applications to search the web
“It provides an API for anyone building AI applications who wants to search the web”
YouTube Unverified1 source· Talk
