LLenny's Podcast
← All frameworks
InnovationPeter Deng

The AI Startup Moat: Data Flywheel + Crafted Workflow

Defensibility for AI apps comes from a proprietary data flywheel and a deeply crafted vertical workflow.

Difficulty
Advanced
Time to result
~ongoing to results
Steps
4
Confidence
92%

Deng's investment thesis for companies building on top of LLMs: raw model access is commoditized, so durable advantage comes from (1) a data flywheel — proprietary data plus a mechanism that keeps generating more of it as the product is used — and (2) ergonomic integration into a specific workflow so good that people switch and tell their friends. Even without unique starting data, usage can generate it.

Origin

Peter Deng's thesis as a general partner at Felicis, drawn from operating at OpenAI and investing in AI startups.

Core principles

  • 01Models get really good at whatever data you show them — AI is malleable, not magic
  • 02Proprietary data is the seed; the flywheel is how you keep generating and compounding it
  • 03Even without unique starting data, product usage can collect it (e.g. accept/reject signals)
  • 04Product craft can overcome incumbent distribution — but only if the product is dramatically better
  • 05Data flywheel and workflow go hand in hand: solving something genuinely valuable earns the usage that feeds the flywheel

How to run it

  1. 1

    Identify your proprietary data seed

    Be mindful of what data you have access to that others don't, and use it to start a flywheel.

    Pro tip If you lack unique data upfront, design the product so its usage collects distinctive data (e.g. which suggestions users accept or reject).

  2. 2

    Engineer the flywheel

    Build the mechanism that continuously maintains and generates more of that data as people use the product, so the model gets smarter at your specific task over time.

    Watch out Assuming AI is a 'magic wand' misses that it only does well what it's been trained on — no data engine, no compounding edge.

  3. 3

    Craft the ergonomic workflow

    Obsess over how the product integrates into people's actual workflow and lives — the sanded-down edges that make it delightful to adopt.

    Pro tip Pick a vertical you understand deeply and solve that specific workflow uniquely well.

  4. 4

    Clear the high bar to beat incumbents

    Accept that incumbents have distribution advantages; only a product that is so much better earns the switch. Aim for craft that makes people install it and evangelize it.

    Watch out A marginally-better product will lose to incumbent distribution; the bar to break through is genuinely high.

In the wild

Windsurf's accept/reject data flywheel

Deng cites Windsurf, which built on Claude 3.5 and captured unique data on which code snippet recommendations users accept and reject, then launched its own model based on that data.

That usage-generated data became a compounding edge that a company without unique starting data could build.

Craft beating Microsoft's distribution

Deng notes Copilot was far ahead with Microsoft's distribution, talent and first-mover advantage, yet Cursor, Windsurf, Lovable and Bolt broke through — and Granola wins against Google Meet/Teams/Zoom distribution — purely on product craft and delightful, sanded-down edges.

Users switch and evangelize products with superior craft despite incumbents' distribution advantages.

Common mistakes

Treating the model as a magic wand

Founders assume AI will just work; but models only excel at what they're trained on, so without a proprietary data flywheel there's nothing defensible and foundational models can eat your lunch.

Competing on a marginally-better product against incumbents

Incumbents' distribution advantages mean a product must be dramatically better to earn the switch; a small improvement won't overcome their distribution.

Is it for you?

Best for

Founders and investors building or evaluating applications on top of LLMs who need durable defensibility

Not ideal for

Undifferentiated wrapper products with no path to proprietary data or a distinctive workflow

From the transcript

having the right data and the right data flywheels is so important like proprietary data

27:30

it's the it's the ergonomics of how does it actually integrate into people's lives

28:00

the the the models will get really good at whatever data you show it

29:00

you have to kind of overcome a pretty high bar of your product has to be so much better

29:00

From the episode

From ChatGPT to Instagram to Uber: The quiet architect behind the world’s most popular products

Peter Deng