What Actually Makes A Startup Durable

Cost & durability of AI-native startups

AI cost economics will self-correct quickly

• Expect ~10x cost reduction per year for “the same intelligence”
• Token/inference costs should decline significantly over the next 6–12 months

AI use beats human labor for programming tasks

• Claim: an engineer with the right model can be ~1,000x better than an engineer without AI
• Implication: it’s increasingly hard to justify “wait for engineers” versus using the best available models

Startup durability requires a “hard bit,” not just AI-enabled software

• As software becomes cheaper/easier to replicate, durable differentiation shifts toward components that are genuinely difficult:
  • Hard-to-execute B2B sales into specific industries
  • Regulatory barriers
    • Example: a banking license near-killed a startup
  • Hard physics / hardware
    • Example: spaceship launch; hardware durability comes from “atoms”/manufacturing complexity
• Avoid thin, easily replicable pure-software ideas
  • Examples framed as likely non-durable if started today: Calendly and DocuSign (“too easy to replicate”)

Durability playbook (implied framework)

• Identify where your startup sits on a spectrum of very easy ↔ very hard
• Choose positioning where your “hard bit” is:
  • Distribution or sales wedge
  • Regulation/approval barrier
  • Deep technical/hardware/complex operations
  • (Often) something competitors can’t copy quickly

---

Building startup communities (especially outside SF)

Prefer joining an existing founder community

• Optimize for joining an existing founder community rather than building from scratch

Go to other hubs with critical mass if you can’t go to SF

• Examples: London or Paris
• Rationale:
  • Being “the only one in a city” reduces peer learning and slows progress

Example / case: London ecosystem traceable lineage

• Example: GoCardless (founded 2011) helped seed a network connecting to TransferWise and Monzo, plus many other companies
• Used as evidence that ecosystems propagate

---

What AI should (and shouldn’t) replace for founders

Judgment delegation boundary

• Don’t “outsource thinking” to AI outputs blindly
• Strategy:
  • Use AI, but maintain an explicit loop where you don’t accept unsatisfactory outputs
  • Iterate until quality meets your bar

Human-centered “not replaceable by AI” categories

• Founder well-being
• Community
• Knowledge sharing via dense peer context
  • Example: YC having ~200 companies obsessing over AI execution
  • Not “content,” but an environment
• “Power of witnessing”
  • Example: virtual office hours / human acknowledgment of what founders are going through

AI-as-advisor guideline (execution principle)

• Use AI as a tool for iteration, but keep:
  • Accountability to your own standards
  • Direct market contact as the truth source

---

“AI-native” operating system: loops, automation scope, and process design

Key success metric for AI process use

• Don’t measure only:
  • Correctness
  • “Reading success/error”
• Measure effectiveness of a self-learning feedback loop that improves outputs over time
• Use memory/harness that captures what you learn from interactions

Avoid the mistake: “use AI, look at result, move on”

• Instead:
  • Iterate until the result changes your plan/behavior
  • Feed back learnings

Practical implementation approach (playbook)

• Start with one narrow loop, not full automation
• Make company knowledge legible/queryable
  • Examples mentioned: emails, Google Drive, support tickets
  • Tools mentioned: Obsidian, GBrain
• Example loops:
  • After each sales call → send follow-up email capturing key points
  • During each sales call → transcribe + have an agent build a prototype so you can demo by end of call
• Last thing to automate:
  • Talking to customers
  • Also suggested: founders shouldn’t delegate it to even a co-founder for core context reasons

---

Research vs business scaling: choose “learning with the market” fastest

Guiding rule

• Confront the market in the fastest possible way

YC regret emphasized

• Not launching soon enough

Operating cycle

• Build ↔ talk to customers (and repeat)
• If users don’t want it → return to building/research

Bias check

• Many founders prefer building/research (comfort zone)
• Recommendation: increase customer conversations to counter that bias

---

Pivoting in the AI era

Lower cost of “writing software” doesn’t mean cheaper pivoting

• Pivoting still often means discarding meaningful work (possibly including the hard bits)

Updated pivot logic (framework-style)

• Pivot when you’ve exhausted evidence against your fundamental hypothesis
• Many “pivots” are really:
  • Loss of enthusiasm
  • Avoidable rejection pain
  • Not a true strategic pivot

Better pivot pattern (example)

• GoCardless-style lesson:
  • The original problem (collect payments for sports teams) wasn’t strong enough
  • The underlying enabling capability led to a broader, better use case

---

Org design: co-founders, agency, and team size effects

AI may compress org sizes, but co-founder value remains

• Coordination cost rises with team size (described via a Metcalfe’s-law-like idea: connections scale with nodes)
• Claim: AI could enable companies to stay under Dunbar’s number (~150 people)

Solo founders

• YC funds solo founders, but statistically they do worse; the bar is higher
• Example scenario:
  • A strong solo founder became sad/confused about direction
  • Co-founder role described as emotional stabilization (“pick you out of the gutter”)

Picking a co-founder (practical checklist)

• Choose someone smart/determined/hardworking
• Add an integrity layer:
  • Do you trust them?
  • Are values aligned?
  • Would you be scared to be on the other team?
• Skill complementarity was downplayed:
  • Business skills are “easy to learn,” while trust/alignment matters more

High-agency (personal/muscle)

• Defined as belief that your actions produce output/impact
• Developed via “hard but tractable” projects
• Often starts with a co-founder/partner

---

Venture capital, capital efficiency, and durability tradeoffs

Seed → VC rounds still matter, but less capital may be needed

• AI-assisted speed may reduce capital requirements versus ~5 years ago

The tradeoff: cheaper building can mean less defensibility

• AI reduces the cost to build similar software
• That reduces durability/defensibility

Therefore

• AI enables founders to attack harder problems
• Harder problems often need more capital, shifting why VC still matters

Analogies/examples

• Small nuclear reactor: requires ~$800M next year (illustrative)
• Banking license: regulated businesses need large capital to build/operate

---

Evaluating “real wedge” vs “AI wrapper” (AI-native marketing risk)

“AI wrapper / AI rapper” framing

• Used as a derogatory term for thin layers on top of generic models

YC evaluation emphasis

• Whether the product is more than “powered by AI”
• Team ability to execute evolution
• Depth of problem space
  • Even if someone goes “deep,” it still matters if the wedge is real

Value-creation routes (explicit)

• Technical wedge
  • Improve the state-of-the-art and create a real technical improvement → margin
• Distribution wedge
  • Sell better / package capability → change outcome ratio

Example evidence

• YC funds both:
  • Deep technical founders (e.g., nuclear physics, satellite re-entry vehicles, RF testing expertise)
  • Also “self-taught vibe coders”
• Suggests the bar may broaden, but “systems thinker + smart/determined” remains important

---

OpenAI / US export restrictions (frontier AI models)

• High-level only: export restrictions are seen as a wake-up call about dependency on US government decisions
• Expected market impact (high level):
  • Increased demand for competitors and “sovereignty” over AI and inference

---

Student builders & AI credits

• Students can get $25K+ in credits
  • Via YC context and sponsor model labs/hyperscalers mentioned
• Recommendation: Go build something cool
  • Credits reduce model cost friction

---

Batch execution habits & YC “user manual” updates (process)

Top trait: launch early and often

• Too much theoretical thinking and fear of rejection delays startup progress

Startup as an empirical loop

• Hypothesis → test with real world data → update approach

Rhythm

• Ambitious goals, reassessed every two weeks
• Each cycle targets the current biggest bottleneck, not dozens of small fires

Batch problem

• Many founders learned “thinking in a library” to win grades
• That approach translates poorly to startup success

---

Founder “early decision” (selection signal)

• Early decision applicants must clearly explain why they need to decide now
• YC reads that answer; the bar is effectively higher if the reason isn’t compelling
• Implied legitimate reason example: inability to afford waiting (financial/job constraint)

---

Key metrics / targets explicitly mentioned

• AI cost trend: ~10x cost reduction per year, expecting improvement within 6–12 months
• Engineering impact claim: ~1,000x better for engineers using the best AI models
• Org size heuristic: keep relationships below Dunbar’s number (~150) (via AI compressing coordination needs)
• Loop cadence / accountability: reassess goals every 2 weeks
• Student credit target: $25K+ credits
• Nuclear reactor capital example: ~$800M next year (illustrative)
• No explicit revenue/CAC/LTV/churn targets were provided

---

Presenters / sources mentioned

• Tom (main YC partner/presenter referenced repeatedly; appears to be Tom Blomfield)
• Matild(e) (partner mentioned)
• Gary Tan (referenced via an “excellent post” about the “power of witnessing”)
• Nikico (mentioned regarding an AI-enabled virtual office hour story)
• Gray (contributor/partner; referenced via cofounder-style commentary)
• Sam Altman (mentioned regarding “one person billion dollar company” sentiment)
• YC (Y Combinator)

Organizations / companies (examples)

• GoCardless
• TransferWise
• Monzo
• Brex
• Cursor founder (referenced via a tweet about launching on Hacker News)