---
id: ins_altman-ai-not-a-business-exception
operator: Sam Altman
operator_role: CEO OpenAI; former president Y Combinator; investor; essayist on startups and frontier tech
source_url: https://blog.samaltman.com/
source_type: essay
source_title: "Startup Playbook — AI Companies Aren't Exempt"
source_date: 2015-09-08
captured_date: 2026-05-05
domain: [strategy, gtm, ai-native]
lifecycle: [strategy-bets, fundraising, positioning]
maturity: applied
artifact_class: framework
score: { originality: 4, specificity: 5, evidence: 4, transferability: 5, source: 5 }
tier: A
related: [ins_intelligence-cost-converging-electricity, ins_altman-singularity-as-smooth-curve, ins_thiel-poor-sales-not-bad-product]
raw_ref: raw/expert-content/experts/sam-altman.md
---

# "We're using AI" is not a business strategy, defensibility comes from domain expertise, customer relationships, and data, not from the model layer

## Claim
The claim "the normal rules of business don't apply to me because I'm using the latest technology" is never true. Using AI itself does not create a defensible business; defensibility comes from non-AI moats, domain expertise, customer relationships, proprietary data, distribution. AI is the new substrate everyone has access to; what differentiates winners is the *non-AI* layer they wrap around the model.

## Mechanism
Foundation models are commoditising rapidly (Altman's own thesis: intelligence cost converges to electricity cost). Any defensibility claim that rests on "we have AI" or "we use the latest model" is therefore a non-claim in 12-24 months. The buyer can switch to a competitor with the same model substrate; the model itself is not a moat. The actual moats, and they are the same ones that have always worked, are: deep understanding of a specific domain that the AI substrate doesn't substitute for; customer relationships and trust that compound over years; proprietary data that the foundation model doesn't have; distribution channels that competitors can't easily replicate. Altman's warning is to AI startups specifically: do the boring strategic work of building moats; don't assume the AI will substitute for it.

## Conditions
Holds when:
- The startup is building on top of foundation models (most AI applications today).
- The category will see rapid foundation-model commoditisation (most current categories).
- The non-AI moats are buildable (most B2B SaaS, most vertical AI, most agent products).

Fails when:
- The startup is itself building a foundation model with credible defensibility (training data advantage, scale advantage, scientific breakthrough).
- The category has structural network effects that compound regardless of model substrate (some marketplaces, some social products).
- Pure short-horizon plays where defensibility doesn't matter (some content / tooling categories with rapid turnover).

## Evidence
> "the normal rules of business don't apply to me because I'm using the latest technology" is never true.

· see `raw/expert-content/experts/sam-altman.md` line 15.

## Signals
- Strategy decks lead with the non-AI moat (data, distribution, expertise) before mentioning the AI substrate.
- Product roadmap invests in defensibility-building work (proprietary data collection, customer-relationship deepening, vertical expertise) alongside model integration.
- Investor conversations don't depend on "we use GPT-5 / Claude 5" as the differentiation argument.

## Counter-evidence
For a brief window in any new technology wave, "we use the latest technology" *can* be a real differentiator, first-mover advantages compound while competitors catch up. The window is shorter than founders think, but not zero. The Altman claim is the right default; first-mover plays are the exception that requires specific evidence.

## Cross-references
- `ins_intelligence-cost-converging-electricity`, the long-run trajectory that makes AI-as-moat fail.
- `ins_thiel-poor-sales-not-bad-product`, Thiel's adjacent claim: distribution beats product, even AI product.
- `ins_specific-knowledge-cannot-be-mass-trained`, Naval's claim on what's actually defensible.