Ask HN: If the models are so good, why doesn't OpenAI just start 100 companies?

Hacker News Discussion ↗

If large language models are so powerful, why aren’t their creators just using them to launch and run dozens of highly profitable companies themselves? Commenters argue that AI labs are following a classic “sell shovels in a gold rush” strategy instead: focusing on core models, chips, and token sales while letting others take the risk of building domain-specific businesses, then potentially copying or acquiring what works. Many also stress that today’s models still lack the long-term judgment, social agency, and operational execution needed to autonomously run complex firms, and that anything simple enough for an LLM to do easily is likely too commoditized to sustain a strong moat.

Shovels vs. Starting Companies

  • Dominant metaphor: frontier labs are “selling shovels in a gold rush” (models, tokens, silicon, cloud capacity) rather than digging for gold (running 100+ end-user businesses).
  • Argument: shovel-selling scales better, spreads risk to customers, and avoids competing directly with those customers.
  • Counterpoint: if the shovels are really that good, why not use them to build a perpetual money machine (e.g., a quant fund or fully automated businesses)?

Model Capabilities and Limits

  • Many argue the models are powerful but far from able to autonomously run complex companies: weak long-term autonomy, judgment, creativity, and social/relationship skills.
  • They’re seen as great at code, math, and bounded tasks, but not at deciding what to build, getting customers, or maintaining key accounts.
  • Some see this as moving goalposts from “AGI ~ human-level on many tasks” to “AGI must run entire businesses without humans.”

Economics, Capital, and Risk

  • Labs need massive, fast-growing revenue to justify huge investment and tight capital markets; starting 100 risky businesses is seen as a distraction.
  • Debate over “investment vs. debt”: legally distinct, but both create intense pressure for revenue and returns.
  • Some invoke Jevons paradox: cheaper intelligence → much more demand for it, so selling tokens/agents is the surest growth path.

Platform Strategy and Data Advantage

  • Consensus that labs prefer platform roles: APIs, agents, app layers, “forward deployed engineers,” and targeted product spin‑offs rather than many standalone companies.
  • Selling tokens lets them observe usage patterns and, with consent, train on customer data; later they can clone or acquire the most successful ideas.

Competition, Moats, and Antitrust Concerns

  • Running 100 companies would pit labs against their own customers (risk of being “Sherlocked”), discouraging adoption and inviting regulators.
  • Some expect eventual vertical integration and spinoffs (consulting, finance, drug discovery, specialized SaaS), but only selectively and when models are stronger.

Impacts on Existing Industries

  • Claims that certain niches (e.g., tutoring/homework help, stock photos, some translation and low-end SaaS/tools) are already badly hit; others dispute how widespread this is.
  • General view: AI makes building software cheaper, so moats shift to domain knowledge, relationships, distribution, and capital-intensive or physical operations.

Future Scenarios

  • Some think labs will ultimately automate and absorb many software markets; others think model commoditization and competition (including abroad) will cap their margins.
  • Timing and feasibility of a fully AI-run company are widely seen as “unclear.”