Ask HN OpenAI 100 Companies: Build What They Won’t

What you’ll walk away with

You’ll leave with a niche filter OpenAI is unlikely to prioritize, a cost stack tied to live API token prices, and a weekend build plan. The viral Ask HN: If the models are so good, why doesn’t OpenAI just start 100 companies? thread stops being a debate and becomes a shipping checklist.

Hundreds of points. Hundreds of comments. Everyone picked a side. You’re going to use the answer instead.

The real constraint isn’t model IQ

ex-aws-dude’s post (HN item 50200997) asked why frontier labs still meter intelligence when app-layer profits look juicier. Strip the title hype and the constraint is simpler: models run bounded work. They don’t carry refunds, sales calls, App Review, or a P&L.

A power plant sells kilowatt-hours. It doesn’t open a hundred bakeries because ovens need electricity. Burned batch? The plant still invoices for the power. That mismatch – capability without responsibility – is why “100 companies” stays a thought experiment.

Waiting for the lab to ship your vertical

Two clocks punish you. Domain data, escalation paths, and buyer trust don’t fall out of a general agent. And strategy from inside the lab already drew a line.

Founders should chase specialized applications off the critical path to AGI – that was Jason Kwon (OpenAI chief strategy officer) at the Madrona IA Summit, per GeekWire’s write-up. On the path, the lab cares. Off it, much less. That’s the gap you use.

Unit economics agree. OpenAI’s API platform (rates as listed for standard contexts under 272K, as of late 2026) meters brains by the million tokens: GPT-6 Luna $0.10 / $0.50 input/output, GPT-6.1 Sol $2 / $10, GPT-6 Astra $10 / $50. You’re renting inference. Not buying a finished company.

5-step playbook: not their 100 companies

Do these in order.

1. Run the Kwon filter

1. One sentence: who pays, what outcome, how often.
2. Does solving this push general intelligence forward, or does it only matter inside one industry’s messy workflow?
3. Keep manufacturing QC checklists, local-regulation filings, niche B2B reconciliation, weird compliance loops. Kill generic “AI that does everything in chat.”
4. Name the human who still owns refunds and “this went wrong.”

Half of thin wrappers die here – before you open an editor.

2. Price the brain, not the dream

Map one happy path to tokens before any landing page.

Tokens per successful job × monthly volume. If that sits above ~30% of what buyers pay for the outcome, redesign first: cache, smaller model, fewer tool loops. UI can wait.

Edge trap: outcome pricing on your side (per accepted letter, closed ticket) while you pay per token is a margin guillotine. A bad week of retries or fat contexts can erase profit even when the product “works.” Token spend is internal COGS you review weekly – never the customer price.

3. Distribution and taste before clever code

Code got cheap for everyone. Thread reactions keep circling the same unpaid labor: picking something people actually buy, showing up where those buyers already hang out, answering support at 2 a.m., eating chargebacks. Email list. Niche onboarding language. A brand that already means “we went first.” Boring stack beats another wrapper.

4. One vertical workflow – not a platform

One buyer. One measurable outcome. Weekend skeleton:

system = """You are a senior ops assistant for independent dental clinics.
Only answer using the clinic's uploaded fee schedule and state board rules.
If uncertain, return {"status":"escalate","reason":...} - never invent codes."""

user_job = {
 "task": "draft insurance pre-auth letter",
 "patient_notes": notes,
 "fee_schedule_id": "clinic_17",
 "must_cite": ["fee_schedule", "state_rule"]
}
# Sol (or Luna draft → Sol final). Log tokens.
# Bill per accepted letter - not per token.

You own fee-schedule import, escalate path, and the clinic relationship. The model owns prose.

Absorption risk: horizontal chat wrappers on generic workflows are easy for a lab to fold into agent/shopping layers later. Vertical workflow + proprietary data + human escalation is a worse swallow. Kwon’s framing again: off the AGI-critical path, less interesting to them.

5. Clone defense on day one

Assume a stranger rebuilds your UI in a weekend once the demo is public. Claim the name first. Post the origin story. Collect waitlist emails. Make switching real – saved history, custom templates, billing already wired. Narrative speed beats code speed. Lock distribution before any launch thread goes loud, not after.

What the token business actually proves

Press summaries put OpenAI’s investor-basis annualized run rate near $50B by late September 2026. That scales. It still doesn’t mean the lab wants your dental-letter niche.

Indie arcs after the same Ask HN energy look repetitive for a reason: model drafts the app fast → launch brings first cash → copies show up within days → survivors already owned distribution and support. The model never picked the idea, never ran the launch, never sat in the queue.

• Log token cost per successful outcome from day one. Retries spike? Fix prompts or retrieval before you touch price.
• Keep a manual override visible. Buyers purchase responsibility theater as much as accuracy.
• Conviction, not feature ideas: the original Ask HN thread.
• Where the lab says it won’t go deep: GeekWire’s Kwon summit coverage.

Will application surface get absorbed over time? Maybe. Docs and commentary still won’t pin when a lab accepts outcome liability at scale – cash flow, regulation, refunds – for open-ended businesses. That missing line is why “100 companies” stays theoretical. Until it moves, sit on the side that still needs a human signature.

FAQ

Is the Ask HN point that OpenAI is “dumb” for selling tokens?

No. Token sales scale without owning every vertical’s support and legal risk. Rational business, not a blunder.

What should I build this weekend so OpenAI won’t crush it?

Vertical job. Proprietary inputs – the customer’s docs, fee schedules, site SOPs. Wire Sol or Luna to one outcome. Escalate-to-human path required. Launch to 20 people who already know you. Concrete: auto-draft a compliance email from a template pack only your buyers use. Not another general writing app.

Don’t models already match experts on lots of tasks?

Strong execution on defined work shows up in benchmarks and enterprise write-ups. The thread’s real gap is agency under uncertainty: who places the bet, who eats the loss, who faces the angry customer. Capability without responsibility is still a utility. Your product wraps that utility in responsibility someone will pay for – and that’s the company they aren’t starting for you.

Next action: Write the one-sentence job + Kwon filter answer now. If it survives, open the OpenAI pricing docs, estimate tokens for one happy path, and block three hours this weekend for the escalate-aware vertical – not chatbot number 100.