New-grad unemployment (22–25), summer 2026. Dead centre of the 6.3–7.8% range of the prior four summers. "No AI hiring crisis."
The Founder
Locks the Electrons
A folio in nine plates · On the moat that quietly moved
Read on your phone. Swipe up. Each plate snaps.
The AI moat is
not the model.
It is the 23-year
electricity contract.
The model layer just became a commodity.
The moat moved down the stack —
to the boring, uncopyable inputs.
Google buys the reactor.
Sep 10, 2026 — WSJ. Google signed a PPA for half of an Ontario nuclear plant's output through 2049. It converts the single least-predictable line item in hyperscale computing — wholesale electricity — into a fixed, knowable cost for the next twenty-three years.
Read the basement, not the blog post: while every founder is chasing GPT-6, Google is chasing kilowatt-hours priced in 2049 dollars. That is what the annual report will not say out loud, but the treasurer knows.
The model became bulk goods.
Same month. OpenAI and Anthropic collapsed frontier inference pricing by up to eighty percent. Institutional investors began muttering the phrase "AI capex fatigue" as hyperscaler buildout crossed US$700B in committed spend.
When your input price falls 80% in a quarter, you are not selling a moat. You are selling grain. The model is not where value accrues any more — the model is what you buy at the wholesaler, not who you become.
Project Jupiter,
New Mexico.
Sep 24, 2026 — Bloomberg. Oracle sent a force majeure notice on Project Jupiter, its New Mexico campus, to defer payments if the site fails to come online in 2028. Per the Financial Times (18 Sep), the roughly $18B in construction loans were already quoted at 89 to 91 cents on the dollar.
The transferable lesson for Steve and Otto: the buildout that skips the boring input contract is the first to be repriced by the market. Ambition without a PPA is a coupon someone else gets to cut.
Two truths, one month.
Ad agencies are stopping graduate hires because AI has already taken the entry-level work — and the execs are saying it on the record.
The average hides the industry. Both are true. Cross-reference — always.
Good at capacity.
Bad at electrons & liability.
FTE-equivalent capacity Wipro absorbed internally with AI — with zero direct layoffs. The capacity was recycled into higher-margin engineering. The head did not shrink; the tentacles grew.
California's AB 316 (in force 1 Jan 2026) killed the "the agent did it" defence. Nippon Life v. OpenAI (Mar 4) tests it for ~$10M+. Cheap ≠ safe. Cheap ≠ liable-proof.
What is your PPA?
The nuclear contract is the metaphor. It is not the plan. The plan is: find the one input in your business a competitor cannot reprice against you next quarter — and lock it for years.
Hyro FMCG · the US-raise pitch
- Multi-year contracted cost of goods with your co-packer.
- Category-exclusive shelf agreements (chain by chain).
- Locked freight lanes at contract, not spot.
The AI of running FMCG is worth more than the drink itself — only if the drink itself sits on uncopyable contracts.
Home Loan Experts / Alaya
- Multi-year lender panel commitments (your cost of capital).
- Retained talent bonded to a 3-year outcome, not a title.
- Tech-stack unit cost fixed via committed volume.
Copyable in a weekend: your model. Uncopyable at any price: your contracts.
Price one
uncopyable input.
Sign it for years.
Not a demo. Not a model swap. Not another pilot.
A contract with a term measured in years, not sprints.
— Steve
“Read the basement, not the blog post.”
Built by Hermes · Sources cross-referenced · 30 Sept 2026