If you’re building or about to build an AI company — this folds the book into founding decisions: which layer to build at, how to dig a moat, which are the ways to die, when to sell to the platform. The core is one line and a day-one checklist. No company names — positions and mechanisms.

The book’s judgment for founders compresses to four words: don’t die in the middle.

The middle is the bare model layer — a qualifying round with an entry fee in the billions, rising each generation, that even model companies might not win. The middle is also the thin wrapper — a layer of prompts over someone’s API, squeezed dead from above (integration), below (deflation), and the side (routing). These two “middles” are the most crowded, highest-mortality spots in AI startups — and exactly where most people pile in, because they’re the easiest to launch and the most dazzling to demo.

The first principle of positioning: leave the middle, hug an end — either the scarcity end (upstream’s seams) or deep into the lock-in end (downstream’s workflows).

The Six Questions for Day One

String the book’s criteria into six questions. Can’t answer any one — you’re not ready to start:

Q1: Which layer am I building at? Not the middle. Hug the scarcity end or burrow into the lock-in end. If the answer is “bare model / thin wrapper,” stop, change position.

Q2: What’s my scenario’s referee? — pick the referee before the scenario. In your scenario, is the right/wrong signal (ground truth) free or paid? Instant or lagged? Is the step you accelerate the real end-to-end bottleneck? The clearer, faster, freer the referee, the realer your data flywheel and the lower your reliability tax. This matters a hundred times more than “which model.”

Q3: Where does my thickness come from? Four kinds — data loop, workflow embedding, compliance barrier, systems integration — own at least one, ideally two stacked. None comes from “the model itself.” If your thickness is all “our model/prompts are better,” that’s not thickness — it’s rented, and it evaporates next year.

Q4: Does the agent inequality hold? Outcome value > token cost + reliability tax (error cost × error rate). If not, your per-outcome model loses money at scale. And make “racing the upstream deflation” a core metric: your margin isn’t set by your pricing, it’s set by how fast token cost falls.

Q5: If the model company natively supports me tomorrow, what’s left? — feature vs company. Left with a deeply embedded workflow, a unique data loop, an unscalable compliance wall — you’re a company. Left with nothing but prompts and a UI — you’re a feature. The sharpest question; it predicts your three-year survival.

Q6: Who holds my distribution entry? Your intent/procurement entry — how users find you, how agents route to you — do you control it, or does the layer above (OS, super-app, aggregation platform) hold it and could intercept anytime? Q1–5 decide how thick you build; this one decides whether users can reach what you built. However thick the moat, it’s moot if the layer above chokes the entry.

Pass all six and you stand on a structured position. Fail any one and you may be building something the trend will roll over.

Four Ways to Die

1. Die in the middle: bare model or thin wrapper — squeezed from three sides. Most common, because launch is easiest.
2. Do the giants’ free market research: fast growth, beautifully proving a demand, thin moat — you’re using your own money and speed to do the model company’s most efficient market research; once demand is proven big enough, native integration arrives to harvest you. Fast growth here is a danger signal, not a victory.
3. Bet the moat on something regulation can zero out: all lock-in built on “high switching cost,” which a single mandated-interoperability regulation can erase. Either stack multiple thickness sources, or lean toward a “compliance-barrier moat.”
4. Require customers to restructure their org first: your value only pays off after the customer changes process, authority, headcount — you die in procurement. The real entry is the one that routes around the org slow-variable.

Exit: Sell the Feature, Hold the Company

Not every AI startup should grind to IPO. If your Q5 answer is “native support leaves me nothing,” you’re probably a feature on someone’s roadmap. For a feature, the best outcome is usually not grinding until integration zeroes you out, but selling for a good price in the window where demand is proven and native integration hasn’t landed yet. But “sell” has preconditions: a buyer exists, buying beats building, the window’s still open. Seeing this isn’t shameful — what’s shameful is mistaking a feature for a company and grinding to zero. If you pass all six and are truly a company — invert it: burrow deep, stack thickness from multiple sources, use the window to dig the moat past where an acquirer can afford or route around.

A last line on riding the trends. Big trends are double-edged for founders; what matters is which edge you stand on: deflation is your engine (driving down token cost) and your nemesis (flattening your capability edge) — eat the cost dividend fully, but never build the moat on “our model is better”; open source both depresses your capability premium and is your weapon against the giants — use it as a tool, not as a rival; the real rival is the layer closer to the user.

The migration map gives investors allocation; it gives founders positioning: silicon→watts, subscription→rent, attention→routing — each arrow’s endpoint is a position that will thicken and isn’t crowded yet. Don’t add one more corpse to a sea that’s already gone red; go to the spot the map marks “deepening, not yet crowded” — quiet now, deep water in three years.

Your company — how many of the six can it pass? Which one stumps you most? Comments open.

Sources (framework argument, June 2026): the six questions and four ways to die are the founder form of the book’s criteria; mechanics appear in Chapters 6 (bare-model qualifier), 7/9/10 (thickness / agent inequality / thin wrapper), 11 (organizational physics), 14 (mandated interoperability).

— From Chapter 16 of a book in progress, working title The Deflation Sandwich

#AI #Startups #DeflationSandwich