Actionable insights — Intelligence as Infrastructure

5:05 1. The penetration-gap screen — measure how early you really are

1. Don't trust the headline user count; estimate the advanced-use penetration — the share actually using the technology the powerful way (here: agentic AI, not "search on steroids").
2. Quantify the gap two ways: depth (only ~10 bips of ~1B white-collar workers use it advanced) and a leading proxy (Claude Code at ~14M DAUs heading toward ~500M).
3. Check the intensity multiplier: the few real users "burn a thousand times as much compute" — so revenue/compute demand scales with depth-of-use, not just headcount.
4. If penetration is sub-1% and rising vertically, treat it as an "L-curve" (straight up) and underwrite years out, not quarters — "we have half of what we need" is the buy signal, not the risk.

• Any technology where headline adoption is huge but advanced use is a fraction of a percent and accelerating; DAU and tokens-per-user as the leading proxies.

7:13 2. Map the new stack, own the value-capture layers

1. For a new compute paradigm, draw the full stack: chips → clouds → foundational models → applications.
2. Ask where value accrues and is defensible, not where the excitement is. Sacerdote's answer: the foundational-model layer (oligopoly economics) and the chip layer (undersupply) — and explicitly not the horizontal application layer.
3. At the model layer, confirm the economics are oligopoly-shaped: a few players who locked up compute early earn enormous incremental margin as token pricing rises on a fixed-cost base (≈18× earnings, not cash-burning).

• Which layer has the scarcity (compute) and the oligopoly; avoid the layer where the moat question is unanswered (apps).

13:57 3. The decommoditization screen — find the part whose spec is now exploding

1. List the components of the old commodity system (the ~$2,000 x86 server: networking, PCB, power, cooling) — for 40 years all commoditized, made by 20-30 firms.
2. Find the parts whose required spec is now rising vertically under AI load: PCB layer counts (10 → 20/30/40/120), network speeds (1 → 400 → 800 → 1,600 → 3,200 Gbps), power draw per rack (+50-125%).
3. Where the spec leap collapses the supplier set to "two or three companies that can do it properly," you've found a decommoditized pinch point — IP-rich, higher-margin, with pricing power.
4. Validate the earnings algorithm: units +50%, ASPs +20-100%, gross margin +300-500 bps, 3-4-year visibility, everything in short supply → earnings can ~double for years.

• Any once-commodity component whose required performance is jumping each generation and whose qualified-supplier list is shrinking.

16:46 4. The demand-vs-efficiency check — does usage outrun the cost curve?

1. Before assuming efficiency kills the compute trade, put the two rates side by side: demand growth vs efficiency gains.
2. Here tokens (the unit of AI compute) grow ~14× a year while chips improve ~100-200% and the estate stretches ~2-3× more from software efficiency — so demand still outpaces efficiency several-fold.
3. Conclusion rule: while the demand multiple exceeds the combined efficiency multiple, the shortage persists and the hardware bet holds.

• The token-growth-rate vs chip-improvement-rate spread — the rotation signal would be efficiency catching demand, not headlines about a cheaper model.

18:54 5. Bifurcate "software" — and use the "are the labs your customer?" tell

1. Never trade "software" as a block. Split it: horizontal application software (in trouble) vs data-driven / infrastructure software (can win).
2. For the at-risk bucket, stack the headwinds: it dropped below AI on the CIO list; token spend eats its budget; pricing power and seat-based models are under threat; vendors' own AI products have been "a fail" (a culture/sales-model problem — selling a service, not software); and "headless" build-it-yourself risk.
3. For survivors, apply the tell: are the AI labs themselves your customers? If the companies driving the boom run on your tool, that's a durable demand signal (debate the multiple separately).

• Infra/observability/data names whose customer logos include the frontier AI labs; horizontal SaaS where AI ARR is <2% of revenue.