Actionable insights — The AI Revolution: Who Wins, Who Loses & the SaaSpocalypse

33:32 1. Hunt dislocations: find two stocks in one sector whose valuations imply opposite futures

1. Pick a sector the market is grading emotionally — one where "on a daily basis" it is "deciding between AI is good and AI is bad." Volatility of sentiment is what creates the mispricing.
2. For each name, back out the implied duration of the cycle rather than judging the multiple in isolation. Ask: how many more years of this cycle does this price require in order to be justified?
3. Sort the sector by that implied duration. A dislocation exists when two names in the same cycle sit at opposite ends: "the market is now valuing some stocks… as if this cycle is continuing through 2030… [others'] valuation implies that the cycle is already over. That is inconsistent."
4. Now add the quality overlay and check whether the ranking agrees with the pricing. Historically "the CPU market's been a little better than memory" — if today the reverse is true, the cheap name is also the better business, and the inconsistency doubles.
5. Take the side where the price already assumes the bad outcome. You are not forecasting the cycle; you are being paid whichever way it resolves, because one of the two prices has to move.

• Two names in one supply chain whose multiples imply different end-dates; a low-multiple name where the low multiple is doing the work of a forecast; the historical quality ranking of two sub-markets quietly inverting while the relative multiples don't.

29:10 2. Bandwagon aversion — treat unanimous positioning as the signal, then find the one fact everyone skipped

1. Track who owns the view, not just the view. "I'm always worried when everybody's on a bandwagon. When I see every last hedge fund, every last long-only on a bandwagon, it makes me uneasy. And then I try not to be on that bandwagon."
2. When a single disclosure re-rates a stock in one session, decompose the number before accepting it. A backlog that jumps 150→450 "in a day" is not a demand trend; it is one contract.
3. Ask the credit question about the counterparty: can whoever signed it actually pay? On Sep 11, 2025 the answer was that OpenAI "did not have money… very little revenue" behind a $300B commitment plus $1.1T more elsewhere.
4. Then run the mirror discipline on the way down. When the price implies the opposite extreme — "the market had reacted too much to the other side because it was saying that OpenAI revenue is worthless" — re-underwrite the counterparty with the new facts (the $122B raise, the "code red" narrowing to compute) instead of staying anchored to your own famous call.

• A metric that re-rates a stock in a single session; backlog or RPO disclosures with no customer breakdown; a counterparty whose funding is announced later than its commitments; your own prior call becoming the reason you can't change sides.

31:14 3. Price the asymmetry: ask what the market is currently valuing at zero

1. Split the company into the established business and the contested one (here: legacy Oracle vs the AI compute backlog).
2. Value the established business on its own multiple, subtract from the market cap, and see what is left over for the contested piece. If the residual is zero or negative, the market has already written the contested asset off.
3. Ask what has to be true for that write-off to be correct — usually one binary (can the customer pay?) — and underwrite that question rather than the whole company.
4. Buy only where the residual is at or below zero, so that being wrong on the binary costs little and being right is uncapped: "clearly it's at least zero."

• A segment whose contracted revenue exceeds the whole market cap; sum-of-the-parts residuals at or below zero; concentration inside the written-off asset (one customer, one contract) that would make the write-off correct.

41:05 4. Run the CIO's budget: AI spend is funded by cancellations, so rank software by "who gets consolidated into"

1. Start from the constraint, not the technology: the IT budget isn't growing as fast as the new AI line item, so "companies have to spend so much on AI right now that they're looking at their budget and saying where can I cut."
2. Put yourself in the buyer's seat and do the arithmetic literally — "if I'm a CIO and I have a hundred software packages I'm managing… I need to go to 30. Guess what I'm cutting?"
3. Sort every software holding into consolidator or consolidated. Consolidators are large platforms that can absorb several jobs and will bundle ("they'll probably bundle because they're increasing my price anyway"). The consolidated are small, single-purpose, or — decisively — not adding value: "Salesforce has not been adding value to them in years and it keeps charging them more and more for that less value every year."
4. Weight the value-per-dollar test above the AI-disruption story. Luria's case against Salesforce works "regardless of AI"; AI only supplies the forcing event.
5. Check the ownership tier separately. Private-equity-owned software is the most exposed because the model is to "buy the software company, milk it" — no product investment, and small scale. "Those companies are going to be gone."

• Seat-count or module churn in a vendor's disclosures; price increases substituting for growth; renewals slipping in small-cap or PE-owned vendors; a customer's AI capex line rising while its total IT budget is flat; a big platform's attach-rate rising as small vendors disappear.

41:54 5. Before extending a sector's credit stress to its listed names, check the balance sheets — "software debt is private equity"

1. When a sector's debt is in the news, first ask who actually borrowed. Distress headlines about "software debt" were being read as a public-equity problem.
2. Check the net cash position of the listed cohort: "the software companies Dan and I cover are in a net cash position. They don't borrow money. There is no software debt for them."
3. Locate the real borrower — here, the leveraged buyout structures. "Software debt is private equity. Bought the software company and levered it up."
4. Then trade the mismatch in both directions: the listed names get sold on a risk they don't carry (opportunity), while the lenders and fund holders carry a risk that isn't in a public price (exposure).

• Sector credit headlines with no issuer named; listed names selling off on a leverage story while carrying net cash; BDC and private-credit marks on software loans; refinancing calendars clustering in one year.

11:39 6. Decompose the theme into layers before applying a threat — most threats only hit one of them

1. Refuse the single-noun framing. "You referred to AI as one business. It's not." Write out the layers: equipment that makes the chips → chipmakers → compute providers → model companies.
2. Take each bear argument and ask which layer it actually lands on. A token price war hits the model layer's margin; it does not touch the equipment, chip or compute layers.
3. Test the threat for second-order effects that reverse it: open-source models "use just as much compute as closed source models," so cheap models are neutral-to-positive for the first three layers — which is why Nvidia funds a free one.
4. Anchor the whole theme on a demand number that is layer-independent — actual money paid for the output. "That is people and companies willing to spend money for AI. So that is real economic activity."

• A bear thesis stated about "AI" without naming a layer; end-customer revenue (not capex) as the health metric; second-order beneficiaries of the disruption you're afraid of; a threat that would raise compute consumption rather than lower it.

37:13 7. Size a theme from the bottleneck outward, using a multiplier and a physical supply check

1. Identify the single component everything else is built around — the one where substitution is genuinely hard. Test the moat by asking what the best alternative is and how far behind it is, in time: "a third-rate Nvidia chip is a year and a half to two years ahead of Huawei."
2. Estimate the spend multiplier — how many dollars of adjacent spend each bottleneck dollar drags along ("$8 to $10 multiplier across the rest of tech": CPUs, memory, networking, cooling, power) — and use it to build the candidate list rather than guessing at beneficiaries.
3. Verify with physical evidence rather than models: factory visits, order-book ratios. "In our recent Asia trip demand to supply is 15 to 1 for chips."
4. Place yourself on the build-out timeline explicitly ("year three of an 8 to 10 year buildout") and pre-commit to treating interim scares as scheduled — "gut check moments three to four times a year" — so a drawdown doesn't get mistaken for a thesis break.

• The technology gap to the #2 supplier narrowing from years to months; demand:supply ratios normalising; the multiplier shrinking as the spend shifts from build to operate; a "gut check" that is actually accompanied by falling orders.

38:58 8. Screen incumbents by whether management misjudged the change — not by how big the moat is

1. Accept that install base and moat are not protection: "you had such a moat, you have such an install base — and they essentially miscalculated what AI is going to do to the business model."
2. Read management's own words for the tell. The failure signature is a company "on a treadmill at 2.5 speed" — the 1995 typewriter maker whose press release said the internet changes nothing, "and then all of a sudden a year later they were bankrupt and gone."
3. Ask concretely what the new technology can do to this product: could a general model "actually do your taxes"? Then ask what share that takes, not whether it eliminates the company.
4. Deliberately test the opposite direction too, and exclude the names that adapted — otherwise the screen degrades into shorting every incumbent. Ives explicitly carves out ServiceNow: "I wouldn't put them" there.

• Management language that minimises the change; product roadmaps unchanged two years into a platform shift; pricing rising while feature velocity falls; an incumbent whose AI announcements are marketing rather than a re-priced product.

48:58 9. Audit who profits from the scary narrative before pricing the risk it describes

1. When a widely repeated forecast is promoted by the companies it would harm, treat that as an anomaly requiring explanation, not as confirmation.
2. Work out who benefits if policymakers act on it. Here: regulation that "stops everybody else from doing AI" leaves the two incumbents "the only winners" — "they're pulling the ladder."
3. Check the claim against first principles rather than vibes: "when employees are more productive, the capital gets better returns. More capital gets invested in those productive employees"; "no technology in the last 100 years has ever been a net job detractor."
4. Then price the policy risk separately from the economic one — a false narrative that succeeds still creates a real, durable moat by statute. That is the exposure to underwrite.

• Incumbents lobbying for rules that raise their own costs; safety or licensing regimes with thresholds that only new entrants trip; open-weight restrictions; a narrative shifting after the lobbying goal is met (Ives: "you've seen definite narrative changes from both of them").

18:18 10. For any AI-exposed business you own, ask what happens if the underlying model disappears

1. Trace each portfolio company's dependency on a single third-party model, the way you would a single-supplier or single-customer risk.
2. Score two distinct failure modes: competitive — the model provider sees your data and "they know how your business operates… and if they decide to compete with you, they can"; and existential — regulatory or provider action removes the model and "if you were a business that built your business directly on top of [it], you're out of business."
3. Prefer the model-agnostic architecture (data and ontology owned in-house, models swappable) and price the insulation layer as a real asset — that is the Palantir pitch in one line.
4. Extend the same question to the model providers themselves: enterprises "switching between models… in order to control costs" is the same substitutability, seen from the other side, and it is why the model layer's margin is the fragile one.

• Disclosure of a single-model dependency in a 10-K; providers deprecating models with short notice; a customer's proprietary data sitting inside a provider that later launches a competing product; enterprises reporting multi-model routing to cut cost.