Actionable insights — These Power Stocks (Not Semis) Are The Real AI Beneficiaries

1:56 1. Map the stack in layers, then ask whose complement just got commoditised

1. Draw the industry as a vertical stack of layers. For AI: chips → infrastructure (clouds) → frontier models → applications. Any technology stack decomposes this way.
2. Identify which layer has been capturing a disproportionate share of the economics. That layer is the one a price shock will hit.
3. Apply "commoditize your complement": when the price of one layer collapses, demand for every other layer rises, because the complement got cheaper. Cheap tokens → more enterprise AI → more cloud consumption → higher GPU rental prices.
4. Add Jevons paradox as the volume check: falling unit cost usually raises total spend rather than lowering it. "If those can get reduced 90% or so, Jevons paradox is still in effect."
5. Conclusion: sell/avoid the commoditised layer, buy the layers whose complement it is. Do not read a sector-wide selloff as applying uniformly across the stack.

• An open-source or open-weight release that credibly matches a paid layer's output; a price cut announced by the layer that had been extracting the rent; enterprises publicly saying a component's cost blew their budget (Uber, Amazon here) — that's the pressure that precedes the commoditisation.

9:43 2. Start from a physical constraint, then walk down to the smallest company that relieves it

1. Find a hard physical bottleneck with a measurable number attached, not a narrative. Here: US grid demand flat at ~4,000 TWh for 20 years, AI heading to ~20% of the grid / ~100 GW by 2030, interconnection queues of 18–24 months.
2. Ask what buyers are forced to do when the bottleneck binds — not what they would prefer to do. Forced behaviour is more forecastable than preference. Here: bring your own power.
3. Ask what already exists that lets them do it today. The US "is very fortunate that they have natural gas pipelines almost everywhere," so on-site gas generation is the path of least resistance.
4. Then walk one rung further out to the next constraint on that workaround — no pipeline at this site; too much load for one campus — and find who relieves that. Each rung is smaller, less covered and cheaper.
5. Put a duration on the theme so you know when the thesis expires: he gives bring-your-own-power "3 to 5 years… until more alternatives can be brought to the market."

• Interconnection queue lengths and utility load forecasts; local moratorium ordinances on data centres; announced campus sizes flattening out at 1–3 GW instead of climbing — that flattening is the trigger for the scale-across trade.

13:24 3. The cheap-twin comp — buy the same tailwind at a fifth of the multiple

1. Take the consensus name everyone already owns for a theme and write down its multiple on a metric that survives loss-making (EV/sales works when there are no earnings).
2. Find the company selling a functionally substitutable solution into the same buyers, and write down the same multiple.
3. State honestly where the expensive one is genuinely better — Bloom's fuel cells are "a bit more efficient"; CrowdStrike and Palo Alto earn far more GAAP profit today. If you can't articulate the premium, you haven't understood it.
4. Then decide whether the gap is bigger than the quality difference. "Valuation is a huge driver in our stock picks and we believe the tailwinds are just as strong."
5. Only then check that the cheap twin has a specific reason to win a share of the demand — not merely that it is cheap. Capstone's is delivery speed; Tenable's is being first into exposure management with a large upsell base.

• Capacity utilisation at the cheap twin — Capstone at ~10% of a ~1 GW/yr three-shift capability is where the operating leverage lives; a multiple gap only pays if volume can actually arrive.

29:36 4. Screen for what screens badly — the SBC normalisation trade

1. Look for companies that fail standard quantitative screens for a mechanical reason rather than a business reason. Screening failure removes a whole class of buyers, which is what creates the mispricing.
2. The specific mechanic here: growth has decelerated but stock-based compensation is still set at growth-era levels, so GAAP earnings look terrible.
3. Treat SBC correctly — "SBC is just like any other variable expense is how we believe investors should view SBC." Don't add it back, and don't treat it as permanent either.
4. Ask one question: is there reason to believe SBC as a percentage of revenue falls quickly from here? If yes, GAAP earnings inflect sharply with no change in the underlying business.
5. Buy ahead of the normalisation, not after the screen clears.

• SBC as a percentage of revenue in the last four 10-Qs, plotted against revenue growth. A falling ratio with stable growth is the setup; a rising ratio with falling growth is a value trap.

30:50 5. The post-AI software moat checklist — an n-of-one test, not a sector call

1. Concede the part of the bear case that is true first: production-ready code now costs 90–95% less to produce, so "it clearly requires a re-evaluation of what it means to have staying power as a software company."
2. Then refuse the blanket conclusion: "the evaluation is an n of one. You have to look at each of them as a case by case deal and not just a broad blanketed, okay, well software is going to get destroyed."
3. Run the checklist on each name: Does it collect unique proprietary data nobody else has? Does it carry security and compliance obligations? Enterprise-grade features? How deep are the integrations into other systems? Is there a real-world component — logistics, construction, physical operations — or is it purely a digital tool?
4. Grade small point tools as replaceable and full enterprise platforms as, so far, not: "none of these big software companies have actually seen any sort of supplanting of their products or services yet."
5. Separately, check whether the multiple compression is about AI at all — often it's just a fast grower slowing down, which is a different (and less interesting) problem.

• The first genuine case of an enterprise platform losing a named large customer to an AI-built substitute. Until that exists, the bear case is a multiple story, not a terminal-value story.

37:49 6. Bound the bear case with the company's own disclosure

1. When a stock is being sold on a plausible structural fear, don't argue with the fear — find where the company already reports the damage.
2. Segment the disclosure until the fear is confined to a measurable slice: here, jobs under $500 versus everything above.
3. Size the slice against the whole. If the damaged band is a minority of the business, the price is discounting the wrong denominator.
4. Then look for the mirror-image company where that same band is the majority — you now have the short or the avoid, for free, from the same piece of research.
5. Add your own primary evidence when you have it, while labelling it as anecdote: "we are big users of the Upwork platform and our spend on that platform has continued to grow… obviously we're an n of one."

• Management commentary that segments the weakness by price band, cohort or product tier — and whether the damaged band keeps its share stable or starts creeping upward into the healthy one. Creep upward invalidates the thesis.

40:08 7. Variant perception is the entry test — including when you agree with the trend

1. Before buying, write down what the market believes and what you believe. If they're the same, there's no return in it regardless of how good the company is.
2. Accept two valid shapes of variance: the market is wrong about a hated name, or the market is right about a loved trend but is underestimating its strength. "You can be with the momentum but if you think somehow the trend is even stronger than the current momentum suggests, that's a variant perspective that qualifies."
3. Scale conviction with the size of the variance: "there's usually opportunities to make a lot more money the more variant your perception is."
4. Frame it as probabilities, not certainty — the edge is "if you're right more often than you're wrong," and "the markets may be right on a lot of them a lot of times."
5. Keep the portfolio deliberately mixed between with-the-tide and against-the-tide expressions of the same conviction so a single regime doesn't decide the year.

• Your own thesis restating consensus back to you. If you can't name the specific belief you hold that the current price contradicts, you have a company you like, not a position.

43:30 8. Three sleeves — let batting average vs slugging percentage set the size

1. Core, 5–15%: the highest-conviction, longest-horizon names, where their historical hit rate is "around 70%." Weight these most heavily — the return comes from being right often.
2. Starter, ~3%: ideas taken before the full three months of work is possible, because "in this modern market, things can move so fast." Explicitly provisional: a starter is expected either to graduate to core or to be dropped.
3. Speculative, 1–3%: names with "very little downside protection" where the point is slugging percentage, not batting percentage — "opportunity for multiples of returns in a short period of time."
4. Match the sizing to the payoff shape, not to how exciting the story is. The exciting story usually belongs in the smallest sleeve.
5. Let cash be the residual of the research process rather than a macro call: if origination isn't producing qualifying names, "we're going to have a higher cash weighting in the portfolio."

• Sleeve drift — a speculative name that has multiplied and is now a core-sized weight by accident. The sizing rule only works if it's re-applied after the move.

46:02 9. Averaging down, conditioned on the fundamentals — 20–30% of the alpha

1. Separate the two variables explicitly: what the fundamentals are doing versus what the price is doing.
2. The add is triggered only by divergence in one direction — "if the fundamentals keep doing what you think they're going to be doing and the price is going the other way, you have to add to that position." Price weakness alone is not the trigger.
3. Treat the add as a research event, not a reflex: at Pernas every portfolio action carries a written communication explaining it, which forces the fundamental case to be restated before capital goes in.
4. Size the contribution honestly so it gets the attention it deserves — they attribute "about 20 to 30% of our alpha" to averaging down when appropriate.
5. The corollary they state elsewhere: this is incompatible with "buy and hold forever." You have to keep re-underwriting, which means you must also be willing to exit when the fundamentals stop confirming.

• Write the disconfirming metric down before you average down. Without a pre-committed fundamental tripwire, "adding on weakness" and "doubling down on a mistake" are indistinguishable in real time.

48:50 10. Turn over rocks at a 1-in-100 hit rate — and publish the passes

1. Treat origination volume as the input variable you actually control. "The hit rate for us is roughly like one in a 100" on companies that get real research — so a portfolio of ten positions implies a thousand looks.
2. Write up the rejections, not just the buys. Pernas publishes a weekly Stock Sonar of three names, many of them passes, "to give our members an inside view into the research process and where we're looking."
3. Use the pass file as the calibration set for your own filter: "it gives people a deep understanding into our filter process," and it is what makes the acceptances legible — five robotics companies were rejected before one was bought, and that context is the actual information.
4. Keep the write-ups short. Two to three paragraphs with "no fat at all" — brevity is what makes turning over a hundred rocks sustainable.
5. Accept the portfolio consequence: when the rocks yield nothing, hold more cash. Origination output sets the invested weight, not the other way round.

• Keep your own pass log with the reason and the date, then revisit it when the price halves. A pass on valuation and a pass on business quality age in completely opposite directions.

1:04:38 11. Sum-of-the-parts with leverage — quantify the hair before the upside

1. Find a subsidiary or segment inside a cheap company that can be valued standalone on a conventional multiple, and check it is genuinely separable — "it's a very portable business as well."
2. Express it as a percentage of the whole enterprise value. If one clean asset is 60–70% of EV at 9–10× earnings, the rest of the company is close to free, and "not a lot has to go right for there to be a rerating."
3. Then state the offsetting defect in the same breath, with numbers: "there is some hair on it with the leverage" — $2.5bn of debt against roughly $600m of book value.
4. Adjust the growth rate for portfolio churn before judging it — after asset sales, "organically it's better than it screens."
5. Ask what makes the cheap part structurally cheap rather than just unloved, and whether the niche is defensible: gaming and prediction-market processing works precisely because "the Stripes of the world, the Adians of the world typically stay away from that kind of business."

• Debt maturity schedule and covenant headroom against the value of the separable asset. In a levered sum-of-the-parts, the question is never whether the asset is worth the money — it's whether the company is allowed to keep the proceeds.