Pernas Research — These Power Stocks (Not Semis) Are The Real AI Beneficiaries

0:00 Cold open — two brothers with an audited record

• Farley's framing: the Pernas portfolio has returned "over 1300% gross or 32% annualized" since inception in 2017 — roughly double the S&P — on an audited track record.
• His argument for the format: learn from people "who've actually done investing well, not just talking well or writing well," which means single-stock fundamental work, not only macro and thematics.

1:28 The setup: memory down 40–50% in a month

• "SanDisk, SK Hynix, that entire memory complex and AI complex has been down 40, 50% over the last month and the market's really questioning if ROI is there anymore."
• The proximate cause is open-weight models. Their claim: the market has the sign wrong.

1:56 Four layers, and "commoditize your complement"

• The AI ecosystem is four layers: chips → infrastructure (clouds) → frontier models → applications. Frontier models had been capturing a disproportionate share of the unit economics.
• "If you're in a technology stack and if your complement gets commoditized, demand increases to all the other layers." Cheaper tokens mean more enterprise AI, which means more cloud, which means higher GPU rental prices.
• Bearish for frontier labs; bullish for every layer above and below them.

3:15 Kimi 3, distillation, and where the air pocket lands

• Moonshot AI's Kimi is "now number three in the world" — the fear being that an open model guts frontier-lab pricing power.
• Dean partly discounts it: Anthropic accused open-weight builders of distilling its models, citing ~4 million exchanged messages, so the gains are "not purely due to innovation, architectural improvements."
• Even so, the demand air pocket at the frontier gets refilled at the hyperscalers, because open weights are only cheap if you rent someone's compute to run them. "Which leaves more pie for the hyperscalers to absorb."

6:34 Why it's "a very special time to be an active manager"

• Deiya: the top 10 names are close to 30–40% of the index weight, and those same names face "a step function change in capital spend… they're essentially existentially forced to make these outlays with uncertain ROI profiles."
• "The leaders of the index are being forced to kind of subsidize this spend for the rest of the economy. And you could find other companies that benefit."
• Average the hyperscalers and you get roughly 10× EV/sales while they spend all their free cash flow and more: "a lot is going to have to go right for that multiple to be roughly correct… it's a lot easier to look elsewhere."

9:43 Theme 1 — bring your own power

• US grid demand has been flat at roughly 4,000 terawatt hours for 20 years — efficiency gains cancelling population growth — so nothing was built for growth.
• AI is "poised to take 20% of the entire US energy grid demand by 2030, roughly 100 gigawatts or so." The grid can't deliver it: transmission is "akin to a two-lane highway which should be an eight lane highway," and data-centre load fluctuation is more than the lines can handle.
• Interconnection queues run 18 to 24 months, so hyperscalers must bring their own power — and America is "very fortunate" to have gas pipelines almost everywhere to plug turbines into. He gives the theme 3–5 years before alternatives arrive; open weights only accelerate it.

11:18 Theme 2 — sites stop scaling up, so bandwidth goes 15×

• 1–3 GW monolithic data centres are "going to be topping out soon": you can't get that much energy to one site, and cities are imposing moratoriums over emissions and water.
• So the build shifts from one 10 GW campus to "10 different 1 gigawatt sites" that must talk to each other continuously to run training and inference — bandwidth needs expand almost 15 times.
• This is the "scale-across" thesis, and the reason a Norwegian-listed optical networking micro-cap is in the portfolio.

13:06 Why not Bloom, Constellation or GE Vernova

• Farley names the familiar behind-the-meter list; Dean concedes "you named the largest players."
• The differentiator is price for the same tailwind: BE at 15–20× revenue versus CEPL at about 3×. Bloom's fuel cells are "a bit more efficient"; Capstone's microturbines are less costly.
• "Valuation is a huge driver in our stock picks and we believe the tailwinds are just as strong."

14:23 What a microturbine is — and why time-to-market is the asset

• "A micro turbine is essentially just a large turbine shrunken down" — 50 kW to 3 MW versus the 200–300 MW machines GE Vernova builds.
• Big turbines carry 18–24 month lead times. Capstone delivers in one to three months with significant spare capacity.
• The whole argument in one line: "if you don't have energy, you don't have anything… as soon as you can find it, you want it."

15:02 Capstone — the numbers, and three ways to win

• Initiated 2025-AUG-05; +1,131% to the 2026-MAY-29 peak, then halved, leaving it up ~545%. Roughly $300–350m fully diluted.
• Running at ~10% of a capacity that could produce about a gigawatt a year on three shifts — the operating leverage is entirely ahead of it.
• Three demand paths: hyperscaler data halls of 20–30 MW (not 300 MW — "that's a bit excessive"); thousands of 1–4 MW edge data centres that still have to move from CPUs to GPUs as video/audio latency starts to matter; and manufacturers that can no longer rely on a congested grid.
• Out of prepackaged bankruptcy three years ago after 40 years in business; first profitable year last year.

17:19 Stabilis Solutions — the virtual pipeline

• For sites with no pipeline you need portable gas: liquefy it, move it in cryogenic trucks, store it, vaporise it on site. "It's called a virtual pipeline."
• A niche that has consolidated over two decades to few providers — and Stabilis just signed a data-centre contract Dean puts at a couple hundred million dollars over two years.
• Second leg: SpaceX is a large customer, with launch cadence going "from 50 or so to 20x that in 5 years time."

18:44 The three sleeves — how a $75m stock gets owned safely

• Positions are rated speculative (1–3%), starter (3–5%) and core (5–15%).
• Stabilis, at a ~$75m market cap, is speculative: "the sizing is smaller with those companies."

19:16 Smartoptics — turning data into light between sites

• Hardware that converts data into light to send across fibre to other data centres, plus software to modulate it accurately.
• Its differentiator is interoperability — the boxes are modified to talk to Broadcom, Cisco and Arista gear — with "best-in-class" software. "Everyone in Silicon Valley knows them."
• Listed on Oslo Børs for about four years, headquarters recently relocated to Sweden.

20:40 Vishay Precision — strain gauges for robot hands

• Dean's next-step thesis: robotics hardware exists, the brain doesn't — and high-fidelity physics simulation is how AI supplies it. If so, "humanoid robotics might be the largest industry in 10 years time."
• VPG makes the strain gauges and sensors that go into humanoid hands, with three large customers already in prototype (he names Tesla and Figure).
• Deiya adds the sizing logic: with an opportunity set this large, "even if you're paying up in multiple slightly… it's just too enormous not to have exposure."
• The host flags the confusion trap: this is VPG, not Vishay Intertechnology.

22:48 Cyber's new chapter — from identity to exposure

• ~80% of recent attacks were identity-based: phishing, "walk through the front door with the keys," which is what CrowdStrike-style agents were bought to contain.
• Frontier models can now read thousands of lines of code and find vulnerabilities, so "the cost of finding vulnerabilities in code has gone down significantly" — traditional hacking is back.
• The new buying question is exposure: where are all my assets, which vulnerabilities exist, which ones matter, and what happens if one is exploited. "When I say CISOs and CTOs are freaking out about it, it's an understatement."

27:13 Tenable — exposure management at 4× sales

• The product a large enterprise now needs: inventory every asset, scan it, surface the hundreds of vulnerabilities, rank the handful that matter, and fix them.
• Tenable started building exposure management in 2022 off a commoditising vulnerability-scanning base, and uses an Anthropic partnership to run frontier models over customer assets and code. Huge upsell base.
• Valuation: 4× EV/sales versus CRWD and PANW north of 20× — for a player that may be better positioned for the platform shift. They expect the ~10% growth rate to accelerate; the culture has "significant urgency."

29:36 Why good software screens badly — SBC and Sprout Social

• Growth has slowed at many SaaS names while stock-based comp stayed "extraordinarily high," so they screen poorly on GAAP. Sprout Social is his named example of a holding that screens badly for that reason.
• "SBC is just like any other variable expense is how we believe investors should view SBC." If SBC as a percentage of revenue can fall quickly, you get "very attractively enhanced GAAP earnings just as a result of SBC normalizing."
• Sprout was priced "for complete decapitation" — "such an easy call to make" — and is up 50–60% since April.

30:50 What actually makes software defensible now

• He concedes the bear case is partly right: production-ready code costs 90–95% less to produce, which "clearly requires a re-evaluation of what it means to have staying power as a software company."
• But no large enterprise software company has actually seen its product supplanted yet. Small tools are replaceable; "fullthroated enterprise" software isn't — yet.
• The moat checklist: proprietary data they collect themselves, security and compliance, enterprise-grade features, depth of integrations, and whether there's a real-world component (logistics, construction — "a company like a Procore") rather than a purely digital tool.
• "The evaluation is an n of one. You have to look at each of them as a case-by-case deal."

34:57 Upwork — the marketplace nobody can rebuild

• "We've forgotten how special marketplace businesses are and how difficult marketplace businesses are to build" — network effects make Upwork "impossible to replace."
• The damage is confined to the bottom: Upwork's own data shows weakness only in jobs under $500 — "I need a logo created, I'm going to give you $300 for it" — which is Fiverr's core, not Upwork's.
• Higher-value consultative work survives because clients "don't even know what questions to ask," and AI makes the expert on the other end more efficient too. "It's a whole lot easier to just talk to somebody who understands what you're doing… instead of having to prompt AI 18 million times."
• Price: a little over 1× EV/sales and a little over 6× EV/FCF after SBC. "I don't know where you find a better bargain than that in this market."

40:08 Variant perception — and why buyside research

• Being "with the tide" still counts as contrarian if you think the trend is stronger than the momentum implies: "there's usually opportunities to make a lot more money the more variant your perception is."
• Why they built a research firm: sellside research is "more coverage than anything," full of biases and without conviction, whereas buyside analysts "have skin in the game, they're invested alongside of their ideas… it reads with more conviction."
• And the discipline nobody else keeps: "track records are important. How has your portfolio done? Are you any good at what you do? If you're not you should just go home."

43:30 Batting average vs slugging percentage — and the 1-in-100 hit rate

• Core positions are the high-conviction, longer-horizon names with a roughly 70% batting average and the heaviest weights. Starters (~3%) are ideas taken before the full three months of work is possible, able to graduate to core.
• The speculative sleeve is the opposite discipline: "very little downside protection, but the upside — it's really about slugging percentage. It's not about batting percentage."
• Origination is brute force: "the hit rate for us is roughly like one in a 100." When the rocks turn up nothing, cash weighting rises — portfolio construction falls out of research output rather than a top-down target.
• Averaging down when the fundamentals hold and the price falls "makes up about 20 to 30% of our alpha." They are "not believers in the buy and hold forever philosophy."
• Deliverables: an initiation report per position plus action updates, a monthly report, and a weekly Stock Sonar of three names — including the ones they passed on, which gives members "a deep understanding into our filter process."

49:08 Meta — in near $90, out above $700, negative now

• Small and mid caps are "incidental" — there are simply more of them — but Meta was once a large position, bought after the 2022 drawdown of ~70% on ad-tracking, TikTok and metaverse fears, near $90 from $300-something.
• Sold above $700 last year on valuation plus disappointment in Llama; he blames cultural mishaps, naming Yann LeCun's position that LLMs are "a cul-de-sac for AI."
• Zuckerberg still rates highly with them as an operator — mobile monetisation, copying Snapchat — but the AI record does not.

51:34 Meta now — "they're just floundering"

• The original strategy was coherent: Meta isn't a cloud company, it's a content company, so open-sourcing AI to make content cheap fed the largest network in the world.
• Now: "it seems like they're just floundering and trying to find the best use of the thousands of GPUs they've acquired," and hundreds of millions spent poaching researchers "didn't really yield anything."
• Meta's capital allocation is "a bit different than Microsoft and Google who are in the cloud business… a lot more measured and a lot more rational" — Meta's spend is partly defensive against TikTok. Conclusion: "the outlook for Meta is definitely negative going forward."

54:48 Why there are no semiconductors in the portfolio

• "With semis it's a lot harder to play. Especially given China. We've learned that quite the hard way."
• China is "no longer just a producer of $1 widgets" — already effectively monopolist in EVs and drones, now going after ASML and the memory makers. CXMT's IPO at a roughly half-trillion market cap is the evidence.
• The ability to make DRAM, CPUs and lithography machines far cheaper "definitely gives us pause regardless of how bullish the overall industry looks."

55:31 Payments — the part they actually want is cross-border

• Banking systems were built national-first; international transfer was an afterthought, leaving the clunky correspondent-banking and SWIFT stack. Moving money that way costs about 5%; fintechs do it near 2%, with tracking and confirmation the old rails never gave.
• The second engine is M2: central banks "are going to continue to print more and more money… traditionally it's grown at around five or six% a year." A company with a durable scrape on that growth compounds hard.
• Everything hinges on defending the take rate — "the competition is fierce and you got to pick the right one."

58:48 Two ways to defend a take rate — Remitly and Wise

• Remitly defends with mindshare: migrants aren't price-only buyers, "trust is a huge deal for them, branding is a huge deal for them," so the take rate erodes only "a bip here, a bip there" while send volume grows 30–40% and revenue 25%.
• Wise defends with infrastructure — direct plumbing into national central banks nobody else has replicated — and then attacks with price: "I'm going to be the loss leader… it's not defense," which Farley captions as "they're the one causing the storm."
• The two barely overlap: Wise is aiming at businesses, banks, correspondent banking and even the card networks; Remitly runs on Visa/Mastercard rails for 90%+ of its volume.

1:03:43 Paysafe — sum of the parts, with hair on it

• A conventional PayPal-style processor with an orientation to gaming and prediction markets — a segment Stripe and Adyen avoid, which is exactly the growth he wants exposure to.
• The rerating case is a digital-wallet subsidiary worth 60–70% of enterprise value at 9–10× earnings, and "very portable." "Not a lot has to go right for there to be a rerating."
• The hair: "significantly overleveraged" — the host cites $2.5bn of debt against $600m book value — with high-single-digit organic growth better than it screens after asset sales, and a new-CEO turnaround paying debt down.

1:06:09 Stablecoins — the threat he isn't worried about

• Corporate stablecoin adoption is a treasury story, and "correspondent banking… is really going to be the loser of all that" — not Wise or Remitly, whose customers are consumers and small businesses.
• Usage is overwhelmingly crypto speculation. Outside broken banking systems in parts of Africa, "there's absolutely zero cases of people actually using stable coins like they're using fiat. It just isn't happening at all."
• The market has already faded the narrative: "I know Circle's come down significantly as a result of this narrative fading."

1:08:04 Where payments value is accruing — and the PayPal pass

• Off a fintech-ETF holdings list (Robinhood, Block, Visa, Global Payments, Affirm, Toast) his one take: processing is commoditising and "the value seems to be accruing to the application layer, to the Toasts of the world" — verticalised software that runs the whole business and embeds payments.
• Rule: "if you're investing in one of these companies they have to do more than just processing."
• Hence the PayPal pass — they liked Braintree's processing (volumes growing fast, scale economics ahead) but not branded checkout, "and put it all together, we couldn't like it just given how profitable the branded side of things were."