SemiAnalysis's Doug O'Laughlin on all things AI, Power, and Corporate Governance
Doug O’Loughlin thinks the AI capital cycle is only beginning to acquire the leverage required for a genuine bubble. Hyperscalers have funded the buildout from operating cash flow, while CoreWeave historically financed GPUs only after winning contracts; Oracle showed what becomes possible when debt enters the equation. Doug points to accelerated depreciation, less growth in capital requirements for risk capital and potentially lower rates; Andrew adds political pressure to announce American data-center investment. Andrew thinks the buildout still has a long way to run, even if a hangover could arrive in three years.
The market may be understating how radically AI changes Big Tech from capital-light platforms into capital-intensive infrastructure businesses. Mid-20s P/E multiples can look undemanding beside 20%-30% EBIT growth, but cash outlays make price-to-free-cash-flow expand as the business model changes. Andrew’s pushback is that today’s AI spending may suppress current earnings to produce tomorrow’s returns; Doug concedes that demand exceeds supply, but warns that “capital-intensive businesses tend to trade at low multiples of earnings.”
Google’s TPU is a credible number-two accelerator—and at the right price, Doug would choose it over NVLink. Doug says TPU’s 3D-torus architecture can be sliced and scaled without the NVL72’s failure blast-radius problem, where one failed GPU can pause the other 71; he also says TPU already handles training and inference in production, with Midjourney allegedly using it entirely. Google’s willingness to sell TPU externally and distribute Gemini beyond its own properties—including on Siri—creates “a real chance of being a true number two.”
The scaling thesis has shifted from ever-larger pre-training runs toward reinforcement learning, test-time compute and verifiable work. Doug accepts that the December case for a pre-training scaling wall was “probably pretty valid,” while treating Gemini 3’s expected release that month as a test. He argues that RL can probably improve tasks with checkable outcomes—from buying a book in fewer clicks to producing superhuman Dota 2 players. The ambition has narrowed from mystical AGI toward making “all white-collar information and knowledge” nearly free at the margin.
Inference hardware already offers extraordinary theoretical economics; selling and monetizing every token is the constraint. Doug estimates a roughly $3.5 million GB200 could generate about $5 million annually if all its tokens were sold, while the stack includes roughly 60% TSMC, 75% Nvidia and 40% neocloud margins before OpenAI’s losses. Supporting perhaps one billion monthly users for several billion to $10 billion should be monetizable—failure would be “a skill issue”—especially if agentic commerce can collect a take rate on purchases.
Power, not semiconductors, is the more structurally offsides bottleneck. Doug believes Taiwan could double chip output in roughly two years, while “we cannot double the power in two years”; conventional operators are still absorbing 120-kilowatt racks as faster players design for 500 kilowatts, one megawatt or even three megawatts. That mismatch benefits miners with usable interconnections and industrial contractors such as Comfort Systems, where roughly 40% revenue growth produced approximately 80% EBIT growth.
In a shortage, specialists can overthink relative winners. Generalists underestimate the physical scale because the “ginormous industrial revolution” is occurring around rural data centers rather than city centers. Specialists can become trapped in questions such as whether the Anthropic-Amazon deal means Trainium is “screwed,” while industry participants see absolute scarcity: even abandoning Trainium for GB200s could send Amazon revenue “through the roof” because every usable unit has value.
Executive compensation is useful signal only when an aligned board ties it to a controllable fundamental lever. Off-cycle PSUs can reveal a board’s intended lever, but they can also be delusional “Hail Mary grants” or incentives to pump a stock. Broadcom illustrates the reflexivity; Opendoor’s new CEO received a $1 salary and awards tied to roughly $9, $13, $17, $21 and $33 share prices while actively promoting the stock. Doug also flags Elastic (ESTC), where a giant off-cycle PSU might be aimed at consumption or an AI-linked search API, while admitting he has no fundamental knowledge of the company.
Target Hospitality combines an incentive tell with two possible demand shocks. The company lost major contracts during a sale process that was later pulled, including the Pecos Children’s Center agreement once worth roughly $2.5 billion. After the stock fell from about $8 to $5, Andrew says the company announced roughly two million PSUs four days later; he described the award as a roughly 5x share-price incentive, with detailed thresholds stated as $20 to $30. The possible levers are roughly 6,000 available detention beds and a West Texas data-center labor shortage—“the two most secular trends of our time” in one company, with execution still uncertain.
1. Debt is the ingredient that can turn the AI boom into a bubble
Doug’s distinction is financing: hyperscalers have so far used the “debit card”—cash already in the bank—while Oracle demonstrated the much larger buildout possible with debt. CoreWeave historically used a delayed-draw term loan and other GPU financing after securing a customer contract rather than building speculatively on spec.
The policy backdrop compounds the cycle. Doug points to accelerated depreciation under the bill referred to in the transcript as OBBB/OBBBA, less growth in capital requirements for risk capital and a potential Federal Reserve shift toward lower rates. Andrew adds political pressure to announce American data-center investment: executives may be asked first how much they are investing in U.S. data centers.
Andrew’s framing captures the reflexivity: once one company can borrow and “bet the firm,” competitors are unlikely to stop at internally generated cash. Doug sees today as relatively early because broad speculative construction—the truly dangerous phase—has not fully arrived. Andrew thinks there may still be a three-year runway before a major hangover.
2. P/E obscures the cost of turning platforms into infrastructure
Big Tech can screen reasonably at mid-20s earnings multiples while growing EBIT 20%-30%, yet that comparison relies on historically capital-light economics. As AI infrastructure consumes enormous cash, earnings and free cash flow diverge: “Price to free cash flow blows out.”
Andrew’s pushback is that Google may be taking AI losses upfront, so its consolidated multiple could represent roughly a 10x Search business, minus roughly 10x of moonshots and plus or minus roughly 5x for AI. If capacity is scarce today and monetization follows tomorrow, current earnings may understate normalized profitability.
Doug grants the accounting argument but retains the valuation warning: every hyperscaler is making the same capital-intensity transition simultaneously, just as depreciation incentives encourage still more spending. Low P/Es do not automatically mean cheap stocks when the underlying business becomes structurally heavier.
3. TPU has become a real competitive vector against Nvidia
Doug calls Google’s TPU “the second best chip of all time,” not a Bing-like distant runner-up. Its 3D-torus topology permits flexible slices across large clusters, whereas he says NVL72’s 72 tightly linked GPUs carry a larger failure blast radius: if one fails, the other 71 may be stopped or paused.
His price-sensitive conclusion is unusually direct: “At some price I would definitely take TPU over NVLink,” assuming comparable software. TPU already supports production-scale training and inference, and Midjourney allegedly runs entirely on it.
The corporate change matters as much as the silicon. Google previously appeared determined to keep its best AI and infrastructure inside Google properties; willingness to put Gemini on Siri and sell TPU externally broadens distribution, improves utilization and gives Google “a real chance of being a true number two.” The recent Anthropic TPU deal is part of that external-TPU opportunity.
4. Scaling continues, but the mechanism has changed
Doug says the December argument that pre-training had reached a scaling wall was “probably pretty valid.” He had nevertheless heard strong excitement around Gemini 3, described as the newest and biggest model, and said he had been told it was due sometime that month; he treated its release as the next “proof in the pudding.”
Reinforcement learning is now the industry’s main enthusiasm because it can probably work wherever success can be verified. An agent can fail repeatedly at purchasing an Amazon book, receive simple good-or-bad feedback, and eventually discover the minimum-click path. Doug expects Amazon’s blocking of agents to become a platform war.
OpenAI’s superhuman Dota 2 players are Doug’s load-bearing example: scaled RL can produce performance beyond humans in a complex, verifiable domain. Pre-training is slowing, test-time compute is probably not slowing but requires a larger budget, and post-training and RL “are definitely not slowing down.”
The destination has also become more practical. Rather than centering AGI that could “zap your little brain,” Doug focuses on pushing the marginal cost of white-collar information and knowledge toward the cost of GPU compute.
5. Token economics work before application economics do
SemiAnalysis’s InferenceMax.ai work leads Doug to estimate that a roughly $3.5 million GB200 could generate around $5 million of annual token revenue if fully utilized and sold. He calls the resulting economics a “well over one year payback period”—his wording—while identifying utilization and customer monetization as the actual questions.
Andrew challenges the circularity: a neocloud may rent GPUs profitably to an AI startup, but the startup itself may have no revenue, creating a dot-com-bubble-style house of cards that feeds on itself.
Doug’s aggregation answers only part of that objection. He cites roughly 60% margins at TSMC, 75% at Nvidia and 40% at a neocloud, followed by perhaps negative 50% at OpenAI; the overall token stack may still be profitable even if the frontier-model provider cannot yet fund its ambitions.
Using a DeepSeek-style model, Doug thinks serving one billion monthly users might cost several billion to $10 billion. “You’re telling me you can’t make $10 billion off a billion users?” His answer is that monetization should be achievable, though capability investment remains open-ended.
6. Agentic commerce could monetize an otherwise deflationary technology
Andrew says he uses free OpenAI access in place of paid food, fitness and tracking apps that might otherwise cost $10, $50 or $100 per year. Doug agrees AI is “insanely, ridiculously, stupidly deflationary,” apart from the continuing need to pay for energy.
The healthier business model is to sell services and collect transaction economics. Doug’s example is a $100 keto grocery basket assembled and purchased through Instacart, with perhaps a 10% platform fee shared with ChatGPT for originating and completing the transaction.
Consumers are already conditioned to pay take rates for convenience, making agentic purchasing a potentially high-margin bridge between cheap intelligence and durable revenue. The broader historical analogy is railroads displacing canals: enormous deflation destroys old workflows but eventually creates new activity on the other side.
7. Physical power cannot scale at silicon speed
Doug believes TSMC and Taiwan could roughly double chip production within two years by adding equipment. Power cannot double on that schedule, making energy and interconnection “much more offsides than semiconductors.”
The cultural mismatch is stark: chip teams discuss one-megawatt racks while traditional data-center operators react, “Are you fucking kidding me?” Digital Realty-type operators are still shell-shocked by 120-kilowatt racks; faster players such as Vantage and Crusoe target 500 kilowatts, while Switch’s Rob Roy is pushing toward three megawatts.
Grid utilization, batteries, backup generation and peak shaving can release capacity, but Doug still sees enormous unmet demand. His scale marker is six gigawatts for New York City; the accelerator orders already placed imply power needs that the data-center buildout has not matched.
The earnings leverage may be clearest in industrials. Comfort Systems’ cited quarter delivered approximately 40% revenue growth, 80% EBIT growth and roughly 50% EPS growth because its fastest-growing segment also carried the highest margins, despite severe capacity constraints.
8. In a shortage, specialists can overthink relative winners
Generalists underestimate the physical scale because the “ginormous industrial revolution” is occurring around rural data centers rather than city centers. Doug cites data centers and GPUs as contributing roughly 190 basis points to the last quarter’s GDP growth.
Specialist investors, meanwhile, can become trapped in relative questions—whether TPU hurts Nvidia, or whether the Anthropic-Amazon deal means Trainium is “screwed.” Industry participants are operating in absolute scarcity: even abandoning Trainium for GB200s could send Amazon revenue “through the roof” because every usable unit has value.
SemiAnalysis encountered the same error comparing Bloom Energy with cheaper, more reliable onsite gas. The customer answer was simply to deploy both: “They want both.” In the present shortage, a rising tide is lifting nearly every boat that can be lifted, including imperfect power and data-center assets.
9. Compensation signals require fundamentals, control and honest boards
Doug cautions that “boards are snowflakes”: some genuinely pursue shareholder value, some are crooks and others are simply wrong. An RSU without performance conditions carries little information, while distressed companies can issue spectacular out-of-the-money PSUs and still go bankrupt.
Broadcom illustrates reflexivity. After Hock Tan rapidly achieved an earlier aggressive share-price award, a new multibillion-dollar package led him to say, “I’ve got to sell a lot of AI”; incentives may drive execution, but they can also encourage public promotion to manufacture the trigger.
Opendoor’s new CEO received a $1 salary and awards tied to roughly $9, $13, $17, $21 and $33 share prices while becoming highly active on X. Promotion can be legitimate investor communication for an underappreciated asset—or “lipstick on a pig” when the business cannot support the story.
Doug also flags Elastic (ESTC), where the company issued a “ginormous” off-cycle PSU that might be intended to incentivize consumption, an AI opportunity or its search API. He explicitly says he has no fundamental knowledge of the company.
The cleanest award attaches pay to an internally controllable lever. Paying executives merely to refinance debt due in 18 months resembles a participation trophy: “If you don’t press the button, you die.” Doug’s acid summary is that independent directors are independent because “they’re looking out for themselves.”
10. Target Hospitality combines an incentive tell with two possible demand shocks
Target Hospitality lost two contracts during a sale process that was later pulled, and the speakers separately discussed the Pecos Children’s Center as the major recent loss; its total contract value was once described as roughly $2.5 billion.
After the stock fell from about $8 to $5, Andrew says the company announced roughly two million PSUs four days later. He described the package as an effectively roughly 5x share-price incentive, with the detailed thresholds stated as $20 to $30 and a potential $60 million payout. Andrew argued that such grants are generally prepared in advance rather than awarded willy-nilly, making them a possible signal of other upcoming levers; Doug also cautions that the board may simply be wrong.
The first lever is detention. Target kept roughly 6,000 vacant beds maintained and ready while ICE sought to expand from around 50,000 beds toward 100,000. OBBBA funded approximately $45 billion for beds, with Doug attributing roughly $30 billion to structures, although the government shutdown delayed new requisitions.
The second is West Texas data-center construction. At about 2,500 workers per gigawatt, 20 of ERCOT’s proposed 80 gigawatts could require 50,000 workers in places where projects exceed local populations; at $120 per worker-day, that implies a roughly $2.1 billion annual housing pool. Target will not win it all, but it has modular inventory, an existing footprint and “a massive incentive to figure it out.”
Full transcript
You're about to listen to the yet another value podcast with your host me, Andrew Walker. I guess before we get into it, it look it means a ton if you can rate, subscribe, review the podcast wherever you're watching or listening to it. But for today, I think you're really going to enjoy this one. We have Doug O’Loughlin. He is from Semi anal semi analyst. He's from Fabricated Knowledge. He is just an absolute absolute expert on all things semiconductor, AI, power, all of that. Just an absolute expert. He is at the center of the field. We have a really fun conversation. And here's a bonus. He's really effing good at corporate governance, springloadads, all that type of shenanigans where if investors are really reading through the AKs and stuff, they might find a management team that all of a sudden flips really bullish or signals that they're going to sell the company. Obviously, nothing's guaranteed. You you can see the full disclaimer. Nothing's invested advice. See the full disclaimer at the end. But we have a really fun conversation on the back end about some of the crazier corporate governance signals and everything we're seeing. So, that's the podcast today. I I'm going to get there in one second, but I'll just roll right into the advertisement. This podcast is sponsored by AlphaSense. AlphaSense is obviously a fantastic sponsor of mine, but one of the reasons Doug is coming on the podcast is because A, he's a friend, but B, he is doing a webinar with AlphaSense that is going to go live. I'm posting this podcast on Tuesday, October 28th. It will go live Tuesday, October 28th. So, if you kind of like what you're hearing from Doug here, I'll include a link in the show notes. I'll include a link on the blog, all that sort of stuff. you should go sign up for the blog because again here's a little secret semi analyst Doug if you were a random paw shop guy who wanted to talk to them about what's happening in semiconductors and AI this podcast is free the AlphaSense webinar is free I promise you Doug's time ain't free you would be paying quite a good bit of money so uh look anytime Doug speaks I listen I think he's super thoughtful I think you're going to enjoy this and if you enjoy this I think you'll enjoy the Alpha Sense webinar so you should go check that out see a link in the show notes and with that said we're going to hop on into the podcast all right hello and welcome Yet another value podcast. I'm your host Andrew Walker
With me today, I'm happy to have on, from Fabricated Knowledge and SemiAnalysis, Doug O’Loughlin. Doug, how's it going?
Good, man. I'm worried that I have to follow up one of the best episodes of Yet Another Value Podcast ever, but we'll try. Maybe I have another 30-bagger accidentally.
I'm laughing because the last episode I published was me just talking to a camera for 30 minutes. Being the narcissist I am, I assumed you meant the episode of me talking, not the episode Doug is referring to—literally the best pitch on Yet Another Value Podcast, Episode 166, AppLovin. Doug pitched it at, what, like 15? It's at 500 or something.
Yeah, something stupid. I don't own it anymore, so it's whatever. I see it and I get sad. Honestly, it's disbelief. I remember that one, too. I think I was super delayed on that one, and then we were like, “Fuck it. We'll do this podcast on that one.” So here we are. I'm a little sick today after the SemiAnalysis retreat, but we're here to chat about AI.
I remember being “sick,” quote-unquote, in college a lot.
To be clear, I was actually hung over. I probably drank 8 of the last 10 days, if I had to guess. So, yeah, I deserve to be sick, and I have been hung over, but I know distinctly what is hung over and what is sick. I am sick. Actually, someone on my team is sick, so everyone got sick, and this is a biorocessing bottom pitch and that's where I'm
All I hear when I hear “drank 8 of the last 10 days” and “Doug, again, at SemiAnalysis” is, “Oh, man, the AI guys are partying so hard, and the value guys are just so sad all the time.”
Doug, tons of stuff we want to talk about today. I want to talk about corporate governance. I want to talk about AI. Do you want to start with AI, just because that's probably the buzziest stuff, and then we'll flip to corporate governance?
Yeah, let's do it, man. As you know, AI is everywhere. AI stands for artificial intelligence, and it's everywhere all the time. I'm just memeing, but, yeah, it's pretty hot, dude. You want to hear my bubble pitch, honestly, because it feels—
I would love to.
Okay, this is something that I feel like I've been—every day, I feel like I have 20/20 vision, and I feel like I'm lucid, and I'm like, “This can't possibly be happening,” yet it continues to happen. This capital cycle is just incredible. After I read the railroad-bubble books, maybe I'll have a different take on this, but I've read a lot of telecom-bubble stuff. I wrote a really good telecom-bubble piece, I want to say, 2 years ago.
I think the one difference between the telecom bubble and this is how much debt is involved, which, until very recently, has been effectively zero. Oracle kind of showed us what's possible. There's a lot of debt available, and whoever gets the next job at the Fed is the guy with the lowest number. U.S. Treasuries are only a 5% discount to the Microsoft 10-year. So you have all these things.
Then there's even more stuff. I saw today that the government doesn't want as much growth in capital requirements for risk capital, and you're just like, “Wow, everything is kind of coming together.” OBBB is effectively an incentive to invest more today. There's just so many things.
Because you have the bonus depreciation, every single possible incentive for you to invest in a capital-heavy GPU data center today is happening today, and everyone wants to do it. So here we are.
Doug, I'm so glad you started with that, because the first note I had is that you wrote, after the Oracle deal was announced, “The bubble is starting.” Exactly what you said: to date, even though everybody says, “Bubble, bubble, bubble,” everything has been out of cash flow, right? Facebook, Microsoft, Google—all these guys. Yes, they're investing tons, but they're doing it out of operating cash flow.
Oracle was the first one who said, “We're going to take out debt. We're kind of going to bet the firm,” by doing this. Once 1 person does it, guess what? It ain't stopping there. Everybody's going pedal to the metal.
We've got people starting to use debt instead of equity. That can blow the bubble up huge. You and I were laughing before: President Trump—literally, if you're an executive and you go into a meeting, the first thing he's going to say is, “How much are you investing into data centers in America?” He asked a CEO, and nobody was like, “We don't do data centers.”
I just feel like it's all speed ahead. Everyone's just pushing this to a big bubble, and everyone can party. Well, yes, there might be a big hangover, as you have right now. There might be a big hangover in 3 years, but it just feels like we've got a long way to go.
It kind of terrifies me that that's the conversation starter. I mean, it's like you can't go bankrupt on a debit card, right? What we've been doing up until now has been all debit-card financing, meaning it's cash in the bank.
Maybe securitize Nvidia GPUs if you're CoreWeave or something.
Yeah, maybe securitize them. But even then, CoreWeave never built on spec, meaning they only build if they have a contract. So they would go out, win a contract, and then find the financing.
The DDTL loan—essentially a delayed-draw term loan—was specifically focused on the GPU-financing side. It was like a flexible revolver focused on GPUs after you win a contract. So it's still different from what would really be kind of crazy, meaning everyone goes and builds on spec.
We're still in these relatively early stages, and everyone has been doing this based on how much cash they've raised from the most valuable businesses in the history of capitalism. All the hyperscalers are still cash-flow positive, and I think the thing that may be slightly underestimated or underappreciated is that you see this and you're kind of going crazy.
For example, you're like, “Oh, price-to-earnings is actually not that high compared to history because it's fine. They're all in the mid-20s, but they're growing EBIT 20% to 30%.” You're like, “Well, this doesn't seem that crazy.” But maybe the one difference that I think is underappreciated is that price-to-earnings is a really, really good metric when everything is super capital-light. Right when something goes from capital-light to capital-heavy, price-to-free-cash-flow blows out.
And so you have these price-to-earnings multiples. It’s pretty hard for price-to-earnings to really expand—or rather, the earnings don’t flow in mechanically—because you have these ginormous cash outflows.
Let me push back on that, because I think a bull would say, “Hey, Doug’s right, right? This huge AI supercluster that Google’s building is a lot more capex-heavy than Google Search, the best business in the history of the world.” But I think a bull would push back and say, “Hey, right now they’re taking all the losses upfront for these giant AI spends.”
So, yes, in the future, cash flow isn’t as good, but they’re bearing the losses now. So when you say 20 or 25 times for Google, that’s probably 10 times for Google Search, minus 10 for all the Google moonshot bets, plus or minus 5 for the AI. I think that would be my big pushback there.
Okay, so that’s probably fair. You cash-burn today for a reward tomorrow, and that works when we know the supply and demand. There’s more demand than supply today. I feel very emphatically yes on that.
And so, yeah, I agree with that. You’re like, “Oh, they’re under-earning, and they’ll over-earn, or they’ll earn their right share tomorrow.” That’s what accounting is for. But I just think it’s really crazy that it happens all at the same time, and the capital-intensity shift happens right as double depreciation kicks in.
So effectively, you get to expense this stuff. It’s just this weird dynamic that everyone’s price-to-earnings multiples are really low when everyone’s business model changes into a more capital-intensive business. You could be like, “Well, what’s the big deal, dude? Capital-intensive businesses tend to trade at low multiples of earnings.” That’s just how it works.
As someone who generally buys capital-intensive businesses, I know you’re really hitting where it hurts. [laughter]
Yeah. The whole thing is really interesting. If we’re talking about that—and actually, while we’re here, because I’m going to do some takes on the Google side—the thing that I think, because I don’t think you definitely follow this space as deeply as we do, is that the one that’s definitely hitting the newswires a lot more right now is this recent Anthropic TPU deal.
We at SemiAnalysis called this out in July or something like that. We were really early on that, and we were very excited about the TPU opportunity because it’s the second-best GPU. It’s the second-best chip of all time.
So if you want to do the real bull case on Google, it’s a trillion for the TPU, whatever GCP accelerates, and that’s like AWS, Waymo, and Search is free, bro. That’s the real bull case. I’m just kidding, but I really think—
TPU—I mean, again, I’m a mouth drooler, but you’re saying the TPU has the possibility to be competitive with Nvidia, is basically what you’re saying.
It is definitely competitive with Nvidia. It is number 2. It is number 2. There’s no other—
Number 2 does not necessarily mean competitive. Bing is number 2 to Google. It is not competitive with Google. Is this Bing, or is this Android and Apple?
Okay, so let’s—it’s time for some AI rumors. The best pre-training chip in the world is the TPU. Everyone says that, and one of the reasons is that the TPU is a much more stable and scalable architecture for pre-training specifically.
So there are a lot of advantages, and the tech tree that the TPU has gone down has a lot of resilience and advantages for software that make it very attractive at scale. I’m going to try to do this justice, but pretty much the NVL72, which is 72 GPUs all tied together, has a really big blast-radius problem that’s complicated, and it doesn’t really work.
Meaning, if 1 of the GPUs fails, they’re like, “What the hell do I do with the other 71?” Seventy-one of them will get stopped or paused. The 3D torus—which you’re going to have to freaking Google—is very complicated. It’s a 3D cube where all the—
You just gave the bull case for Google, right? You said “Google this” while you were pitching the Google TPU.
Yeah, there you go. Sorry—ChatGPT this. It’s an all-to-all mesh, so you can make slices of TPUs and scale this very, very big.
There are 2 different philosophies on how to scale into really, really large clusters, but they are both very compelling and valuable philosophies. I think TPU could be very price-competitive. It’s just a margin question. I would argue that at some price, I would definitely take TPU over NVLink, if all things were equal with software.
It’s actually a product that runs inference in production at scale today, and training and inference in production at scale today. It’s a real competitor. You can argue it’s number 2, and I think at the right price it would be number 1 for me.
Midjourney, for example, allegedly uses all TPUs. There are companies at scale today that are stoked to use TPUs, and I think that’s going to be a big and new, interesting competitive vector that people don’t appreciate, or haven’t been appreciating.
Google’s go-to-market is like, you don’t want to bet on those PMs because they get killed by Google. “Killed by Google” is one of my favorite websites. But I think something has changed in the last year, because in the beginning they really were like, “Oh, we’re going to have the best AI. It’s going to be done on the best infrastructure that’s ours, and you’re only going to be able to use it in our Google properties online.”
I think they’ve been willing to open the aperture by being willing to do Gemini on Siri and sell TPU externally. They’re willing to open it up so that they can have more customers. That means they can have a real chance of being a true number 2.
Killed by Google. It’s one thing to kill Google Fi or Google Meet or whatever, when it would be a nice business, but it’s not. It’s another thing with AI: It’s existential for them. Every big tech company has clearly realized that, and I’m sure they had a vision last year, but at some point you start to realize, hey, we can’t do this all internally. It’s too much of a scale game. It’s too much of a moat.
I’ll be honest, that was super interesting, but we’re at the far limits of my technical knowledge. Even though, if I were a better interviewer and better at this, I would dive into so many more questions, I’m going to back up to my dumb-dumb brain and ask you something else.
For the past year, when I would talk to a value-investor friend—and I’d be more of a value investor than a SemiAnalysis growth investor—they would say, “AI bubble.” They would all point to the models stalling out. Inference is scaling out, the models are stalling out, and everything.
I don’t know where we are in that, but I just want to ask you: If I got an AI bear on here, the first thing they would say is that the models are stalling out and they’re not improving. We’re hitting the limits of scaling. What would you say to that question?
Okay, that’s a great question. It’s pretty hard to answer. In December specifically, the idea that scaling walls were done on the pre-training side was probably pretty valid.
I think there’s a lot of hype right now around Gemini 3, which is supposed to be coming. I was told Gemini 3, which is the newest and best model, is going to be the newest and biggest model in existence, and I think that comes out sometime this month. That’s the proof in the pudding. A lot of people are very excited about it. We’ll see.
There have been multiple scaling vectors for how these models are getting better. I think we’ve become a lot less focused on AGI—superhuman intelligence that’s going to zap your little brain because you can’t even understand it—and more focused on the idea that essentially all white-collar information and knowledge is free, or the marginal cost goes to your GPU. That’s really the focus of where things are going right now.
Specifically, what the entire industry is very excited about is RL. RL means reinforcement learning. In the same way you and I learn how to do stocks, good or bad, on a very lowend size data set—our history, our careers, whatever—we’re like, “Dude, this setup really works for me, and I do a really good job.”
Pretty much, the goal is that you do training on these models to be like, “Hey, your job is to check out this book from Amazon in the least amount of clicks.” In the beginning, it literally clicks everything. You can just say, “Hey, no. Bad job, bad job, bad job. Good job, bad job.”
Then all of a sudden, it gets to the point where it can click on all the things and check out your book really easily. Anything that is verifiable—meaning that there is an outcome to be desired and you can effectively guide it along a path to figure out the answer—is probably a solvable problem.
And you're like, “Well, that works for an Amazon book,” or something like that. Amazon is already blocking all the agentic stuff, by the way. That's going to be a war, I'm sure. But you can say that. The reason people really liked OpenAI to begin with was the original RL thing they did, using reinforcement learning, to make these superhuman Dota 2 players. I was going to use that as an example.
Yeah.
And so these superhuman Dota 2 players are better than humans. You're like, “Okay, well, they won't be able to be as good in these complex things.” I've played enough Dota 3 to be like, “This is really complex, I'm going to be bad at it, and I know I'll never be better at it than these bots.” That's probably enough.
Meaning, you use RL to kick the shit out of a lot of verifiable domains and tasks, and you just scale it to the moon. That's the plan going forward. You'll see multiple iterations—there are multiple roads of progress. Pretraining is slowing down, but it will get a little bit better. Test-time compute is probably not slowing down, but no one really wants to spend the compute budget, so everyone's like, “Give us more compute budget.” All this post-training and RL stuff is definitely not slowing down, and that's where people are excited.
What happens is this new vector kind of hits an asymptote, and then all the stacked vectors get better in aggregate. It definitely is happening faster, meaning the returns to scale are slowing down a lot quicker than in AGI cases. But when you look at the history of improvements in technology, it's pretty damn fast, dude.
I agree with almost everything you said. Let's just say almost everything. I do have one question. You said the returns on investment are slowing. How are they measuring returns on investment? I hear even an OpenAI guy talk about the returns on investment stuff, and I'm like, “How the fudge are you measuring this?” You don't have any. It's not like they have any notable revenue or anything, right?
I'm sure they're getting better, and you'll hear all these guys talking about their returns on investment in AI. The only one I really believe has any is Facebook, and maybe Google, because I know Facebook is using it for a lot of internal training and targeting stuff. But I hear returns on spending coming down while still being very positive, and I'm like, “I don't know how they're measuring it.” So, yeah, what else?
Okay, so this is actually a really good question. InferenceMAX.ai tells you how many tokens you can get and how much they cost. That assumes it's paid.
We made this benchmark at InferenceMAX.ai. It's really cool. It shows AMD versus Nvidia and how AMD is catching up. We're not going to go into that, but the takeaway is that if you could sell the tokens generated on 1 GB200 in a year, Nvidia took this and turned it around in its marketing material: 1 GB200, which costs, let's say, $3.5 million, makes you $5 million a year. That's well over one year payback period.
The raw unit economics of this stuff are extremely high—insanely, incredibly high. The problem is how much you're able to give it to actually get paid for the entire stack. That's the real issue.
Let me push back in one way there. I hear you, because you're renting it out. You take the Nvidia GPU, buy it for $3.5 million, and rent it out to—well, CoreWeave would be the one renting it out—but they're renting it out to some AI startup, right? That's how they're going to generate their return.
I think where people would say, all the classic value investor things, is, “Hey, they're renting it out to an AI startup that has no revenue. So how is that AI startup generating a return on investment?” That's where you kind of get into the dot-com bubble vibe, where it's all just a house of cards, a circular reference blowing up on itself.
Okay, so, yeah, let's talk about it, because I think that, in aggregate, a token that is sold is profitable by a meaningful amount. TSMC makes a 60% margin. Nvidia makes a 75% margin. A neocloud makes a 40% margin just buying those and renting them out. Then all of a sudden, OpenAI makes a negative 50% margin.
So I think if you aggregate it, the tokens sold are actually pretty profitable. Specifically for OpenAI, you're probably right—they can't afford it. But on a unit-economic basis, how do you justify spending all this money? You have to improve the technology; the capabilities have no end.
Going back to that rack example, if you were just to buy a rack, run inference on the models, and sell it to people, even on a subscription-service basis, it's pretty profitable. Assuming it's a DeepSeek model, you can kind of do the entire global infrastructure of OpenAI—all the free users—for, I don't know, probably a few billion, maybe $10 billion.
Let's just say you're supporting 1 billion monthly active users for $10 billion, and you're telling me you can't make $10 billion off 1 billion users? I feel like that's a skill issue. They'll be able to monetize that. That's not a crazy number to me.
Is OpenAI—and I'm asking this from personal experience because I'm actually going to do a post at some point—a lot of apps that I would pay the App Store $10, $50, or $100 per year for, whether it's food tracking, workout tracking, or other stuff, I just put it all into OpenAI for free. Is AI insanely deflationary?
Yes. Yes, I think it is. This is something I've had this whole debate about for a long time. I think AI is insanely, ridiculously, stupidly deflationary, with the exception of if you build—
If you're trying to power your lights.
Yeah. Yeah. Well, yeah, but you have to keep paying for energy.
I think the next leg that makes everything a lot healthier is if you sell things as services, and that's why we've been very bullish on this agentic purchasing thing. My favorite example of how ChatGPT-5 is set up for monetization is a post we did on SemiAnalysis. Pure schizo-posting, straight up: Fidji Simo works at Instacart, gets ChatGPT and Instacart to work together, then leaves Instacart to go work at OpenAI. She's like, “We are definitely going to be monetizing this stuff soon.”
That's an example where I think you can have a massively profitable business that is done on a service basis, isn't deflationary, and has high margins. We're pretty programmed to pay take rates if it's convenient. I want to go buy my entire keto weekly shopping list on ChatGPT and purchase it on Instacart, right?
Let's say it's $100. Instacart will take some platform fees and share some with ChatGPT—let's say 10%—because it does increase all this stuff. I want to have a vision checkup, and it does all that top-of-funnel stuff where you just take a vig on it. Effectively, that's a very valuable thing.
You could be massively deflationary, but it's a service, and you don't deflate the entire economy away. That's the real concern, man. We have these things that are kind of terrifying, but this same thing has happened over and over again throughout history. Railroads were an order of magnitude better than canals.
The classic, right? The queen didn't want to bring in sewing machines because, “What happens to the poor sewers?” Well, guess what? You get with the times or you get run over.
Yeah. Yeah, so it's going to get run over. It's going to be really deflationary, and then on the other side of it, at some point, it blows up and there's all kinds of new stuff. I hope it's not that we're all just talking to our AI girlfriends, and there's something more positive to play than that.
My 8-month-pregnant wife knows I use AI a lot, and she had a friend who actually had a friend who got really into it. She was like, “Please tell me you're not using AI as a girlfriend.” I'm like, “Girl, have you met me?” But I can understand the concerns.
I have 2 more questions on AI, and then let's hop into some of the corporate governance that I think you and I both have a special place in our hearts for.
Yeah. First question. Actually, I'm going to skip the power stuff. I will just tell you guys, I'll ask 1 power question just so I can give this anecdote. A year ago, you and I probably talked once a quarter, but you can tell how highly I think of you because I remember the Oracle thing and I remember this conversation.
I was doing a lot of work on the Bitcoin miner today at Power Play, and you were doing it separately from me. I called you up and said, “Hey, man, I feel like it’s got a long way to run, but these things have really run, and I think people don’t understand how much worse the Bitcoin miner assets are than normal data centers. How much room is there to run?” And you were just like, “Unlimited demand. Buy them all to the moon.”
APLD is about an 8-bagger since then. I think that’s the one we were really talking about. You can pick your share of any one, but I want to give that anecdote. I’ll ask 1 question: How much more room can this power trade possibly have to run?
This is something we had a DM about before we started this, talking about where I think things are maybe offsides in the markets. My belief is that power is much more offsides than semiconductors, because semiconductors are really scalable. Actually, TSMC is really scalable. They can add a ton more capital equipment and probably make double the chips. Taiwan can do that. Taiwan can make double the chips in 2 years. I 100% believe that.
We cannot double the power in 2 years. Just straight up. I don’t think people appreciate or understand that at all. I think that’s where the biggest disconnect comes from when I talk to investors and other people. All the chip guys are like, “Yeah, the orders are good. Nvidia is going to crush it. Taiwan’s going to figure it out.” There are a lot of technology trees, and they’re like, “Yeah, I think we can get to a megawatt per rack.”
You talk to a data center guy, and they’re like, “Are you fucking kidding me? A megawatt per rack? This isn’t real.” And you’re like, “I don’t know what else to tell you, man. Everyone else has bought in except for a data center guy.” There are a lot of really fast-moving data center operators who are doing very, very, very well in the space.
I don’t know if you’ve met Switch, the Rob Roy guy. He’s a billionaire. I remember that IPO very vividly from when I sat at my buy-side shop at Buie Capital. I remember reading it and thinking, “This guy’s a crackhead.” He was like, “We’re going to make high-density racks. That’s the future.” And I was like, “I guess.” He has always been about high-density racks, and he’s been crushing it because the second he heard about this, he was like, “Higher density. In fact, I’m going to 3 megawatts.”
“We’re taking the dials to 11.”
Exactly. He’s literally like, “I’m at 8, and you’re telling me I can go to 11.” He’s going straight there. I think the DLRs of the world are still shell-shocked that we’re at 120 kilowatts, while the really fast-moving players, like Vantage or Crusoe, are all asking, “How do we get to 500 kilowatts per rack?”
To me, 500 kilowatts per rack is a crazy number. My mental math here is that 6 gigawatts is New York City. Everyone is saying that all the chips being purchased will support all this power. We know the power that they should support, and the data center side—the power side—is just not moving fast enough.
There are a lot of ways you can wring efficiency from the grid. The grid is focused on peak versus trough, and utilization is going up. Batteries, backup power, shaving the hottest days or the coldest days out of the year, and using peakers are all working. But even with all that, we’re still talking about a lot of demand.
Just add P × Q—that equals more power from here. The last thing I want to say specifically on this, stock-wise, is that the power miners probably still work from here. Every weird little miner that hasn’t run probably still works, to whatever extent. We’ll say IREN is the craziest, or Applied Digital and IREN are the crazier ones.
The other thing is that the incrementals on industrials interest me the most. One of my favorite companies is probably Comfort Systems. You can look at their print yesterday: revenue goes up 40%, EBIT goes up 80%, and EPS goes up maybe 50%. All of these businesses are growing the fastest in their highest-margin segment, and they’re often massively capacity-constrained.
What that does for EPS is relatively modest, and what I think it does from here, as power accelerates, is that you get acceleration in these businesses with these crazy incremental EPS stories. How you underwrite it is very, very hard.
I’m laughing because I can’t cry on a podcast publicly. I actually looked because somebody was tweeting about Comfort, the ticker FIX, and their earnings this morning. I had notes from about 4 years ago, and the stock price I had in there started with a 9. The stock price today, after this big earnings beat, starts with a 9. There’s just an extra digit in there now. It’s a 10-bagger in 3 years.
But, okay, 2 last questions on this. I seeded these to you, so hopefully you’ve had time to prepare. If not, we’ll just go to the corporate governance. I am a generalist investor. What is 1 thing I, as a generalist investor—and you talk to everyone in the space, from generalists to pod shops to the companies themselves—would not understand or might have a perverse perspective on versus a specialist investor who really spends all their time here?
I still think the power and scale of the infrastructure stuff is really, really, really hard to wrap your mind around. I’ve done enough data center tours to be like—
Like, literally, I do a data center tour and they’re like, “Yeah, there are 100 megawatts in here,” and in my mind I’m like, “This is a baby data center.” The scale really helps you understand how much we’re trying to deploy here.
I also think about Taiwan’s capacity and willingness to make chips. It feels like this entire ginormous industrial revolution is happening, but it’s not happening in the cities. It’s usually in rural areas or outside the core, so no one gets to see it. But tens of thousands of people are working on these data centers and buying GPUs.
You saw the last quarter of GDP growth: 190 basis points of contribution from data centers and GPUs, driving everything. It’s just insane.
Okay, let me ask the question slightly differently. I’m referring particularly to specialist investors. I think you guys talk to a lot of pod shops and tech-focused investors who are probably trading the quarter more than anything else, but they’re very heavy in the weeds here. What is 1 thing that you think specialist investors have a different opinion on versus the industry people—the hyperscalers and the actual people doing this—when you talk to them? Pods versus industry.
I think design lock-in and the inertia of big products are important. We’re probably past the point where a good chip or a good product can win scale on its own anymore. Maybe we can do the HBM drama live, because with HBM4 there’s been all this discussion about whether it’s going to qualify, whether it’s not, whether Samsung is screwed, whether it’s not. Then Sam Altman goes and says, “Hey, I need 50% more.”
I think what’s happening here between specialists and industry is the difference between the absolute and the relative. A lot of people I talk to on the specialist side are focused on, “Well, isn’t this bad for someone else?” And you’re like, “A rising tide is lifting every fucking boat that could possibly be lifted.” Every industry guy is like, “Yeah, I’ll take that total garbage data center power stuff because I need it. I need it. I need it.”
A good example is maybe the Anthropic-Amazon thing. A lot of people are freaking out about it because they’re like, “Well, doesn’t that mean Trainium is screwed?” One of the most galaxy-brain theses is that if they just give up on Trainium and buy GB200s, revenue is going to go through the roof. They need as much as possible, and any supply is going to be massively valuable.
I feel like there’s almost too much relativism going on here. People are saying, “This is going to be good or bad.” We got a little tripped up on this at SemiAnalysis. We were very bullish on Bloom Energy based on time to market, and then we got really deep into the on-site gas space and thought, “On-site gas is so much cheaper and more reliable. Why would you ever do this?”
And the answer is that you’re going to do both because they want both. I think that’s where a lot of specialists have tripped themselves up in the last year, specifically.
I’m not a specialist, but going back to our discussion on energy, I remember having this call with you. I was like, “Look, Applied Digital’s big plant—they’ve got it out in North Dakota, if I remember correctly, right?”
Yeah, it’s got a lot of space, but nobody wants to do North Dakota.
I was ticking through all of these Bitcoin miners and their assets, and I was basically like, “They’re not all A-plus assets, right? This is a B-minus asset.”
This is a D-plus asset, and I think you were rightly like, “Dude, it just doesn’t matter. The demand’s there, and we were early to it.” Every single one that we talked to is, at worst, a 7, and at best—I mean, oh my God—the lost fortunes.
Putting that all to the side, people say I do a lot of stuff, but Doug, I have no clue how you have the time to run SemiAnalysis, be an analyst, and be a corporate governance/shitco master. So, let’s put our corporate governance hats on. I’ve got some questions here, but I’ll just let you cook for a second if you want to talk about anything—corporate governance signals, spring-loading, whatever you want to go with.
Okay. So, I will admit, I do love it. SemiAnalysis has been taking a larger and larger share of my brain and time. Unfortunately, I think my absolute peak in terms of edginess was well over a year ago, because I was really sharpening the edge until then. It’s a little dull, but let’s be real: I love it. I just love boards. What can I say? I just love when the people part of the equation get into it and you can really, really, really see how weird decisions get made.
Doug has some master model for detecting off-cycle PSU grants and RSU grants and everything, and I’m trying real hard to recreate it. I have not had a lot of success, but one day I will come for you on the trackers.
Yeah, please, please do. I mean, I guess then it can be the beta era. The PSU hits, the stock just rips it into the thing, and then we’re good to go.
That was my first question for you, actually. When we’re referring to this corporate governance stuff, we’re talking about companies where, you know, the classic example is that a company normally gives out its share grants in February and then gives out an off-cycle grant in August. Well, guess what? Why are they going to do that? Because they’re going to smash earnings when they report them 2 weeks later.
Have you seen companies start to game investors? You and I would see that and think, “Hey, this is probably a buy,” and buy them. Have you seen companies start to try to game investors by giving off-cycle PSUs to get people excited and then hit them with an equity offering?
I will tell you, with spin-offs about 10 years ago, companies realized investors would buy spin-offs hand over fist, no matter what. I think they started spinning off garbage segments with toxic liabilities to take advantage of that. So, I could see how everything is game theory. I could see how companies respond to this to take advantage. Have you seen any of that?
I think that’s a pretty good observation and something I’ve been thinking about for a long time. Can I mention the event we all went to?
Yeah, please. Anything you want.
Cool, cool, cool. We went to this event in New York. It was cool; it was an activist thing. Thank you, by the way, for the invitation. It was very interesting to see how the investors thought about it, because I was like, “Oh, everyone’s pretty clued in. Everyone’s pretty clued in on the PSUs, man. Show me the incentives, I’ll show you the outcome.” A lot of activists are very focused on throwing in incentives to make sure outcomes happen.
At the end of the day, it’s hard because I think boards are snowflakes, right? Some boards are really, really, really upstanding, and they’re actually trying to practice shareholder-value corporate governance. Some boards are total crooks. Sometimes the compensation packages aren’t even aligned, right? That’s one of the issues that I think is huge: What happens if you’re just getting PSUs and there’s not even performance attached to them? It’s just share units—guys just getting RSUs and getting paid out. You see that occasionally too, and you’re like, “Wait, wait, wait. How does that have any signal?”
There’s this whole problem where board alignment is almost the third thing. How much grift—or how much alignment—do they even care about at the end of the day? There’s a lot of control over some of this, and that’s where I think the art of this comes in.
I’ve seen Hail Mary grants where you’re just like, “Dude, a company is about to go bankrupt, and they’re like, ‘Let’s just give them more out-of-the-money, crazy PSUs.’” Then they just go bankrupt, and you’re like, “Is that the board that has bad judgment? Is it the management team pitching their compensation package to the board? Is it compensation?”
There’s this whole thing where I think there’s a layer of fundamental analysis to be done, because the board could also just be stupid as hell. That happens all the time too. It’s like, “Oh, they’re about to spring-load. It’s definitely coming.” The same board is saying, “Well, we’re going to turn around this company,” even as it’s been slowly eroding into a total shitco. Are you trusting the same judgment of the board that got you there?
That’s one of the issues that makes this become really complicated and fuzzy. I think a lot of the time, to make it a higher-signal game, you really have to focus on when the fundamentals actually line up. If there isn’t a fundamental story and it’s just a board giving PSUs, hoping to say, “Hey, we’re all aligned,” that can get tricky really quickly.
Actually, I’m going to bring this into the semiconductor world, because Hock Tan recently had this ginormous PSU where he’s like, “Dude, I make$10 billion”—or I forget the number. It’s a multibillion-dollar package.
The Broadcom guy.
The Broadcom guy. To be clear, he had these crazy share-price targets that he hit in a little over a year. It was like, “Oh, if you could double the share price—or if you can 2.5x the share price in a year and a half—you get paid out what is now effectively $1 billion.” He did it in a year and a half. All this acceleration, all this stuff.
He goes to the conference and says, “You just saw my compensation package. I’ve got to sell a lot of AI.” That also kind of lines up with the Anthropic thing I was saying earlier, with the TPUs externally. He’s like, “Look, I’ve got to sell a lot of AI revenue ASAP in order for me to get paid.”
You’re in this weird spot where you’re like, “That feels a little too reflexive.” The feedback loop is almost too closed. When this really works, it’s when a board and management team that are really good and understand the challenges in their business try to align the management team on what really matters, to pull the lever to make shareholder value.
Instead, when the board says, “I’m going to essentially pump up a target, give it to the CEO,” and the CEO then goes out into the public to pump it, it’s almost reflexive.
You know which one this reminds me of? I actually had this on my list. Opendoor—I don’t know if you looked at it or not.
Yeah, I’ve seen it.
It goes from $0.90 to $9. They bring in the new CEO, pay the new CEO $1, and give him these RSUs—or PSUs, I’m sorry—that vest at crazy prices: $9, $13, $17, $21, and $33.
If you look at the fundamentals of Opendoor, you’re like, “No way. There’s no way,” unless you’re galaxy-brained—and maybe I’m wrong, because some of these have worked. But I have wondered: Are companies starting to bring the memeification of certain companies into their PSUs and encouraging them, as you’re saying, to go out and be pumpers? The incentives get really perverse and weird.
Yeah, for sure, bro. I mean, the Opendoor CEO is super active on Twitter, or X, right?
And he is pumping the crap out of his stock.
I would argue that if you’re in some of the really aggressive compensation schemes, you’re also pulling a third lever, which is: pump your stock. That could be good or bad. If you have a really cheap, underappreciated asset, maybe pumping your stock isn’t a bad thing. You can argue that’s just investor communications—getting your story out there.
But if you have a really, really, really crappy fake company and you’re like, “I want to put lipstick on a pig and tell you that it could fly. Trust me, bro,” that’s just promoting. I think that’s probably the short-term versus long-term problem.
There are a lot of these compensation packages where I’ve looked at them and been like, “Dude, this was a really good job. This actually made a lot of sense. You created a lot of value.” A really goated board effectively says, “My company is really hurting because of XYZ.”
Probably the best example of this is leverage. I really like the levered companies, because the problem is: What are the external versus internal factors, right? The best possible compensation package you could have is when you have a lever that is internally pullable, and then you essentially award them to pull that lever.
Making a compensation package work on a challenged business is very hard. It’s like, “Okay, I need you to magically accelerate McDonald’s organic sales to double digits.”
Yeah, I don’t think you can do that.
Okay, so I’m completely with you. The other one that sucks is when it’s something that the company so clearly needs to do, and they get paid for doing it anyway. I can’t stand it. I know companies like, “Hey, we have debt that’s due in 18 months. We’re going to get a bonus if we refinance that debt.”
I don’t see why you should get a bonus for doing something like that. The company goes bankrupt if you don’t refinance that debt. Now, if there’s some lever, some magic lever you can pull, but you’re kind of just doing your job. But I hear you.
I understand, because you can argue that that’s just raw incompetence. It’s like, “Do I have to literally have the hamster press every single button to get fed?” Right? It’s like, hey, if you don’t press the button, you die. It should be pretty clear.
It’s almost like, if we’re talking about the participation trophies—the real participation trophies that they’re handing out—these board members are handing out participation trophies to management teams across the nation. Bro, look, earlier this year I went crazy on all these biotechs that were trading below cash, and I was like, all these boards—they own no stock. Many of them—I had some letters that were ready to get real spicy.
I was like, “Hey, guys, you’ve made $3 million as a board member over the past 12 years, and your stock is trading for half of cash, and you own 0 shares. How are you going to tell me that you’re here for shareholder value? You’re not here for shareholder value. You’re going to tell me you’re here for science? You’re not here for science. All your science failed. All you’re here for is—you’re a 74-year-old man. You’re here for a retirement package. That’s all you’re doing, and you’re just trying to prolong it.”
One of my favorite takes is that independent board directors are independent because they’re looking out for themselves.
That’s a really good one. Okay, we’re almost at the end of our time. Again, for those who don’t know, Doug charges quite a bit for his time if you’re a pod-shop analyst, but fortunately he’s a friend, so we get him for a few. Give me one spicy corporate governance example you’re looking at right now.
So many. Should I do my work hat—the semiconductor ones?
No, no. Let’s go with something more fun.
I’m going to parrot one that Non-GAAP just posted. The Elastic one, I think, is really interesting.
ESTC.
ESTC. If you’ve ever been involved with this name, I actually remember the IPO. I remember trying to convince my boss to buy it. It’s been quite a bit of brain damage, but they did this ginormous off-cycle PSU, and I do think that there might be some kind of lever in consumption, specifically focused on AI, that they might be trying to really incentivize—or say that, no, no, no, this search API, or this search, is really going to work out for us.
That one is—I mean, to the extent—I have no fundamental knowledge on that one. I’m trying to think of another one. Do you want to talk about—well, no. I don’t even know enough about this company, Target Hospitality. You schooled me on what the hell they do.
The ticker there is TH. I unfortunately don’t follow it anymore. Go ahead.
Yeah, and Andrew knows quite a bit about it. I think they are a very incentivized management team to get the share price higher. Not only are they extremely incentivized, they’re extremely closely owned by a private equity firm that owns over 50% of the stock.
Historically, private equity is probably one of the hardcore offenders of executive compensation, because if you own the majority of the thing, you can do whatever the hell you want. Who cares about the shareholders? The only shareholder who really matters is the one guy who sits there and really controls the board. Often, for private equity, if they’re a public company and private equity owns the majority of it, they’re looking for a liquidation so they can line up some really interesting and weird things.
I think the thing that I’m most interested in, or kind of focused on, is that they’ve done really aggressive PSUs in the past, and they’ve won some aggressive contracts from the government. They also had a sale process. They tried to almost go private again. You were involved with that.
Yeah, that’s where I got really hurt. They had a sale process. The private equity owner, who owns 60% plus, offered to take them private. They lost not 1 but 2 contracts in the middle of that sale process, if I remember correctly.
The process got pulled. That’s kind of where the pain for me came. But my pain might be Doug’s gain here.
Yeah. So they lost not 1 but 2 contracts, and then they lost, I think, the real one that they just lost—
Their big daddy one. Yep.
The Pecos Children’s Center. That was the really, really painful one. I think the total contract value on that thing at one point was like $2.5 billion. That one got canceled by DOGE.
If I can back up, for those who maybe don’t know—obviously, you and I are very well-versed in it—what happens is they lose the contract. From memory, I think the stock goes from like $8 to $5. Four days later, they announce this giant PSU grant. They give the CEO 2 million PSUs that don’t kick in until the share price hits $20 on the low end and $30 on the high end.
So let’s just say he’s going to hit the high end. They give him $60 million if he can hit this. I can’t remember the timing, but the timing is generally 3 to 5 years on these. They give him $60 million if he can effectively 5x the share price in 5 years.
What Doug is saying is, hey, these pay packages are not awarded willy-nilly, right? They have to be in the works for a while. So they’re probably working on it. These pay packages are not just awarded willy-nilly. You generally don’t want to award a CEO a pay package that they have no shot of hitting, a lot of times, because these are kind of friendly.
You award a CEO a pay package that calls for the stock to go up a lot, and you kind of know in your back pocket there are a lot of things on the come that are going to help get there. So I think that lays out why. Don’t you tell me: They just lost their big Pecos contract, right? What are the things on the come for this company that has lost 2 to 3 big contracts in the past 18 months? What are the big things on the come that could get them to a 5x stock price?
So this is where you have the 2 most secular trends of our time in 1 public company: ICE detentions and data centers in West Texas. That’s the story.
They lost the Pecos facility. As you know, it’s a modular housing play specifically focused on man camps. Their historical business is oilfield services. If you’re familiar with the shale boom, they reaped money then. I think Civeo was public, or someone else was.
I think of Civeo. I think that’s who you’re—
Yeah, I’m thinking of Civeo. I’m thinking of Civeo. They have modular man camps, and you can go look on Google and see them. They’re literally just bunks where people can sleep.
They won this ginormous contract with the government. The government pays about $60 a day, and it’s a little bit lower average daily rate, but it’s pretty good margin—historically, about 50%. They lose this contract, which is the majority of their revenue, but now they have 6,000 beds. They’ve been telling anyone who will listen, “We’re keeping those beds warm. We’re paying the maintenance capex on them. We’re paying the energy. We don’t want the facilities to go bad,” because they’re trying to keep them active until ICE potentially takes them. And ICE has been telling anyone who will listen as well: They want 100,000 beds this year.
100,000 beds is a lot of beds. I think detention beds before OBBB were like 50,000, so they want to double the beds. There aren't many beds quickly available. They have 6,000 ready to go.
OBBBA passed and funded $45 billion for beds. I mean, that's the entire ICE budget, $30 billion of which is attributable to structures. You're telling me you have 6,000 beds available, you have $30 billion of funding to buy some freaking beds, they're ready to go, and you have a management team that's incentivized to win those damn beds.
If you go look, they're reactivating many of these contracts for old prisons they turned offline. For example, I think GEO and CoreCivic announced some new beds on October 1, which is the fiscal new year. There was even an ICE detention center where they're doing over $100 a bed for these detention beds. They'll tell anyone who listens that they think they will win this contract. They're on a special purchasing vehicle for detention beds.
Now we roll into the current problem. ICE was completely out of money year to date. They overspent their entire budget, and on October 1 they could spend some of the budget, but they had to requisition this new asset. This is going to be a totally new contract. The government has been shut down, so this is a weird government-shutdown play. When the government reopens, in theory they should be able to requisition with the $40 billion they have just lying around, saying that they're going to buy beds.
The management team is obviously chock-full of incentives now that this fixes their big contract. I mean, it's going to come in at a much lower rate, which is bad for them, but that's a big, big, big hole at, let's just say, 100% utilization, because these guys will figure it out. I don't think they can sell all of the beds. Actually, some of the beds will be support, and some of the beds will be primary. Even so, we're talking about $80 million to $100 million of revenue a year at 50% EBITDA margins on a run-rate basis. That's like 25% of their revenue today, and it would be a huge swing in the company.
That's the first leg. But I think the thing that gets me more excited, actually, is the data-center opportunity. As you guys know, I'm all over the data-center AI story. They're trying to add 80 gigawatts of power to ERCOT. Eighty gigawatts of power is a lot of power. That's hundreds of data centers, effectively.
These data centers take a lot of HVAC technicians, a lot of electricians, and a lot of builders. My estimate is that, on average, you need about 2,500 workers per gigawatt. A lot of these are in the middle of nowhere. That's the average rate, not peak; peak is a little higher. Let's just say 20 gigawatts of them are in West Texas, which I think is reasonable based on the work I've done. You're talking about 50,000 workers who need to be in West Texas building these freaking data centers. A lot of them are in places so remote and rural that the workforce is larger than the county population. Shackelford is a good example.
Yeah. No, when I was doing the work on this—I remember doing the work last year, when the data-center stuff was really starting to ramp—I was just like, “Hey, there are all these huge data centers getting announced right down the street from basically where these big things are.” At the time, we were worried about government contracts, so could they switch it into housing for the data-center workers? I was kind of like, “I don't know.”
But now it's so in demand. You can go on Google Maps, look at one of their places and then a data center, and you're like, “Oh my God, there's nothing.” People are going to be driving an hour and a half out here, or a company can say, “Hey, we'll pay you guys a rounding error of a rounding error of one cluster of GPUs to house this entire workforce for 6 months.”
Yeah. And that's the plan. To be clear, they've already announced a big contract. That first big contract is in their most recent 8-K. The average daily rate is twice the government rate, so these are really big numbers. Let's say 50,000 people, potentially—they're not going to win all of it—but if they're all $120 a day, that's $2.1 billion of annual data-center housing in West Texas. That's not even counting the rest of everything else.
I think there's just a huge opportunity there. They have all the modular stuff, they have the historical footprint, and they have a massive incentive to figure it out. Effectively, if they don't, this is the biggest bag they've ever fumbled in their entire lives.
Yeah, that's my pitch. But I love stories like that. That's what I really like: where you have a clear opportunity, compensation that really aligns them to the opportunity, and a clear catalyst, which makes a lot of sense. I feel like we have all the letters in the alphabet except X or something, and you just need an X. If they can't figure it out, then shame on them.
It's good. All right. Well, hey, we've got to hop. It's been over an hour. You're sick, but you've thrown in an MJ flu-like-symptoms-type performance here. Dolo, where can people find you? You've got multiple hats on. Where would you like people to pop in?
Follow me on Twitter, Fab Knowledge. That's one way. I still write Fabricated Knowledge when I remember to do it.
As a subscriber, that hurts, man.
Hey, hey, I have some good stuff. I actually have some good stuff in the pipeline. Of course I'm freaking sick, but I think the other thing I was going to say is SemiAnalysis. I'm very, very proud to be part of the team. It's a really cool company. We're building something really special. I think we're going to be a singular research firm of a generation. That's really where I spend all my time.
You guys are, bar none—I mean, I'm just a dumb generalist, but you've built it out, right? You are the go-to place for semiconductor analysis, news sources, analysis, and all that sort of stuff in the space.
And we want to scale to pretty much everywhere AI touches. That could be anything, apparently. So, everywhere. Yeah, it's like when software ate the world—that's the plan. I'm really proud of the stuff we've done already and what we're planning to do. The team's pretty great.
Well, look, thank you so much for coming on. One of the reasons Doug came on is a he's a friend and I told him he had to, but b he's doing a webinar with Alpha Sense who's sponsoring this podcast. If you want to listen to the webinar, hear all of what Doug has to say on the future of AI and semis. Follow the link. You can sign up there. But Doug, this has been awesome. Thanks for enduring the flu and coming on. I'm looking forward to chatting soon.
Yeah, nice seeing you, Andrew. We'll do this soon.
A quick disclaimer, nothing on this podcast should be considered investment advice. Guests or the hosts may have positions in any of the stocks mentioned during this podcast. Please do your own work and consult a financial adviser. Thanks.