Summary

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.

Deep dive

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.”

The Woman Who BEAT Maduro: Venezuela's Next Leader? Nobel Peace Prize Winner María Corina Machado Nvidia CTO Michael Kagan: Scaling Beyond Moore's Law to Million-GPU Clusters