Tech analyst and Stratechery founder Ben Thompson recently voiced a stark warning on the "Invest Like the Best" podcast, saying, "I believe in AI, but I also worry whether we can hold out until it actually generates sufficient returns."

During the podcast, Thompson cautioned that AI infrastructure investment suffers from a severe capital timing mismatch: revenue has yet to arrive, while the financing chain has already exhausted free cash flow, burned through the debt market, and is now moving to equity financing. He drew a direct comparison between the current AI boom and the 19th-century railroad bubble, noting that bubbles may burst, but the underlying technology endures. He also pointed out that Nvidia's high profit margins face hidden erosion, with Google and Amazon emerging as its true long-term competitors, and that if the bubble bursts, the most lasting legacy will be power infrastructure.

The Money Question is More Urgent Than Power or Compute

Current market discussions around AI often focus on whether compute and electricity supplies are sufficient. Thompson argues that both of these questions rank behind a more pressing one: Is there enough money?

"We start with free cash flow, and then tech companies burned through the debt market at an astonishing pace, probably within about a year. Now Google is issuing equity, and Nvidia is piecing together a $500 billion project to tap into pension funds and insurance float. What comes after that? Where does the money come from?"

The ideal scenario is for AI application revenue to feed back into investment, creating a positive cycle. But Thompson points to the problem of "timing mismatch"—infrastructure buildouts take years to translate into revenue, and if capital runs dry before that window closes, we could see a "big bang."

"Even if the explosion happens, AI won't disappear. It will continue to develop," he said. "But for many people, that process will be extremely painful."

The Railroad Bubble Burst, But the Trains Kept Running

Thompson directly compares the current AI investment cycle to 19th-century railroad construction.

In terms of scale, railroad construction accounted for a share of US GDP comparable to current AI capital expenditure, making it one of the largest infrastructure investments in history. In terms of structure, both feature severe "duration mismatches." "Building a railroad and making money from it is a decade-long or even multi-decade endeavor. But you have to issue debt in the short term to pay for construction, and as a result, the world ran out of money."

Regarding the outcome, the bubble burst, but the railroads remained and continued contributing to GDP for a century afterward. "He believes that even with a bubble, what matters is what's left behind after it bursts. The internet bubble left behind fiber optics, and Google built its empire on cheap 'dark fiber.' The most likely legacy of this AI cycle is power infrastructure."

"If all of this collapses and we find ourselves with a huge surplus of electricity, that's actually a wonderful world to live in. We've always been in a state of energy scarcity—energy is the foundation of everything."

Nvidia: Margins Look Solid, But Risk Has Moved Off-Balance-Sheet

Thompson believes Nvidia has so far maintained high profit margins, but there's a neglected structure behind this:

"They use various deals to sustain their apparent margins—like providing a 25% backstop in exchange for new cloud companies committing to buy compute through 2030. Why can these companies get lower financing costs? Because Nvidia is taking on the risk. And taking on risk comes at a cost."

He characterizes this as a disguised price cut: "The price discount isn't reflected in book margins, but if you look at it from an overall discounted cash flow perspective, that discount is real."

Nvidia's real long-term threats come from Google and Amazon:

"They're not just building their own chips; they're also selling them externally. Google has already sold about 20% of its TPU capacity to Anthropic, and Andy Jassy has basically confirmed Trainium will eventually be sold externally. What's their advantage? Lower capital costs. This is a war of capital."

The best-case scenario for Nvidia might be for electricity to become the hard constraint. "Once compute is limited by power rather than chip count, the most efficient chips regain their premium, and Nvidia is still the most energy-efficient supplier per token."

"Power constraints are arriving more slowly than anyone expected, including Jensen Huang himself. During this time, Amazon has more room to improve Trainium, and Google has more time to bring TPU up to competitive levels."

Google Is Like Berkshire: High-Margin Business Funding Bigger Bets

Thompson offers a striking analogy: Google is following the Berkshire Hathaway playbook.

The logic goes like this: Berkshire's See's Candies has extremely high margins, but its scale ceiling is obvious, leaving profits with nowhere to reinvest. So Warren Buffett acquired BNSF Railway—low margin, but massive in absolute profit terms, generating more free cash flow in a single year than See's Candies has accumulated in its entire history.

"Google Search is one of the most perfect business models ever created. Zero marginal cost, no investment needed, everything scales. But AI requires enormous amounts of cash."

He argues that if AI's total addressable market encompasses all white-collar work and everything robots can touch, then even with lower margins, the absolute profit scale could far exceed search. "When we look back then, Google Search will be See's Candies."

Within this framework, Thompson sees Google's equity issuance not as a sign of weakness, but quite the opposite—"You trade a smaller share of a much bigger pie, and ultimately no one will complain." He specifically notes that Berkshire participating in Google's equity offering carries deep symbolic weight: "This is almost literally—railroad money is flowing into Google."

Compute Scarcity Unexpectedly 'Rescues' Intel

Thompson's take on TSMC is that it has inadvertently created its own future competitors.

TSMC has long maintained a highly conservative capacity expansion pace, a strategy that was rational during periods of ample supply—building more capacity means decades of fixed cost drag. But after the AI demand explosion, this conservatism created a severe compute shortage and shifted the risk entirely onto big tech companies:

"Risk doesn't disappear; it just moves. TSMC doesn't want to bear oversupply risk, so tech companies bear the cost of the compute shortage—which is massive revenue they could have earned but didn't."

His key insight: It's precisely this scarcity that makes big tech companies willing to endure the pain of propping up Intel and Samsung's logic businesses.

"Normally, no one wants to go to Intel. TSMC is just too convenient—who would put in that extra effort? But when the shortage is severe enough, and the revenue you're losing due to lack of compute is large enough, you become willing to do it."

He even predicts: "I expect some major company to announce a significant partnership with Intel in the near future, and that will be a big deal. But at the end of the day, TSMC created this situation itself. The cure for high prices is high prices."

Who Has the Deepest Moat Among the Major Players?

Amazon: Thompson's answer is "always Amazon." The core logic is its "first best customer" model—using products internally first, then selling them externally once proven. AWS, Graviton, and Trainium all follow this pattern. "Their core business is almost immune to AI model-level disruption, and their ability to organically incubate new business lines is the most compelling."

Apple: Thompson believes Apple's advantage lies in "owning the user entry point," so suppliers come to them, not the other way around. AI can be integrated on demand. "I support Apple continuing to do what it does best—making great devices." He also raises an interesting risk comparison: Microsoft didn't "miss mobile" so much as treat phones as PC accessories. Could Apple treat the phone as the center of the AI era and miss a larger paradigm shift? "I think it's possible, but doubling down on hardware is also entirely reasonable."

Microsoft: Thompson characterizes its strategy as "a remake of IBM in the 90s"—acting as AI's middleware and implementation consultant, helping enterprises get on board rather than betting on frontier models. Rational, but also fraught with crisis: "Codex, Claude Code—these products are like arrows aimed directly at the heart of Microsoft's core business. Their strategy is sound, and it's desperate, but they might succeed precisely because of that desperation."

Meta: Thompson gives it surprisingly high marks, with the core point being that the advertising business itself is becoming one of AI's biggest beneficiaries. "The biggest AI monetization right now might not be Anthropic or OpenAI, but the incremental gains that Google and Meta's ad systems are capturing." He also criticizes Zuckerberg for never truly confronting the social value of the ad business—"this is a company that seems a bit embarrassed about how it makes money."

OpenAI vs. Anthropic: "Don't underestimate the power of conviction. The most consequential things in history have usually been driven by something religious. OpenAI is like the mainstream church, and Anthropic is like the evangelicals—they're all in, it's their core belief."

What Happens When the AI Boom Runs Out of Money?

In a wide-ranging conversation, Thompson addressed the economic logic, geopolitical landscape, and business models of the AI era. Topics covered include whether massive AI infrastructure investment can deliver returns before capital runs dry, what the railroad boom teaches us about the current investment cycle, why Google increasingly resembles Berkshire Hathaway, how AI commoditizes intelligence, why Amazon has the deepest moat in tech, why Apple's hardware-focused strategy is correct, how AI threatens Microsoft's core business, why the current compute shortage might save Intel, and how Google and Amazon will ultimately become Nvidia's most important competitors.

AI's Funding Dilemma: Lessons from the Railroad Boom

"If you could know one thing in advance to judge the future direction, what would it be?"

Ben Thompson: "What worries me is that as panic grows about model openness, and as related events unfold, the eventual outcome might not be reduced risk, but halted releases, making it increasingly unclear where the true frontier actually lies—creating a false sense of security. Also, the question of recursive self-improvement in AI—could this capability eventually trigger some kind of leap forward? These are things I don't have certainty about right now."

"What I worry about most is the timing mismatch: can actual investment returns generate revenue quickly enough to sustain continued investment? We're moving down the capital curve: starting with free cash flow, tech companies entered the bond market at an astonishing speed, burning through it in about a year; now Google is issuing more stock, and Nvidia is putting together that $500 billion plan to tap pension funds and insurance float. What's next? Where does the money come from?"

"The ideal scenario is returning to a free-cash-flow-driven model. But if we don't reach that point in time, there could be a massive crash. Of course, even with a crash, AI won't disappear—it will continue to develop. Just as we look back at the internet bubble, the railroad boom, or various historical bubbles, those losses seem trivial on the macro scale of human history—though devastating to many people at the time."

"What can the railroad boom teach us? Measured by GDP share, this round of AI investment could be one of the largest infrastructure buildouts in history."

"Railroads had a classic duration mismatch: building a railroad and profiting from it is a decade-long or even multi-decade project; but financing had to be done in the short term, and so the world ran out of money. That's probably why people invoke the railroad story today—we're not just asking whether compute is sufficient or power is sufficient; the more fundamental question is: Is there enough money? It sounds absurd, but that's exactly what happened in the 1870s—the world ran out of money."

"Interestingly, the railroads kept operating, kept developing the West, and contributed profoundly to GDP for generations, right up to today. The money BNSF Railway earns is flowing to Google through Berkshire Hathaway—and that's not a metaphor; it's literally true."

The Berkshire Model and the Google Analogy: "For me, Nvidia's deal and Google's equity issuance are deeply connected—the latter genuinely shocked me when it happened. Why would Google issue equity? Can't it raise money? Why dilute shareholders? Doesn't it believe in itself?"

"The Berkshire analogy is fascinating. They have See's Candies, which has extremely high margins but limited absolute scale, with no room for reinvestment, so cash just accumulates. The genius of BNSF Railway is that they took See's profits and invested them in a much lower-margin business, but with such massive absolute scale that the lower margins still produced more absolute profit than See's entire lifetime earnings combined. When capital gets large enough, you stop thinking in percentages and start thinking in absolute numbers."

"That's why I find this story so interesting—it seems to capture exactly where Google is heading. Google has a search business with astonishing margins, almost the perfect business model—no investment needed, zero marginal cost, pure aggregation. Meanwhile, AI represents an opportunity that requires burning through enormous amounts of capital. But if AI is intelligence, and the market for intelligence covers all white-collar work and eventually, through robotics, almost everything, then even with lower margins, the absolute profit scale will far exceed search advertising."

"In that world, we might look back and call Google Search 'See's Candies.' In that world, you exhaust free cash flow first, then tap hundreds of billions in debt, and finally issue equity. Issuing equity dilutes your percentage of the pie, but if the pie grows to astronomical proportions, no one will complain. Berkshire participating in Google's equity raise carries profound symbolic weight—they're not just Google investors; they're a mirror and model for where Google is heading."

AI's Capability Boundaries: Verifiable vs. Unverifiable Domains

"Setting aside business and competition, just on the technology itself, where do you stand on AI capabilities relative to others in the industry?"

"My view is: extremely optimistic in some areas, relatively conservative in others."

"AI's performance in programming is breathtaking. Thinking back to a year ago when everyone was hand-coding, it's hard to imagine now. Mathematics is similar. But the obvious rebuttal is: these are 'verifiable domains'—you can confirm whether the answer is correct. The real question is: does excellence in verifiable domains cleanly transfer to 'unverifiable domains,' or domains with extremely long verification cycles? That remains to be seen."

"When I raise this question, some people at certain labs respond: 'People thought we couldn't solve chess or Go, and we solved them all.' But my counter is: chess and Go are themselves bounded, knowable domains. Scaling eventually solved them, but they're bounded. The real question is: is there a genuinely open, 'unknowable' domain where AI has accomplished something previously thought impossible? That's where I'm not yet fully convinced."

"That said, an AI trained on all internet data is distillation—a distillation of the final states of human thought, not the process of thought itself. It doesn't capture the traces of thinking, the emotions and experiences behind the writing. If technologies like Neuralink can truly capture the process-level traces of human thinking, that might fundamentally expand these models' capability boundaries. Under that assumption, my concerns about verifiability might be resolved by the answer of 'more data.'"

"My other judgment is that a vast amount of work and economic activity doesn't actually depend on the domains I'm not yet convinced about. In reality, many people are essentially like 'conscious AIs'—they function well in verifiable domains, receiving tasks and executing them. This might be a somewhat pessimistic view of human nature, but the market size is enormous. Even if models stopped improving today, the economic opportunity would still be astronomical."

Aggregation Theory's Extension and Evolution in the AI Era

"You published aggregation theory in 2015, which basically defined the winners and losers of that technological era. What principles or theories do you think will define the winners in the AI era?"

"That's a good question. I've been weighing whether aggregation theory still applies equally."

"The core of aggregation theory is zero marginal cost. One key dimension is distribution: in a world of abundant content supply, what's scarce isn't distribution but discovery. Companies that solve the 'discovery' problem dominate their markets, creating a positive flywheel. Another key element is transaction costs—the vast majority of Google and Meta's advertisers have never interacted with a human salesperson; everything is done by machines, which is the perfect zero-marginal-cost business model."

"AI clearly changes this equation—inference costs are real and cannot be ignored. But the question is, how 'heavy' is this cost?"

"My observation is that most AI users today use AI as a Google replacement or a recipe generator, and serving these users costs very little—not much different from serving them a webpage. But at the other extreme, users who truly leverage test-time compute scaling—they let models spend more time thinking about a problem, which directly equates to marginal cost: every extra second of thinking costs an extra cent. So the question 'does AI have marginal costs?' has answers that aren't even in the same universe for different users."

Implications for Microsoft's pricing model: "Microsoft launched an enterprise plan—$100 per user per month, with a usage allowance and overage charges. This strategy is actually quite a dilemma. Microsoft's core value proposition is: everything you need is bundled together, uniformly priced, and it all works together. Once you introduce usage-based billing, you break that logic."

"First, it decouples from headcount. In the past, buying a license for a new employee at $50 or $100 per month was a 'no-thought revenue stream'—it was already built into the cost of hiring."

"Second, usage-based billing requires monthly decisions. But most companies budget annually, and this 'watch the bill every month and decide how much to spend' model doesn't fit corporate budgeting logic at all."

"Third, once you're looking at the bill monthly, you start asking: is every dollar worth it? Are these products good enough? Should we spread the budget elsewhere?"

"Microsoft is doing this because heavy AI users cost far more than $100 per month, so they had to adjust. But they don't want a large number of ordinary users to start thinking deeply about this, because that would fundamentally undermine their business model."

On AI consumer business models: "Silicon Valley relearns two things every decade. First, consumers don't want to pay for software. Second, consumers don't care about productivity."

"Dropbox is the classic case: a stunning consumer product that ultimately had to be completely rebuilt and pivoted to enterprise, because consumers simply wouldn't pay. Enterprises pay because employee time has value—improving employee efficiency is improving ROI. Consumers are the opposite: they're already exhausted from work, and when they get home, they just want to scroll short videos."

"OpenAI made the same mistake, just magnified a hundredfold: insisting on selling subscriptions to consumers. They've sold a lot, but not nearly enough. If you're in the consumer market, advertising is the correct business model."

"OpenAI is finally getting into advertising now, but the timing is a bit odd—at the same time, they've realized they must capture the enterprise market because Anthropic is competing fiercely. If they had jumped into advertising immediately after ChatGPT went viral, I think they'd have a killer ad product today, and Google and Meta would be under much more pressure."

"The fundamental advantage of the advertising model for consumer platforms is that the pressure to raise prices falls on advertisers, not users, so there's no demand elasticity problem. When Netflix raises subscription prices, they worry about users downgrading or churning; but with advertising, giving users free access actually makes it easier to grow."

Compute, TSMC, and Intel

"This year's capital expenditure is around $800 billion, and next year it's projected to be $1.3 trillion, still climbing, and newly built compute capacity is absorbed almost as soon as it comes online—demand fully keeps up with supply."

"Several timing mismatches are happening simultaneously. The 'not enough compute' narrative stems from underinvestment in 2023 and 2024, which is correct. But TSMC has been reducing capacity expansion speed throughout 2023, 2024, and 2025. That means today's compute gap will only worsen in the coming years, because the lead time for building a fab is much longer than building a data center."

"All the money being invested today won't turn into compute tomorrow—it will materialize in 2028 or 2029. So we hear Andy Jassy and Satya Nadella say, 'We're building data center shells, and we'll procure GPUs when there's actual demand'—that's a nice story, but I don't quite believe it. Once you've committed fixed costs, how can you just let them sit idle?"

"This brings us to commodity market logic, which the tech industry generally lacks understanding of."

"Take shipping: once you buy a ship, the cost is depreciation, and marginal costs are just fuel, crew, and port fees—very low. That means you'll definitely run that ship at full capacity, keeping it running even at prices that only cover marginal costs, even if it shows accounting losses—because depreciation is just an accounting concept; the money's already spent. Containers are pure commodities, priced by market marginal costs. During the pandemic, container prices went from $3,000-4,000 to $17,000-18,000, and shipping companies made fortunes—but everyone simultaneously ordered new ships, and two years later, a flood of new vessels hit the market, prices crashed, and another bloodbath followed."

"The memory industry is even more so, having gone through countless boom-bust cycles. Samsung's rise came precisely from counter-cyclical investment during downturns, expanding aggressively while Japanese players hesitated, and then sweeping competitors aside when the next cycle arrived. The market eventually consolidated to three oligopolists, and everyone stopped repeating past mistakes, forming a sort of 'unspoken understanding.'"

"But that logic collides directly with the current moment: it took them a long time to realize that AI represents a structural change in memory demand, not an ordinary cyclical fluctuation. Memory problems will eventually be solved, but memory manufacturers are now making a strategic error—turning themselves into a giant target. Apple is lobbying to bring in Chinese memory, and one of the hottest research areas is reducing memory usage. It's like Iran and the Strait of Hormuz: a blockade is a card that has maximum deterrent power when not played—once you play it, people will find ways around it. Try to blockade in 2035, and UAE and Saudi pipelines and ports will already be built. I worry memory makers have maneuvered themselves into the same position."

On TSMC: "TSMC's problem is more severe because it's the only player at the leading edge. TSMC's core risk-avoidance logic is: if there's overcapacity, it's not just wasted fixed costs—it locks in excess capacity for thirty years. So they're naturally biased toward conservatism."

"But they've actually made the same mistake as memory makers, and the same mistake as Iran—they haven't expanded aggressively in the past few years. This compute shortage has become severe enough that big tech companies are giving up enormous revenue and profits. When that pain becomes real enough, there's an economic incentive to do things they wouldn't otherwise do—help Intel and Samsung's logic businesses catch up faster."

"I've been writing about Intel since 2013: you must do foundry. By then it already felt late, but their stock kept soaring on the cloud computing wave, and they didn't truly recognize the gap until 2020. Intel's fundamental problem is: no customer-service mindset, culture, organization, or full IP portfolio, and they can't find initial customers willing to bear the friction of working with them—going to TSMC is just so much easier."

"But it's precisely this extreme compute scarcity that's forcing big tech companies to become willing to bear the pain of getting Intel back on track. That's the logic of 'high prices self-heal': scarcity creates economic incentives for people to do things they wouldn't otherwise do. I expect a heavyweight partnership announcement in the near future, and that will be a milestone."

Amazon and Apple's AI Moats

"Among the top fifteen tech companies, which ones do you find most interesting today?"

"The answer is always Amazon."

"Amazon is remarkable because they always make themselves the first and most important customer. They use their scale to make anything work, then sell it to others. AWS is the classic example. Contrary to popular belief, AWS wasn't Amazon's idle capacity—it came from a profound internal insight: our internal retail team needs a scalable, pure-API compute infrastructure, no meetings, no negotiations, the interface is just there. If it works internally, it works externally. Actually, AWS served external customers before internal teams, but ultimately it served both."

"Amazon's logistics is another side: first use third parties (UPS, FedEx, USPS), then build in-house, then open to others. Marketplace, Graviton chips, Trainium chips—all the same model: be your own first and most important customer, refine the product through internal use, and when it's good enough, take it to the external market. Their AI products are following the same path—not all will succeed, but the methodology is extremely elegant."

"Amazon's core business—e-commerce and logistics—is deeply rooted in the physical world, making it inherently difficult for AI to replace, and it will only benefit from AI. I believe their moat depth exceeds that of any tech company."

On Apple: "Apple seems to be sitting on the sidelines in this AI wave, which might be a case of 'luck is better than being good.' But Apple has a complete ecosystem, and most importantly, they have direct access to users—that's the classic aggregator advantage: you own the user, so suppliers come to you. They can integrate external AI providers on demand."

"From a consumer demand perspective, people don't want to increase productivity; they just want a working assistant—a genuinely good Siri. This can be done entirely on-device, and Amazon doesn't even need to bear inference costs because they're using the user's own battery."

"Apple's true advantage is that the smartphone itself is nearly perfect. Small enough to fit in a pocket; big enough to view any content; capable of running your entire digital life. TVs have become accessories; everything is on your phone. I don't see anyone displacing the phone's central position."

"The real question is: will the phone always be the center? In the home environment, 'ambient computing'—where you just talk to AI and it tells you what you need to know—is a genuine opportunity, and Apple has the best foundation here, provided they don't over-rely on their own frontier model capabilities."

"There's also a Microsoft-style trap here: Microsoft didn't miss mobile internet—they entered early—but their mobile products were 'shrunken PCs' because they always assumed the PC would be the center. Apple recognized that the phone isn't an accessory to the Mac; the phone is the phone—they started building this understanding from iPod and iTunes."

"Could Apple repeat this mistake—assuming the phone is always the center and adapting AI around it? Or is AI truly ubiquitous, presentable through phones, devices, and computers alike, and would that actually be disruptive to Apple? It's possible. But I also think Apple doubling down on what it does best is entirely reasonable."

"One final point: AI is a probabilistic endeavor, while Apple is the king of deterministic products. Launching an iPhone, you only get one shot—you have to get it right the first time; any quality issue costs billions. Apple has never recalled a single iPhone. Behind that is a completely different decision-making culture and supply chain philosophy, entirely distinct from what's needed to build great AI. I tend to let companies focus on what they're good at—Apple continuing to make great devices is perfectly reasonable."

The Frontier AI Player Landscape

"Among the five frontier AI competitors—OpenAI, Anthropic, Gemini, xAI, and Meta—which one has the most interesting position?"

"OpenAI and Anthropic have the highest risk, but also the largest upside. Don't underestimate the power of conviction—the most influential causes in history have typically been driven by some kind of 'religion.' Silicon Valley has two 'religious organizations': OpenAI is like the mainstream church, attending regularly every Sunday; Anthropic is like fervent evangelicals, with belief internalized to the bone—it's their core reason for existence."

"Google just needs search not to die too quickly; Meta has a massive advertising business as a backstop, but if Meta were run by anyone other than Mark Zuckerberg, they wouldn't be at the frontier. It's one of the purest manifestations of founder energy."

"Meta building a new team from scratch and doing AI from zero is honestly quite crazy, but hats off to Zuckerberg. And there's a deeper logic: for a purely digital company, not being at the frontier is actually the more reckless choice."

Microsoft and the cautionary tale: "Microsoft had $20 billion in free cash flow last quarter and paid $10 billion in dividends, but they're not at the frontier. Their strategy is: provide a middleware layer across all models, help enterprises build on their platform, manage model changes and iterations, and provide backward compatibility. That's IBM's playbook from the 90s."

"When Gersner took over IBM, his insight was: IBM was mediocre at everything. Breaking it up wasn't the answer because the individual parts were worth even less. IBM's greatest asset was its 'bigness'—the internet was coming, and every company needed to get online but didn't know how, so IBM became the integrator: building a middleware layer between enterprise legacy mainframes and modern web services, and building out a massive consulting arm alongside. That bought IBM another thirty years."

"Microsoft is following exactly this path: we'll help you with AI, help you transform without handing over your core data, we're the stable, reliable platform. It's a rational strategy, but it's also a defensive posture born of necessity."

"Microsoft's existential threat: Microsoft's products are essentially 'the interface between users and computers.' And products like Codex and Claude Code are arrows aimed directly at Microsoft's heart. Long-term, all digital companies face this threat, but Microsoft is the most direct target. Their strategy is rational, but it's also a last stand—and precisely because of that, they might succeed."

On xAI: "The space data center theory is interesting, but if that's truly their core competency, I'm not sure they need their own model. If so, why burn tens of billions on model development before that? The fact that they're selling compute to Anthropic is itself quite telling. The Cursor acquisition was a good move—the two companies are highly complementary."

Nvidia and the Commoditization of Intelligence

"How do you connect the lack of understanding of commodity markets to Nvidia and the commoditization of intelligence?"

"Nvidia's current profit margins, I believe, are not natural. Let me explain why."

"There's a lot of discussion about 'circular financing' right now—Nvidia providing 25% capital backing in exchange for equity in new cloud companies that commit to buying Nvidia compute through 2030. On the surface, Nvidia maintains its prices and margins; but in substance, they're taking on risk. If those new cloud companies can't sell their compute capacity, or if hyperscalers' own capacity is more than sufficient, Nvidia ends up paying for compute nobody needs."

"This investment has an expected value—not zero, not 100%, but somewhere in between. That expected value is effectively a discount on Nvidia's margins—though that discount doesn't appear in the reported margin figures, it genuinely exists in the company's overall discounted cash flow analysis."

"This is what I mean by 'price discounts manifesting in strange ways.' Generally speaking, moving things off the balance sheet works in practice most of the time; but you need to see the full picture, not just gross margins."

"Nvidia's ultimate challenge: Nvidia's real competitors aren't AMD, but the hyperscalers, especially Google and Amazon. They're not just building their own chips—they're selling chips to external customers. Google has already sold 20% of its TPU capacity to Anthropic, and Andy Jassy almost confirmed on the last earnings call that Trainium chips will eventually be sold externally. Their logic is clear: this isn't a differentiation play, it's commodity sales, and it won't cannibalize their cloud business appeal while improving R&D leverage."

"The core advantage hyperscalers have over new cloud companies is lower capital costs. This is a war of capital costs. Elon says 'we'll always buy Nvidia'—the real reason is that Nvidia is the most flexible and easiest to sublease, and CUDA's moat has also significantly weakened because models themselves don't care what hardware they run on."

"Nvidia's real hope might lie in electricity becoming the bottleneck. If power is truly severely constrained, users will seek maximum token efficiency, and Nvidia is still the most efficient choice—that would be favorable for Nvidia."

"But what disappoints Nvidia is that the US has actually brought more power online than expected in recent years—Elon's behind-the-scenes solar, West Texas natural gas, nuclear plant restarts—these developments have gone surprisingly smoothly. And that gives Google more time to refine TPU efficiency, and Amazon more time to improve Trainium. If we end up in a world where electricity isn't the bottleneck, maintaining those profit margins will be extremely difficult."

After the AI Bubble: What Will Remain?

"What I've been thinking about throughout the AI boom is: after the bubble, what will remain?"

"The internet bubble left behind fiber optics. Google's rise was largely due to buying massive amounts of dark fiber at bargain prices after the bubble—today's internet backbone still runs on WorldCom's fiber. The railroad boom left behind BNSF, still operating today, and its money is flowing to Google. Good bubbles always leave behind something truly lasting."

"What can the AI bubble leave behind? GPUs won't last long, data centers can, but most importantly—electricity."

"If all of this collapses and leaves us with a massive new supply of electricity, that would be an incredibly valuable legacy. We've always lived under energy scarcity constraints; energy is the foundation of everything. What it would be like to live in an energy-abundant world is hard for us to imagine, because our thinking has been framed by energy scarcity."

"I think America has delivered a stunning performance on this front. Electricity is already a bottleneck, and will continue to be, but it's become a real constraint much later than anyone expected—including Jensen Huang himself, who I suspect thought power shortages would become Nvidia's moat much earlier."

"This is one of the most encouraging signals of our era: watching America tackle these challenges genuinely lifts my spirits."