During a recent episode of the Invest Like The Best podcast, tech analyst and Stratechery founder Ben Thompson shared a striking sentiment: "I believe in AI, but I worry whether we can hold on long enough for it to actually generate sufficient returns."

Thompson issued a warning about a serious temporal mismatch in AI infrastructure investment: revenue streams have yet to materialize, yet the financing chain has already exhausted free cash flow, blown through debt markets, and is now turning to equity financing. Drawing a direct parallel to the 19th-century railroad bubble, he suggested that while a bubble may burst, the technology itself tends to endure. He also pointed out that Nvidia's high profit margins are facing hidden erosion, with Google and Amazon emerging as its true long-term competitors. Should a crash occur, he argues, the most lasting legacy could be the power infrastructure built along the way.

The Money Question Looms Larger Than Power or Compute

Current market discourse on AI heavily focuses on compute capacity and electricity availability. Thompson contends that both of these concerns are secondary to a more pressing issue: the availability of capital. "We started with free cash flow, and then tech companies burned through the debt market at an astonishing rate, taking about a year. Now Google is issuing equity, and Nvidia is stitching together a $500 billion project to tap into pensions 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. However, Thompson highlights the core problem of "temporal mismatch"—infrastructure build-outs take years to translate into income. If capital dries up before that window arrives, a "big bang" could occur. "Even if it explodes, AI won't disappear. It will keep developing," he said. "But for many people, that process will be very painful."

The Railroad Bubble Burst, Yet the Tracks Remained

Thompson directly compares the current AI investment cycle to the railroad construction boom of the 19th century. In terms of scale, railroad construction as a percentage of US GDP at the time is roughly equivalent to today's AI capital expenditures, making it one of the largest infrastructure investments in history. Structurally, both share a severe "term mismatch."

"Building a railroad and making money from it is a decade or multi-decade undertaking. But you must issue debt to fund construction in the short term, and as a result, the world ran out of money." When considering outcomes, he notes that the bubble burst, but the railroads remained, contributing to GDP for the following century. He argues that what matters is what a bubble leaves in its wake. The internet bubble left behind fiber optics; Google built an empire on cheap "dark fiber." The most probable legacy of this AI cycle is electrical infrastructure. "If it all collapses and we find ourselves with a massive surplus of electricity, that's actually a very good world to be in. We've always been in a state of energy scarcity; energy is the foundation of everything."

Nvidia's Margins Look Strong, But Risk Has Shifted Off-Balance-Sheet

Thompson believes Nvidia's high margins persist, but they conceal a significant structural dynamic: the company is engaging in various transactions to maintain its apparent profitability—for instance, offering a 25% guarantee in exchange for new cloud companies committing to purchase compute power until 2030. Why can these companies secure lower financing costs? Because Nvidia is absorbing the risk, and taking on risk comes at a cost.

He characterizes this as a disguised form of price reduction: "The discount isn't reflected in the book margins, but if you look at it from a discounted cash flow perspective, that discount is very real." Nvidia's genuine long-term threat comes from Google and Amazon. They are not just building their own chips; they are also selling them externally. Google has already sold about 20% of its TPUs to Anthropic, and Andy Jassy has essentially confirmed that Trainium will eventually be sold to outside customers. Their key advantage is a lower cost of capital. This is a war of capital.

What Nvidia might most hope for is for electricity to become a hard constraint. If compute becomes limited by power rather than chip supply, the most efficient chips would regain their premium, and Nvidia remains the most efficient provider per token. However, power constraints are emerging slower than anyone anticipated, including Jensen Huang himself. This gives Amazon more time to refine Trainium and Google more time to bring TPU to competitive levels.

Google is Becoming the Berkshire Hathaway of AI

Thompson proposed a compelling analogy: Google is following the Berkshire Hathaway playbook. The logic is as follows: Berkshire's See's Candies boasts extremely high margins but has a clear ceiling on scale, leaving profits with nowhere to reinvest. So Warren Buffett acquired BNSF Railway—a business with lower margins but enormous absolute profit, generating more free cash flow in a year than See's Candies has 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 demands massive capital. Thompson argues that if AI's total addressable market encompasses all white-collar work and everything robots can touch, then even at lower margins, absolute profit could far exceed search. "At that point, we'll look back and see Google Search as See's Candies."

Within this framework, he views Google's equity issuance not as a sign of weakness, but quite the opposite—"you trade a smaller share of something for a much bigger pie, and in the end, no one will complain." He specifically notes the profound symbolism of Berkshire Hathaway participating in Google's equity offering: "This is almost literally—railroad money flowing into Google."

Compute Scarcity Unexpectedly Rescues Intel

Thompson's assessment of TSMC is that it inadvertently created its own future competitors. TSMC has maintained a highly conservative capacity expansion strategy, which was rational in times of abundant supply—building more capacity means decades of fixed-cost drag. But after the AI demand surge, this conservatism has created severe compute shortages, shifting all the risk onto big tech companies: "Risk doesn't disappear; it just transfers. TSMC didn't want to bear the overcapacity risk, so tech companies are now paying the price for compute scarcity—that's a lot of money they could have earned but didn't."

His key insight: it is precisely this scarcity that is motivating large tech companies to endure the pain of supporting Intel and Samsung's logic businesses. "Normally, no one wants to go to Intel. TSMC is just too easy to work with; why would anyone bother with the extra effort? But when the shortage is severe enough, and you're losing enough revenue for lack of compute, you're willing to do it." He even predicts: "I expect a major announcement of a significant partnership with Intel in the near future; that will be a big event. But ultimately, TSMC created this situation itself. The cure for high prices is high prices."

Analyzing the Moat Landscape of Major Players

Regarding Amazon, Thompson's answer is "always Amazon." The core logic lies in its "first-best customer" model—using its own products first, then offering them to the market once they work. This pattern applies to AWS, Graviton, and Trainium. "Their core business is almost immune to AI-level disruption, and their ability to organically incubate new business lines is the most compelling."

On Apple, Thompson believes its strength lies in "owning the user interface," which means suppliers come to them rather than 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 raised an interesting risk comparison: Microsoft didn't "miss mobile" so much as treat phones as PC accessories; could Apple make the mistake of treating the phone as the center of the AI era, thereby missing a larger paradigm shift? "I think it's possible, but doubling down on hardware is also entirely rational."

He characterizes Microsoft's strategy as a "1990s IBM remake"—acting as middleware and implementation consultants for the AI era, helping enterprises adopt rather than betting on frontier models. It's rational but fraught with danger: "Codex, Claude Code—these products are arrows aimed directly at the heart of Microsoft's core business. Their strategy is sound, desperate, but might succeed precisely because of that desperation."

Thompson gives Meta unexpectedly high praise. The core point: the advertising business itself is becoming one of the biggest beneficiaries of AI. "The biggest AI monetization right now might not be Anthropic or OpenAI, but the incremental gains seen by Google and Meta's ad systems." He also criticizes Mark Zuckerberg for never truly confronting the social value of the ad business, describing Meta as "a company a little embarrassed about how it makes money."

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

The Full Interview: When AI Hype Runs Out of Money

Program: Invest Like The Best | Release Date: August 19, 2026 | Duration: 1 hour 25 minutes 53 seconds

In this episode, Ben Thompson engages in a deep conversation about the economic logic, geopolitical landscape, and business models of the AI era. Topics cover: whether massive AI infrastructure investment can deliver returns before capital runs out, 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 current compute shortages might save Intel, and how Google and Amazon will ultimately become Nvidia's most significant rivals.

AI's Funding Dilemma: Lessons from the Railroad Boom

Host: If you could know one thing in advance to predict the future direction, what would it be?

Ben Thompson: My worry is that as panic over model openness escalates and related events unfold, the outcome won't reduce risk but stop releases, making it increasingly unclear where the true frontier lies—creating a false sense of security. Also, the question of AI's recursive self-improvement—could that capability eventually trigger some kind of leap? These are things I'm not certain about. What worries me most is the temporal mismatch: can the returns on investment generate revenue fast enough to sustain continued spending? We're moving down the capital curve: starting with free cash flow, tech companies stormed into the bond market, burning through it in about a year; now Google is issuing 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? Ideally, we return 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 if it crashes, AI won't disappear; it will keep developing. Just like we look back at the internet bubble, the railroad boom, or any historical bubble, on a macro-historical scale those losses seem trivial—though devastating for many people at the time.

Host: What lessons can the railroad boom offer? By GDP share, this AI investment cycle could be one of the largest infrastructure builds in history.

Ben Thompson: Railroads had a classic term mismatch: building a railroad and profiting from it is a decade-to-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, but more fundamentally: is money sufficient? It sounds absurd, but that's exactly what happened in the 1870s—the world's money ran out. Interestingly, the railroads kept operating, kept developing the West, and contributed enormously to GDP for a long time, even to this day. The money BNSF Railway earns is, through Berkshire Hathaway, flowing into Google—and that's not a metaphor, it's a literal fact.

On the Berkshire model and the Google analogy: To me, that Nvidia deal and Google's stock issuance are deeply connected—the latter genuinely shocked me. Why is Google issuing equity? Can't it raise money? Why dilute shareholder equity? Doesn't it believe in itself? The Berkshire analogy is interesting. They have See's Candies, with extremely high margins but limited absolute scale, no room for reinvestment, just accumulating cash. The genius of BNSF Railway was investing See's profits into a business with much lower margins, but of such enormous absolute scale that the lower margin yielded more absolute profit than See's total profit over its entire lifetime. 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 perfectly capture where Google is headed. Google has a search business with incredible margins, almost the perfect business model—no investment, zero marginal cost, pure aggregation. At the same time, AI represents an opportunity that requires burning through enormous capital. But if AI is intelligence, and the market for intelligence covers all white-collar work and, eventually through robots, almost everything, then even at lower margins, the absolute profit scale would far exceed search advertising. At that point, we might look back and call Google Search 'See's Candies.' In that world, you exhaust free cash flow, tap hundreds of billions in debt, and finally issue stock. Issuing stock dilutes your percentage of the pie, but if the pie becomes astronomically large, no one will complain. The fact that Berkshire participated in Google's issuance is deeply symbolic—they're not just Google investors; they're a mirror and model for Google's future direction.

AI's Capability Boundaries: Verifiable vs. Unverifiable Domains

Host: Setting aside business and competition, just on the technology itself, where does your assessment of AI's capabilities sit relative to others in the industry?

Ben Thompson: My view is: extremely optimistic in some areas, relatively conservative in others. AI's performance in coding is remarkable. Thinking about how a year ago everyone was writing code by hand, it's hard to imagine now. Same with mathematics. But the obvious rebuttal is: these are all 'verifiable domains'—you can be sure whether the answer is right. The real question is whether excellent performance in verifiable domains cleanly transfers to 'unverifiable domains,' or those with very long verification cycles. That remains to be seen. When I raise this, some people in certain labs respond: 'People thought we couldn't solve chess or Go, and we did.' But my counter is: chess and Go are bounded, knowable domains; scaling eventually solved them, but they're bounded. The real question is: is there a truly open, 'unknowable' domain where AI has accomplished something previously thought impossible? That's what I'm not fully convinced about. That said, an AI trained on all internet data is distillation—a distillation of the end-state of human thought, not the process itself. It doesn't capture the trace of thinking, the emotions and experiences behind writing those words. If technologies like Neuralink can truly capture the processual traces of human thought, it might fundamentally expand these models' capabilities. Under that assumption, my concern 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 those domains I'm not yet convinced about. In reality, many people are essentially like 'conscious AIs'—they function well in verifiable domains, receive tasks, and execute them. This might be a slightly pessimistic view of human nature, but the market size is enormous. Even if models don't improve from today, the economic opportunity is astronomical.

The Aggregation Theory's Extension and Evolution in the AI Era

Host: You published Aggregation Theory in 2015, which basically defined the winners and losers of that tech era. How do you now view what principles or theories will define winners in the AI age?

Ben Thompson: That's a good question. I myself am constantly weighing whether Aggregation Theory still applies equally. The core tenet is zero marginal cost. One key dimension is distribution: in a world of extremely abundant content supply, what's scarce isn't distribution but discovery. Companies that solve the 'discovery' problem dominate their market and create a positive flywheel. Another key is transaction costs—the vast majority of Google and Meta advertisers have never interacted with a human salesperson; everything is done by machines, a 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 it merely as a Google replacement or a recipe generator; serving these users costs very little, almost the same as serving them a webpage. But at the other extreme, users who truly leverage test-time compute scaling—they let models think longer on a problem, which directly translates to marginal cost: every extra second of thought costs more money. So, the question 'does AI have marginal costs?' has answers that aren't even in the same universe for different users.

On Microsoft's pricing model implications: Microsoft launched an enterprise plan—$100 per user per month, including a certain amount of usage, with overage charges. This strategy is quite a dilemma. Microsoft's core value proposition is: everything you need is bundled together at a flat price, 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 was $50 or $100 a month, an 'unthinking revenue stream'—it was already built into hiring costs. Second, usage-based billing requires monthly decisions. But most companies budget annually; this 'decide monthly how much to spend' model doesn't fit corporate budget logic. Third, once you look at the monthly bill, you start asking: is each dollar worth it? Are these products good enough? Should we spread the budget around? Microsoft does this because heavy AI users cost far more than $100 a month, so they have to adjust. But they also don't want ordinary users to start thinking deeply about this, because it would fundamentally undermine their business model.

On consumer business models for AI: Silicon Valley has to relearn 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: an amazing consumer product that ultimately had to be completely rebuilt and repositioned as an enterprise product because consumers simply wouldn't pay. Enterprises pay because employee time is valuable—improving employee efficiency is improving ROI. Consumers are the opposite: they're tired after work and just want to lie down and scroll short videos. OpenAI made the same mistake, just a hundred times bigger: insisting on selling subscriptions to consumers. They sold a lot, but nowhere near enough. If you're in the consumer market, advertising is the correct business model. OpenAI is finally doing ads now, but the timing is odd—at the same time, they realize they must capture the enterprise market because Anthropic is competing fiercely. If they had jumped into advertising immediately after ChatGPT's viral success, I think they'd have a killer ad product now, and Google and Meta would be under more pressure. The fundamental advantage of the ad model for consumer platforms is that the pressure of price increases falls on advertisers, not users, so there's no demand elasticity problem. Netflix worries about downgrades when it raises subscription prices; but with ads, giving the product away free makes it easier to scale.

Compute, TSMC, and Intel

Host: Capital expenditure this year is around $800 billion, next year projected at $1.3 trillion, and it keeps climbing. Newly built compute capacity is absorbed almost as soon as it comes online; demand fully matches supply.

Ben Thompson: Several temporal mismatches are happening simultaneously. The claim that 'compute isn't enough' stems from underinvestment in 2023 and 2024, which is correct. But TSMC has been reducing capacity expansion from 2023 through 2025. That means today's compute gap will only worsen over the next few years because a fab's lead time is much longer than a data center's. All the money invested today won't become compute tomorrow, but in 2028 and 2029. So we hear Andy Jassy and Satya Nadella saying, 'We're building data center shells and will procure GPUs when there's actual demand'—that's a nice story, but I don't fully believe it. Once you've invested the fixed cost, how can you just leave it idle? This is where commodity market logic comes in, and the tech industry generally lacks understanding of it. Take shipping: buy a ship, the cost is depreciation, and marginal costs are just fuel, crew, and port fees—very low. This means you'll always run the ship, pricing low enough to cover marginal costs, even if the books show losses—because depreciation is just an accounting concept; the money's already spent. Containers are pure commodities; prices are set by marginal cost in the market. During the pandemic, container prices went from $3,000–4,000 to $17,000–18,000, and shipping companies made a fortune—but everyone simultaneously decided to build ships, and two years later, a flood of new ships hit the market, prices crashed, and there was another bloodbath. The memory industry is even more so, having gone through countless boom-bust cycles. Samsung's rise was exactly through counter-cyclical investment during downturns, expanding aggressively when Japanese players didn't dare, and sweeping competitors when the next cycle came. The market eventually consolidated into three oligopolies, and everyone stopped repeating past mistakes, forming an 'unspoken understanding.' But this logic has directly collided with the current moment: it took them a long time to realize that AI represents structural memory demand changes, not ordinary cyclical swings. Memory issues will eventually be resolved, but memory makers are now making a strategic mistake—turning themselves into a huge 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 whose deterrent power is greatest when unused; once deployed, others find ways around it. Try to blockade in 2035, and the pipelines and ports in the UAE and Saudi Arabia will already be built. I worry memory makers have backed themselves into the same corner.

On TSMC: TSMC's problem is more severe because it's the only player in leading-edge processes. TSMC's core risk-avoidance logic: if there's overcapacity, it's not just wasted fixed costs; it locks in excess capacity for thirty years. So they're naturally conservative. But they've actually made the same mistake as memory makers, the same mistake as Iran—not actively expanding capacity in recent years. This compute shortage has become severe enough that large tech companies are giving up massive revenue and profits. When that pain is real enough, people have economic incentives to do things they wouldn't otherwise do—like helping Intel and Samsung's logic businesses catch up. I've been writing about Intel since 2013: you must do foundry. At the time, it already felt late, but their stock soared under the cloud computing wave until 2020, when they truly realized the gap. Intel's fundamental problem: it lacks the customer-service mindset, culture, organization, and full IP portfolio, and can't find first customers willing to bear the friction of working with them—going to TSMC is just easier. But it's precisely this extreme compute scarcity that's forcing big tech companies to accept the pain of helping Intel get back on its feet. This is the 'high prices cure themselves' logic: scarcity creates economic incentives for people to do what 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

Host: Among the top fifteen tech companies, which do you think are in the most interesting position today?

Ben Thompson: 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 most typical example. Contrary to popular belief, AWS wasn't Amazon's idle capacity, but came from a deep internal insight: our internal retail team needed a scalable, API-only 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 eventually both. Amazon's logistics is the other side: first using third parties (UPS, FedEx, USPS), then building their own, then opening to the outside. Marketplace, Graviton chips, Trainium chips—all follow the same pattern: 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 this path too; 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, hard to replace by AI, and only benefits from it. Their moat depth, I believe, exceeds any other tech company.

On Apple: Apple seems to be sitting on the sidelines in this AI wave, which might be a case of 'luck being better than skill.' 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 suppliers on demand. From a consumer demand perspective, people don't want to improve productivity; they just want a usable assistant—a truly good Siri. This can be done entirely on-device; Amazon doesn't even need to bear the inference cost because they use the user's own battery. Apple's real strength: the smartphone itself is near-perfect. Small enough to fit in a pocket; big enough to view any content; it runs your entire digital life. TV has become an accessory; 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'—you just talk to AI and it tells you what you need to know—is a real opportunity, and Apple is best positioned here, provided they don't over-rely on proprietary frontier model capabilities. There's also a Microsoft-style trap here: Microsoft didn't miss mobile internet; they got in early, but their mobile products were 'shrunk-down PCs' because they always assumed the PC would be central. Apple realized the phone isn't a Mac accessory; the phone is the phone—they started building this understanding with the iPod and iTunes. Could Apple repeat this mistake—assuming the phone is always central, then adapting AI to center on the phone? 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 it's entirely rational for Apple to double down on what it does best. One last point: AI is a probabilistic engineering endeavor; Apple is the master of deterministic products. When you launch an iPhone, you get one shot, and you must get it right the first time; any quality issue costs billions. Apple has never recalled an iPhone, which reflects a completely different decision-making culture and supply chain philosophy, distinct from everything needed to build great AI. I tend to let companies focus on what they're good at—Apple continuing to make great devices is entirely reasonable.

The Landscape of Frontier AI Players

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

Ben Thompson: OpenAI and Anthropic have the highest risk but also the highest upside. Don't underestimate the power of conviction—the most influential undertakings in history are usually driven by some form of 'religion.' Silicon Valley has two 'religious organizations': OpenAI is like the mainstream church, attending every Sunday; Anthropic is like fervent evangelicals, with faith internalized to the bone—it's their core reason for existing. Google just needs search not to die too quickly; Meta has a massive advertising business to fall back on, but if Meta were run by anyone other than Mark Zuckerberg, they wouldn't be at the frontier—this is one of the purest expressions of founder energy. Meta building a new team from scratch and doing AI from zero is, frankly, quite crazy, but kudos to Zuckerberg. And there's a deeper logic: for a pure 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 to provide the middleware layer between all models, help enterprises build on their platform, manage model changes and iterations, and provide backward compatibility. This is the IBM playbook of the 1990s. When Gersner took over IBM, his insight was: IBM did everything mediocrely. Breaking it up wasn't the answer because the individual parts were worth even less. IBM's biggest asset was its 'bigness'—the internet arrived, and all companies needed to get online but didn't know how, so IBM became the integrator: building a middleware layer between legacy mainframes and modern web services, and building a massive consulting army alongside. This gave IBM another thirty years. Microsoft is taking this exact path: we'll handle AI for you, help you transform without handing over your core data, and we're the stable, reliable platform. It's a rational strategy, but also a desperate defensive posture.

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 possibly precisely because of that desperation, they might succeed.

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

Nvidia and the Commoditization of Intelligence

Host: How do you connect the lack of understanding of commodity markets with Nvidia and the commoditization of intelligence?

Ben Thompson: Nvidia's current profit margins, I believe, are somewhat unnatural. Let me explain why. There's a lot of discussion about 'circular financing'—Nvidia providing 25% capital support in exchange for equity in new cloud companies, which commit to buying Nvidia compute through 2030. On the surface, Nvidia maintains its price and margins; but in substance, they're taking on risk. If those new cloud companies can't sell the compute they've built, or if the hyperscalers have plenty of their own capacity, 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 it doesn't appear in the reported margin line, it truly exists in the company's overall discounted cash flow analysis. This is what I mean by 'price discounts appearing in strange ways.' Overall, 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 them to external customers. Google has already sold 20% of its TPUs to Anthropic, and Andy Jassy essentially confirmed on the last earnings call that Trainium will eventually be sold externally. Their logic is clear: this isn't a differentiation play but commodity sales; it won't cannibalize their cloud business's appeal, and it increases the leverage on their R&D investment. The hyperscalers' core advantage over new cloud companies is a lower cost of capital. 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 been greatly weakened because models don't care what hardware they run on. Nvidia's real hope might be for electricity to become the bottleneck. If power becomes severely constrained, users will seek the highest token efficiency, and Nvidia is still the most efficient choice—that would be good for Nvidia. But disappointing for 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 restarts; these have progressed surprisingly smoothly. This gives Google more time to refine TPU efficiency and Amazon more time to improve Trainium. If we end up in a world where power isn't the bottleneck, maintaining those margins will be extremely difficult.

What Survives the AI Bubble?

Ben Thompson: What I've been thinking about throughout the AI boom is: what will survive the bubble? The internet bubble left behind fiber optics. Google's rise was largely because they bought up massive amounts of dark fiber at bargain-basement prices after the crash—today's internet backbone still runs on WorldCom's fiber. The railroad boom left BNSF, still operating today, with its money flowing to Google. Good bubbles always leave behind something truly enduring. What can the AI bubble leave behind? GPUs won't last long, data centers might, but most importantly—electricity. If this all crashes and what's left is a massive new supply of electricity, that would be an incredibly valuable legacy. We've always lived under the constraint of energy scarcity; energy is the foundation of everything. What a world of energy abundance looks like is hard to imagine because our thinking has been framed by energy scarcity. I think America has delivered an amazing response on this front. Power is already a bottleneck and will continue to be, but the time it truly becomes the binding constraint is arriving later than everyone expected—including Jensen Huang himself, who I suspect thought power shortages would become Nvidia's moat sooner. This is one of the most encouraging signals of this era: watching America tackle these challenges is genuinely uplifting.