AI Like in AI-rlines? Why the Hyperscalers Lose — and You Win
July 27, 2026
13 min read
AI Like in AI-rlines? Why the Hyperscalers Lose — and You Win
Chinese AI now sells intelligence for cents on the dollar, and the market is asking whether America's 700-billion-dollar capex bet will ever pay off. The fear is justified — but it is bearish for the wrong companies and bullish for almost everyone else. This is not a story about the end of AI. It is a story about who pays for it, and who gets paid. Our analysis for The Market by NZZ, plus the extended investor's view for arvy readers.
In 30 seconds — the core thesis
- The AI revolution is real — the question is who gets paid. Microsoft, Alphabet, Amazon and Meta will spend a combined ~$700bn in 2026 — two-thirds of Switzerland's GDP, in a single year. Chinese models like DeepSeek deliver "good enough" performance at 1% of the price. When intelligence becomes a commodity, the profit migrates from the producer to the applier.
- Free cash flow is the mother of all red flags. Bank of America expects aggregate hyperscaler FCF to collapse from $191bn (2025) to $19bn (2026) — and turn negative in 2027. Some of the most profitable business models in history are about to start burning cash.
- The airline lesson: the perfect demand story that ruined its shareholders. The airline market nearly doubles to $1.2tn by 2033 — and cumulatively earned next to nothing for its owners. The money went to the passenger and the engine maker. We own Safran and GE Aerospace, not an airline.
The original analysis — the excerpt
«A durable competitive advantage has proven elusive since the days of the Wright brothers.» — Warren Buffett, Berkshire Hathaway shareholder letter (2007)
It's 2007. Warren Buffett sits down to write his annual letter and reaches for his favourite punching bag: the airline industry. A far-sighted capitalist present at Kitty Hawk, he jokes, would have done investors a lasting favour by shooting Orville down. Not because the planes didn't fly. Because the shareholders never did.
Fast forward to last week. Mr. Market is in a foul mood. The Nasdaq corrects sharply, momentum darlings are taken behind the woodshed, and the headlines quickly find their culprit: China. More precisely, Chinese AI models that deliver nearly the same performance as their American rivals — for a fraction of the price.
The fear is real. And my take may surprise you: the fear is justified. But it is bearish for the wrong companies, and bullish for almost everyone else. Welcome to the world of AI-rlines.
→ Read the full article on The Market by NZZ
01The Greatest Arms Race in Business History
To see why the fear is justified, follow the money. All 700 billion dollars of it. Microsoft, Alphabet, Amazon and Meta have guided towards combined capital expenditures of some $700bn for 2026 alone. For perspective: that is roughly two-thirds of Switzerland's entire annual GDP — spent in a single year, mostly on chips, data centres and electricity.
The cost blocks are staggering: tens of thousands of latest-generation Nvidia GPUs, with single clusters running into the billions. AI researchers commanding salaries that would fund entire engineering departments elsewhere. And energy — above all, energy.
| Dimension of the arms race | Magnitude |
|---|---|
| Hyperscaler capex 2026 (MSFT, GOOGL, AMZN, META) | ~$700bn |
| For comparison: Switzerland's annual GDP | ~$1,050bn (≈ two-thirds of it) |
| Data-centre electricity growth to 2030 (US + China) | ~80% of global growth |
| Heading towards | ~1,000 terawatt-hours ≈ Japan's entire annual use |
Sources: BCA Research, Bloomberg Finance L.P. (capex); Nature (energy). Both cited in the NZZ column.
High fixed costs demand high prices. Whoever spends hundreds of billions must charge accordingly per token to ever earn a decent return. Which is precisely where the problem begins.
02Cents on the Dollar. Literally.
Across the Pacific, a price war is raging. Baidu, Alibaba, Tencent, ByteDance and DeepSeek are undercutting each other relentlessly — solid models are sometimes given away for free to win market share. Add open-source models as a construction kit, subsidised energy, cheaper engineering talent and a fast-maturing domestic chip ecosystem around Huawei, and you get a radically different cost base.
How different? A different business model
The most capable American frontier models charge 25 to 50 dollars per million output tokens. DeepSeek's V4 Flash charges 28 cents. That is not a discount. That is a price gap of roughly 90 to 180 times — up to two orders of magnitude.
This is where the "good enough" threshold enters. If a model delivers 90% of the performance for 1% of the price, the rational CFO chooses the cheap option for customer service, standard code and routine analysis. Intelligence becomes a commodity — like electricity, like bandwidth, like a seat on a flight from Zurich to London.
And when the foundation becomes a commodity, the competitive edge migrates from building the model to applying it. In application — products, workflows, ecosystems — America has always been strongest. That is why the world got Amazon, the iPhone and Microsoft Office. But note carefully: that strength belongs to the users of intelligence. Not necessarily to its producers.
Why a moat made of software only partly protects against exactly this commoditisation — and why we walked away from software when the uncertainty became permanent — is covered in our companion analysis Software's Good Story — and Why It's Not Good Enough.
03Free Cash Flow: The Mother of All Red Flags
At arvy, two numbers sit above everything else: free cash flow and return on invested capital (ROIC). Both are currently telling the same uncomfortable story about the hyperscalers. Cash conversion is sinking while the capital base explodes. Even with revenues growing, ROIC is almost arithmetically condemned to fall before it can ever rise again.
| Aggregate hyperscaler free cash flow | Forecast (BofA) |
|---|---|
| 2025 | $191bn |
| 2026 | $19bn (−90%) |
| 2027 | negative |
Source: BofA Global Investment Strategy, Bloomberg. Some of the most profitable business models ever devised will collectively start burning cash from 2027.
Let's go back to the year 2022. Meta is pouring billions into the metaverse. Free cash flow collapses — and the stock loses roughly three quarters of its value, peak to trough. The business is intact. The users are still there. But Mr. Market has stopped believing that the spending will ever return a dollar. Only when Mark Zuckerberg declares the "Year of Efficiency" and capital discipline returns does the stock stage one of the great comebacks in market history.
The lesson and the prisoner's dilemma
The market punishes capex without visible returns brutally — and rewards discipline practically overnight. "Then they should simply spend less," you might object. But whoever taps the brakes first risks losing touch with the frontier — and would get stamped "loser of the race of the century," complete with a valuation discount. Whoever keeps spending loses margin. A prisoner's dilemma, played out with the largest balance sheets on earth.
This, quite honestly, is the part that keeps me up at night. Not for the economy. For the valuations.
Why free cash flow and ROIC sit above everything at arvy — and how we value on FCF yield rather than growth narratives — is covered in our foundational piece The True Cost of Investing. And how to apply the same price discipline to big-tech valuations is the subject of What Were You Thinking?.
04Monopoly Erosion: Everyone Hunting the Same Dollar
There is a second, underappreciated dimension. For decades, the tech giants were de-facto monopolists — but each in its own territory.
| Company | Then: its own monopoly | Now: the same product |
|---|---|---|
| Search | All building "intelligence," selling to the same customer, monetising via ads + subscriptions | |
| Meta | Social media | |
| Microsoft | Enterprise software | |
| Amazon | E-commerce |
Four quasi-monopolies. Four separate hunting grounds. Four beautiful, wide moats. And today? Four monopolies are merging into one oligopoly locked in a price war. Buying the hyperscalers no longer means owning four independent money machines. It means betting four times on the same AI-and-advertising dollar.
Monopoly erosion, I call it. Combine it with record capex and Chinese price pressure, and you have a demanding cocktail for stocks trading at premium multiples. History has seen this constellation before. It even had wings.
05AI-rlines: The Greatest Growth Story Nobody Made Money On
Commercial aviation was one of the great growth markets of the twentieth century — and it still grows today. Industry revenues of roughly $650bn are projected to nearly double to $1.2tn by 2033. Passenger numbers multiplied over decades, the world grew closer, tourism and trade exploded. The demand story was, and remains, close to perfect.
And yet, over its entire history, the industry has cumulatively earned next to nothing for its owners. Which brings us back to Buffett and Kitty Hawk. Aviation is no outlier.
As Edward Chancellor — the historian who has studied speculative manias like no one else — argued just days ago in a remarkable interview with The Market, the great speculative bubbles — from the railways of the 19th century to the automobile and electrification boom of the 1920s to the technology bubble of the late 1990s — were all investment booms followed by a bust. His verdict on the product itself: large language models are turning into a commodity — and commodity prices gravitate towards marginal cost, which contains no adequate return on invested capital.
Schumpeter said it a century ago: in a hypercompetitive environment, competition grinds profits towards zero. The railways carried the freight. The fibre optics carried the internet. The shareholders carried the losses.
Why did the perfect growth story ruin its investors? The ingredient list reads suspiciously familiar:
| Ingredient | Airlines then | Hyperscalers now |
|---|---|---|
| Enormous capex | Aircraft fleets | GPUs & data centres |
| Commodity product | A seat is a seat | A token is a token |
| Merciless price competition | Low-cost carriers | Chinese labs & open source |
| Permanent co-pilots | Regulation & geopolitics | Regulation & geopolitics |
Who profited? Everyone else. Passengers, whose real ticket prices collapsed. The economy, which could build on global supply chains. And, often forgotten, the suppliers: engine makers earn on every flight hour, no matter which airline goes bankrupt next.
Why we own engine makers, not an airline
It is no accident that we prefer owning GE Aerospace and Safran to owning any airline. Safran needs planes to keep flying — a low bar. An airline needs fuel prices, labour unions, weather, regulators and competitors to cooperate simultaneously — a very high bar. Safran and GE Aerospace are Heavy Assets, Low Obsolescence. HALO, as regular readers know.
Transfer the analogy: the model providers are becoming airlines. They fly intelligence instead of passengers, deliver spectacular revenue growth — and fight a structural battle for their margins. They will not disappear, just as airlines never disappeared. But between "indispensable to the economy" and "attractive to shareholders" lies, as ever, an ocean. Preferably crossed at 36,000 feet.
Why physical assets — engines, grids, infrastructure — are reasserting their value in an AI-obsessed world is the subject of our companion analysis The Return of the Old Economy.
06The S&P 500: An AI Index in Disguise
Why should this concern you even if you own no hyperscaler directly? Because the index long ago stopped being one. Semiconductors alone account for roughly 20% of the S&P 500. Add the other fifty-odd names with direct AI exposure — hyperscalers, software, energy and infrastructure plays — and well over half of the index is riding a single thesis.
The S&P 500 has become an AI index in disguise. This is not a critique of passive investing — ETFs remain a wonderful instrument. But you should know what you own: whoever buys "the market" today buys a concentrated bet on the return on AI capex. With one foreseeable side effect: index volatility will rise structurally. Every twitch in the AI story — a price shock from China, a disappointing capex guidance — feeds straight through to the headline barometer.
The foretaste
The recent momentum washout was a foretaste. Goldman Sachs' high-beta momentum basket surrendered a third of its value in under a month. Momentum, as always, is a double-edged sword: what can rise 1,000% can also fall 50%.
And yet, beneath the surface, something remarkable is happening.
07The Breadth Awakens — and That's the Good News
Here is the good news — and it is considerably bigger than the bad. When intelligence becomes cheap and ubiquitous, nearly every other company wins. The insurer automating claims. The industrial group making maintenance predictive. The healthcare company accelerating development. They all buy intelligence without carrying the capex — like business travellers who profited from the airlines' price wars without ever owning a plane.
The market has started to trade exactly that:
| Market-breadth signal | What it shows |
|---|---|
| Advance-decline line (NYSE) | New all-time high — more stocks rising while the old locomotives consolidate |
| Russell 2000 vs. Magnificent 7 | Breaks its multi-year downtrend — regime change in the making |
| A/D trend since 2023 | Intact uptrend, accelerating |
Sources: TrendLabs, StockCharts.com (A/D line); Adaptiv Investment Management System (Russell 2000). The advance-decline line is the "blood pressure monitor" of market breadth.
For active investors, this is the real punchline. When more than half of the index hangs on a single thesis, Active Share stops being a buzzword and becomes a tool. A stock picker's market is beginning — one that favours companies that consume cheap intelligence over companies that must produce expensive intelligence.
We prefer businesses where the bar is low: quality with real free cash flow, high ROIC, and moats no API key can replicate. Geography, regulation, physics. Good Story, Good Chart — you know the recipe.
What "Good Story & Good Chart" means in practice, and how we hold quality across full market cycles, is covered in The Tortoise Problem of Quality Investing. Which positions we currently hold is documented in the arvy Strategies Quarterly Report Q1 2026.
08The Passenger Wins
Let me give the bulls their strongest argument, because it is a good one, and a piece that ignores it is propaganda, not analysis. At the frontier — scientific research, complex agents, strategic planning — a premium will continue to be paid, and there America leads. Switching costs and data sovereignty are real barriers; critical Western infrastructure will hardly ever run on Chinese models. The market stays bifurcated, and the hyperscalers own ecosystems that took decades to build.
All true. And I agree with most of it. But notice what it defends: the businesses, not the economics. Airlines had loyalty programmes, hub monopolies and bilateral route protections too. It protected the companies. It never protected the returns.
The big picture seems clear to me. Cheap, ubiquitous intelligence is one of the most bullish developments for the world economy since the internet itself — just not necessarily for those producing it at record capex. Aviation changed the world and impoverished its shareholders. AI will change the world too. Who gets rich along the way will be decided not in the data centre, but in the application.
"Your margin is my opportunity," Jeff Bezos famously said. The phrase Amazon once used to raid entire industries is now pointed at the hyperscalers themselves. This time, China is doing the talking. And you, dear passenger, are applauding.
Buffett wished someone had shot down the Wright brothers. Nobody will shoot down AI — nor should they. The planes will fly. The tokens will flow. The economy will soar. Just remember, when you board, who historically made money in this business. It was never the airline. It was the passenger. And the engine maker.
09Frequently Asked Questions
Does this mean I should sell Nvidia, Microsoft and the other hyperscalers now?
No — this is not a sell recommendation or a prediction. The core message is more nuanced: the hyperscalers' business models are excellent, but with record capex, collapsing free cash flow and premium valuations, the risk-reward is demanding. The question is not whether AI is real (it is), but who captures the return on invested capital. Historically, that was rarely the capital-intensive producer.
Why is cheap Chinese AI bullish rather than bearish?
Because the benefit migrates to where intelligence is applied — not where it is produced. When an insurer, an industrial group or a healthcare company can buy intelligence at 1% of the former price without carrying the capex, its productivity and margins rise. That is bullish for the broad economy and for the appliers — and bearish only for the capital-intensive producers trapped in a price war.
What is the AI-rlines analogy in one sentence?
Commercial aviation was a perfect demand story (market nearly doubling to $1.2tn by 2033), yet cumulatively earned next to nothing for its shareholders — because a commodity product, enormous capex and a price war ground returns towards zero. The AI model providers risk playing the same role: indispensable to the economy, but structurally difficult for shareholders. The money went to the passenger and the engine maker.
I invest passively in an S&P 500 ETF. Does this affect me?
Yes. Semiconductors alone account for roughly 20% of the S&P 500; together with the AI-linked names, well over half of the index rides a single thesis. This is not a critique of passive investing — ETFs remain a wonderful instrument. But you should know that "buying the market" today is a concentrated bet on the return on AI capex, with structurally higher volatility. One possible answer is complementary diversification into companies that benefit from cheap intelligence rather than producing it expensively.
How does arvy invest specifically in this environment?
We prefer "HALO" businesses — Heavy Assets, Low Obsolescence — with real free cash flow, high ROIC and moats no API key can replicate: geography, regulation, physics. Concretely, we own engine makers like Safran and GE Aerospace rather than airlines, and we favour companies that consume cheap intelligence over those that must produce expensive intelligence. The full positioning is in our Quarterly Report Q1 2026.
Read on — the thematic anchors of this analysis
Software's Good Story — and Why It's Not Good Enough
Why a moat made of code doesn't survive commoditisation
What Were You Thinking?
The arithmetic behind 10x sales — the price is the problem
The True Cost of Investing
Why FCF yield and ROIC sit above everything
The Tortoise Problem
Why slow quality compounders win
No Pain, No Premium
The mental model behind patience: destination vs. weather
Why arvy Is Your Long-Term Partner
Quality strategy, skin in the game, FINMA-regulated
Who Really Makes Money in the AI Boom
The AI revolution is real — as real as the internet, as real as aviation. Which is exactly why the decisive question is not "whether" but "who." Who carries the record capex, and who collects the productivity dividend? History gives an uncomfortable but clear answer: it was never the airline. It was the passenger. And the engine maker.
At arvy, that means concretely: we own the "Heavy Assets, Low Obsolescence" businesses with real free cash flow and moats of geography, regulation and physics — and we favour companies that consume cheap intelligence over those that must produce expensive intelligence. It's the kind of investing we always wanted for ourselves. And now we share it with you.
Original written by Thierry Borgeat, Co-Founder of arvy, for The Market by NZZ. The extended arvy version was reviewed by Patrick Rissi, CFA and Florian Jauch, CFA. Data sources: BCA Research, Bloomberg Finance L.P. (capex/FCF); Nature (energy); intelligentliving.co (token prices); BofA Global Investment Strategy, Bloomberg (free-cash-flow forecast); Goldman Sachs FICC & Equities, Bloomberg (momentum basket); TrendLabs, StockCharts.com and Adaptiv Investment Management System (market breadth); market.us (airline market). Warren Buffett quote from the 2007 Berkshire Hathaway shareholder letter. Edward Chancellor interview in The Market, July 2026. Last updated: July 2026.
Disclaimer: This article is for general educational purposes only and does not constitute personal investment advice. The securities named are illustrative and not a buy or sell recommendation. Past performance is no guarantee of future results. arvy is a FINMA-supervised asset manager with a CISA license (Art. 24). Imprint & Legal Notice.
