Big Tech is pouring unprecedented sums into chips, data centers, and power. The technology may change everything, but the business case still has to survive the cost.
The shape of the bet: Four companies will spend roughly $725 billion this year building AI infrastructure, and none has yet shown returns large enough to clearly justify it. A widely cited MIT study found 95 percent of corporate AI pilots show no measurable financial return. The European Central Bank warned in August that a market correction is likely even if AI succeeds as a technology. None of that means AI fails. It means the technology bet and the price bet are two different questions, and this piece is about the second one.
The scale of the bet
Four companies are about to spend more building AI infrastructure this year than most countries generate in annual economic output, and none has yet demonstrated returns large enough to clearly justify the scale of the investment. Microsoft, Amazon, Alphabet, and Meta are on pace to spend about $725 billion combined on capital expenditure in 2026, up 77 percent from about $410 billion the year before. Add Oracle to that group, and analysts at Futurum put the five largest US cloud and AI infrastructure providers at a combined $660 to $690 billion, a lower figure than the four-company total because it isolates AI and cloud infrastructure spending specifically rather than each company’s total capital budget. Either way, projections for 2027 already cross a trillion dollars, and central bankers have started warning openly that the stock prices built on top of that spending may be getting ahead of themselves.
That money is moving into GPUs, custom silicon, server farms the size of small towns, and power purchase agreements that stretch decades into the future. It is happening not because AI has already produced trillions of dollars in profit, but because the companies building it believe the cost of falling behind is worse than the cost of overbuilding. That belief is rational for each company individually. Whether it is rational for the industry as a whole is the question this piece is built around.
Why nobody can afford to stop
Four companies each fear the same outcome: a rival pulls ahead in AI and captures the next major computing platform before anyone else can respond. None of them knows whether the underlying economics will work out. What their spending suggests is that stopping first looks like the riskier move to the executives making these calls, because a rival’s lead in compute is far easier to see coming than a return on capital is to prove.
This is the logic driving the current buildout. If Microsoft slows its data center construction and Google does not, Microsoft risks ceding enterprise cloud contracts it may never win back. If Meta pulls back and OpenAI’s next model needs more compute than Meta has available, Meta’s own AI ambitions stall while competitors scale. Every major player is making a bet that looks reckless in isolation and defensible only in the context of what everyone else is doing.
That dynamic explains something that would otherwise be strange: companies keep raising their spending guidance even as questions about returns get louder. Alphabet raised its 2026 capital expenditure ceiling to as much as $190 billion during the year. Meta raised its own guidance twice, ultimately citing rising memory chip prices and additional data center costs. Amazon is projecting negative free cash flow this year for the first time in years, a direct consequence of capital spending that outpaces what the business is generating in cash. On recent earnings calls, executives at all three companies have framed the bigger risk as under-building while a competitor pulls ahead, rather than overbuilding.
The result is a self-reinforcing cycle. As long as every competitor keeps spending, no individual company can afford to be the one that stops. That collective momentum is what turns a bet into an arms race, and an arms race is exactly the kind of dynamic that produces overbuilding even when every participant is acting on reasonable information.
Where does the money come back from
Strip away the technology and the AI boom is a capital allocation problem. Money moves from investors into Big Tech balance sheets, from Big Tech into chips and data centers, from data centers into AI models, and from AI models into whatever businesses and consumers are willing to pay for. At some point, for the investment to make sense, money has to travel back up that chain in the form of revenue and profit large enough to justify what went into building it.
Right now, the money flowing down that chain is measured in the hundreds of billions per year. The money flowing back up is far smaller, and the gap between the two is one of the central facts of the AI economy in 2026.
Part of the reason is that businesses have not yet decided how much AI is actually worth to them. A July 2025 study from MIT’s Project NANDA, based on interviews with more than 50 executives, a survey of over 150 leaders, and an analysis of close to 300 public AI deployments, found that 95 percent of generative AI pilots inside companies delivered no measurable impact on profit and loss within six months of launch. Only 5 percent of pilots produced results the businesses running them considered a financial success. The study has been criticized for its narrow definition of success and its reliance on a relatively small set of interviews, and it likely understates value that shows up in ways other than a direct P&L line, such as productivity gains that are hard to isolate. But even a more generous reading of the underlying pattern points to the same conclusion: enterprises are experimenting widely with AI and struggling to convert that experimentation into the kind of durable, budget-justifying returns that would support hundreds of billions in annual infrastructure spending.
The study also found that tools built by external vendors succeeded about twice as often as tools built in-house, and that most AI budgets are still going toward sales and marketing pilots even though the researchers found the strongest returns in back-office automation. That mismatch matters because it suggests the industry has not yet worked out where AI actually creates value, let alone how to price and sell that value at the scale the infrastructure spending assumes.
A cost problem traditional software never had
Software has historically been one of the most profitable businesses in the world because of a simple mechanic: once you build it, serving one more customer costs almost nothing. A company can spend years and hundreds of millions building a product, and then distribute it to a million new users for a fraction of a cent each. That mechanic is what allowed software companies to sustain gross margins north of 80 percent for decades.
Generative AI breaks that mechanic. Every time a user sends a query to a large language model, the company serving that query pays for real computing power, drawn from real electricity, running on real hardware that depreciates. There is no equivalent of shipping a copy of software for free. The cost shows up again every single time.
Inference costs, the price of running a trained model to answer a query, have fallen dramatically, which complicates the picture rather than resolving it. Jefferies research, reported by the South China Morning Post, put average enterprise inference prices at about $2.04 per million tokens in May 2026, falling to $1.16 to $1.18 by early August, driven by intensifying competition and the rise of low-cost open-source models out of China. Broader industry tracking shows mid-tier model pricing falling around 36 percent year over year even as frontier-model pricing has risen, because each new generation of top-tier models adds capability that providers are pricing at a premium.
That split is the crux of the problem. Commodity-tier models, the open-source options and budget tiers that businesses use for routine tasks like summarization or basic drafting, are under intense and growing margin pressure as competitors undercut each other on price. Frontier-tier models, the small number of systems capable of complex reasoning or long agentic workflows, keep getting more expensive to train and serve even as their prices rise to match. Companies betting on AI infrastructure are being squeezed from both directions: falling prices at the commodity tier and rising costs at the frontier tier, with no guarantee that enough customers will pay frontier prices to cover what frontier capability actually costs to build.
Will the hardware hold its value
Data centers are built to last for decades. The chips inside them are not, and how long they stay useful is now a public fight over whether the industry’s earnings are real.
Investor Michael Burry, best known for his early bet against the 2008 housing market, spent late 2025 accusing hyperscalers of overstating earnings by extending the assumed useful life of their AI hardware. His argument was specific: Nvidia’s chip generations turn over about every two to three years, but companies including Meta, Oracle, Microsoft, Google, and Amazon have been depreciating that same hardware over five to six years on their books, a choice that lowers annual depreciation expense and inflates reported profit. Burry estimated the practice could understate industry-wide depreciation by roughly $176 billion between 2026 and 2028.
The companies involved argue the longer schedules are defensible. Nvidia has said customers consistently report four to six years of useful economic life based on real-world utilization, since older chips often get repurposed for inference workloads even after they stop being competitive for training frontier models. SEC filings show a genuinely mixed picture: Alphabet extended server useful life from four to six years in 2023, Microsoft made a similar extension in 2022, and Oracle extended its own schedule that same year. Meta went further still, extending its schedule again in January 2025, to five and a half years. Amazon moved the other way in 2025, shortening its assumption from six years to five and citing the accelerating pace of AI hardware development, a choice closer to what Burry’s critics would call the honest answer.
Whichever side of the accounting debate turns out to be right, the underlying economic tension is real, and it connects to a second, larger uncertainty: how much of this hardware the world will actually need. One version of the future has AI becoming more efficient faster than demand grows, leaving both the depreciation assumptions and the broader buildout looking oversized for the workloads they need to support. A second version has the opposite effect: cheaper AI gets used far more often, in far more places, and total compute demand grows even as the cost per query keeps falling. Economists call this the Jevons paradox, first observed in the 1860s when more efficient steam engines led to more coal consumption rather than less, because efficiency made coal-powered machinery cheap enough to use in far more applications than before.
Global data center electricity demand offers some evidence for the second scenario. The International Energy Agency’s most recent estimate has global data center electricity consumption growing 17 percent in 2025, with AI-focused facilities alone growing 50 percent, even as the cost of running any individual query has been falling for three straight years. That is Jevons in action: cheaper compute has not slowed the growth in total compute demand, it has accelerated it.
Whether that pattern holds is the single largest variable in the entire AI investment thesis. If efficiency gains keep outrunning demand growth, today’s buildout risks becoming overcapacity, and the aggressive depreciation schedules built on assumptions of years of useful demand for older chips risk being exposed at the same time. If demand keeps outrunning efficiency gains, both the spending and the depreciation assumptions built on it may turn out to be conservative rather than reckless. No one building data centers or setting those depreciation schedules today can yet know which future they are building for.
The physical world
AI lives mostly in the abstract for most people who use it. Its costs show up in transformers, land, and water.
Gartner projects worldwide data center power demand will rise 27 percent in 2026 alone, reaching 132 gigawatts, up from 104 gigawatts in 2025, with AI-optimized servers expected to surpass conventional servers as the largest single driver of consumption by 2027. Goldman Sachs Research has flagged a structural power shortfall in the United States of about 9.3 gigawatts in 2026, a gap the firm projects widening to 45 gigawatts by 2028. That shortfall is large enough that most major hyperscalers have now signed at least one nuclear power agreement or small modular reactor deal to secure supply outside the ordinary grid.
The scale gets more concrete at the local level. Electric Power Research Institute data shows data centers already consumed 26 percent of Virginia’s total electricity supply in 2023, with that share projected to reach as high as 59 percent by 2030. In Dublin, data centers already account for nearly 80 percent of the city’s electricity use, and the strain is not confined to the United States and Ireland. Saudi Arabia’s Public Investment Fund and the UAE’s G42 are each backing gigawatt-scale AI campuses, including a 5-gigawatt UAE project positioned as the largest AI data center outside the United States, precisely because sovereign capital can secure land, water, and power commitments faster than most American or European grids can currently deliver them. Land, transformers, cooling water, and grid interconnection capacity have become, alongside chip supply and capital, one of the industry’s hardest constraints on how fast AI infrastructure can be built. In several of the largest build markets, the grid itself, not money, is now what is limiting the pace of construction.
The dot-com lesson, and a better parallel
The comparison to the dot-com crash of 2000 gets made constantly, and it is usually made lazily. The internet was real, the transformation it caused was real, and the bubble that formed around internet stocks was real too. Investors did not misjudge whether the internet mattered. They misjudged whether every company racing to build internet infrastructure would be worth what the market was paying for it.
A closer parallel than the dot-com startups themselves is the telecom industry that built the physical network underneath the internet boom. Companies including WorldCom, Global Crossing, and Qwest laid over 80 million miles of fiber optic cable across the United States during the late 1990s, believing internet traffic would keep doubling at a pace that justified the buildout. Global Crossing alone spent close to $15 billion constructing more than 100,000 miles of cable before filing for bankruptcy in January 2002 with $12.4 billion in debt. WorldCom’s bankruptcy that same year, at $107 billion in assets, remained the largest in American history until Lehman Brothers in 2008. Across the sector, an estimated $2 trillion in market value was wiped out and more than 500,000 telecom jobs disappeared between 2001 and 2003.
The fiber itself, though, mostly survived. Much of the cable laid in that period was engineered to last 25 years or more, and most of it eventually got lit and put to use, providing the cheap bandwidth backbone that powered broadband, streaming, and cloud computing for the next two decades. The technology the telecom industry built turned out to be exactly what the internet needed. The companies and investors who built it too early, with too much leverage, mostly did not survive to benefit from it.
That is the version of history worth worrying about with AI. GPUs are a different kind of asset than fiber optic cable, and the difference cuts against today’s buildout:
| Boom | Physical asset | Typical asset lifespan | What happened to early investors | Long-term technological outcome |
|---|---|---|---|---|
| 1840s British railways | Iron rails and locomotives | Decades | Wave of bankruptcies among rail financiers | Permanently reshaped national trade and transit |
| 1990s telecom buildout | Dark fiber optic cable | 25 years or more | Sector lost roughly 2 trillion dollars in equity value | Became the backbone of the modern internet |
| 2020s generative AI | GPUs, clusters, custom silicon | 2 to 6 years, contested | Unsettled: capital spending has outpaced revenue so far | Still forming |
Fiber laid in 1999 could sit dark for years and still be worth lighting a decade later, because the glass itself did not go obsolete. A GPU purchased in 2026 is a depreciating asset on a much shorter clock, competing against next year’s chip even if it is never used. That makes today’s buildout more vulnerable to obsolescence than the telecom boom was. But the core lesson still transfers: a technology can be exactly as transformative as its believers claim, and the specific companies and investors financing its early infrastructure can still lose everything, because the technology being right and the financing being sound are two separate questions.
The market is already asking the question
That separation between technology and valuation is what a European Central Bank blog post raised on August 17, 2026, when five ECB researchers wrote that a correction in current stock market valuations is likely, even if AI proves to be exactly as transformative as its most optimistic backers believe. Their reasoning rests on the Shiller CAPE ratio, a measure that compares current stock prices to a ten-year average of inflation-adjusted earnings. As of August 2026, the S&P 500’s CAPE stood at about 41, its highest level since the dot-com peak of December 1999, when it reached 44, and above the 33 level it reached just before the 1929 crash. The researchers compared the AI rally explicitly to previous technology booms, including the 19th century railway expansion, noting that each case followed a similar arc: genuine transformation, rising valuations, and a correction that eventually separated the technology’s long-term value from the price investors had paid for it in the moment.
The exposure they flagged is concrete. Euro-area households hold approximately 440 billion euros in equities tied to the so-called Magnificent Seven, the group of large US technology companies that includes Nvidia, Microsoft, Alphabet, Amazon, Meta, Apple, and Tesla, mostly through funds and ETFs rather than direct ownership. A sharp correction in those stocks would reach European pension funds and ordinary savers who may not realize how much of their exposure runs through a handful of American technology companies. The Bank of England and the IMF issued similar warnings in October 2025, with the IMF’s managing director noting that valuations are approaching levels last seen during the internet bubble a quarter century ago.
None of these institutions are predicting when a correction happens. The ECB researchers were explicit that their analysis is independent research rather than official policy, and that if AI proves transformative enough, valuations could ultimately end up higher after a correction than they are today. What they are saying is narrower: the current level of optimism embedded in AI stock prices is historically unusual, and history suggests optimism at this level tends to get corrected eventually, regardless of whether the underlying technology succeeds.
The strongest case against all of this
The argument above has a real counterargument, and it deserves to be stated at full strength rather than waved away.
Civilization-scale computing infrastructure cannot be built after demand arrives. It has to be built in advance, on the belief that demand will show up, because data centers, chip fabrication capacity, and power plants take years to construct. Companies that wait for proof of demand before investing risk losing years of lead time to competitors willing to build first. Seen this way, today’s spending is not speculative excess. It is the necessary cost of being ready for a technology that adoption curves suggest is still in its early stages.
There is also a case that today’s spending will look conservative in hindsight rather than reckless. If AI agents mature enough to handle substantial categories of white-collar work, or if physical robotics powered by the same underlying models opens markets that barely exist today, the revenue eventually generated could dwarf what current infrastructure was built to support. Enterprise adoption curves for genuinely new technology categories have historically taken longer than early investors expect, largely because organizations move slowly, not because the technology underperforms. Under that scenario, the error would lie with observers who assumed demand would stay near today’s levels rather than accounting for how much larger it typically becomes once a technology matures.
That scenario also exposes what the infrastructure bet actually depends on. The clearest path to revenue large enough to justify hundreds of billions in annual spending runs through labor costs: AI that reliably replaces expensive human work, at scale, across enough industries to matter. That is a different and higher bar than AI that assists workers or speeds up individual tasks, and it is the bet embedded in every optimistic revenue projection for this infrastructure, whether or not the companies making it say so directly.
Both of these arguments could be true. Infrastructure could be necessary and still be financed badly. Demand could eventually justify today’s spending and still arrive too late to save the specific companies and investors who bet on an earlier timeline.
Who gets hurt, and who comes out ahead
If a correction does arrive, the damage would not fall evenly, and the telecom parallel is instructive about where to look.
The most exposed are companies with a single AI bet and no other business to fall back on: AI-native startups burning cash on compute they lease rather than own, and chip suppliers whose revenue depends entirely on hyperscalers continuing to buy at the current pace. A slowdown in hyperscaler orders would hit that layer first and hardest, the way equipment suppliers like Nortel and Lucent collapsed alongside the telecom carriers they sold to in 2001, even though the carriers were the ones that had overbuilt.
The hyperscalers themselves sit in a different position. Microsoft, Alphabet, Amazon, and Meta generate enormous free cash flow from search advertising, cloud contracts, and e-commerce that has nothing to do with whether AI ever turns a profit on its own. That diversification is what let Level 3 Communications buy up bankrupt fiber networks for cents on the dollar after the telecom crash and later carry an estimated 70 percent of global internet traffic. One plausible outcome, if AI valuations correct the way telecom did, is the same pattern repeating: weaker, single-bet companies fail or get acquired, while cash-rich survivors buy the infrastructure and talent left behind at a fraction of what it cost to build. Under that scenario, ordinary AI users could end up with cheaper, more accessible tools as a direct result of other people’s losses, the same paradox that made bankrupt fiber networks the foundation of cheap broadband a decade later.
Investors in the 1840s railway boom lost fortunes building track that still carries freight today. The people financing AI infrastructure in 2026 are making a similar bet, on a much shorter clock, with far less certainty that the asset survives long enough to pay them back.
Two separate bets
The AI boom is really two different bets wearing the same headline, and collapsing them into one question, is AI a bubble, obscures more than it reveals.
The first bet is on the technology itself: whether large language models and the systems built around them will meaningfully change how businesses operate and how work gets done. That bet appears to be resolving in the technology’s favor. Adoption is real, capability keeps improving, and even skeptical enterprise research finds pockets of genuine value, concentrated for now in back-office automation rather than the sales and marketing use cases getting most of the budget.
The second bet is on price: whether the specific companies currently building AI infrastructure, at the specific valuations investors are currently paying, will generate returns that justify what has already been spent. That bet is far less settled. It depends on questions that remain genuinely open: whether enterprises will pay enough to close the gap between infrastructure cost and enterprise revenue, whether commoditization at the model layer erodes margins faster than efficiency gains lower costs, whether the hardware being built today retains its economic value on the timeline companies are currently assuming, and whether the physical grid can deliver power fast enough to support the buildout already committed to on paper.
History suggests these two bets can resolve in opposite directions at the same time, as the railway boom, the dot-com crash, and the telecom fiber glut all did in their turn. The pattern has repeated often enough to deserve serious weight: transformative technologies tend to attract more capital than their early economics can support, and the reckoning that follows tends to separate the technology’s real value from the price the first wave of investors agreed to pay for it.
If a correction in AI valuations arrives, it will not settle the question of whether artificial intelligence works. It will settle a narrower and, for the people holding the stock, more consequential question: whether the price paid to build it arrived years ahead of the revenue needed to justify it.
