TL;DR
- The Bank for International Settlements warned that the five biggest hyperscalers are on track to spend over $1 trillion on AI infrastructure across 2025–2026, echoing past investment manias.
- BIS drew explicit parallels to 19th-century railroads and the late-1990s dot-com bubble, framing the AI data center build-out as a potential systemic macro risk.
- Tech investors counter that AI infrastructure already has substantial demand and monetization, unlike speculative railroads and dot-coms.
- The warning targets Alphabet, Amazon, Meta, Microsoft, and likely NVIDIA, whose aggressive AI capex is setting the pace for the entire ecosystem.
A Global Financial Authority Calls Out the AI Build-Out
The Bank for International Settlements just threw cold water on the AI infrastructure gold rush. In its latest annual report, the BIS — often called the central bank for central banks — cautioned that the five largest hyperscale tech firms are on track to spend over $1 trillion on AI-related capital projects between 2025 and 2026. That’s not a typo. A trillion dollars.
The report doesn’t mince words. “The worldwide competition to establish artificial intelligence infrastructure is currently enhancing economic expansion; however, the surge may resonate with some of the most challenging investment collapses in history,” the BIS said. The institution explicitly invoked two historical manias: 19th-century railroads and the late-1990s dot-com bubble.
This isn’t a think-tank white paper or a skeptical analyst note. It’s a warning from one of the most influential financial institutions on the planet, and it frames AI data center spending as a potential systemic risk. That’s a big deal.
Why the BIS Thinks Hyperscalers Are Building a House of Cards
The core argument is straightforward. A handful of companies — likely Alphabet, Amazon, Meta, Microsoft, and possibly Apple or NVIDIA — are turbo-charging near-term growth by pouring unprecedented capital into AI infrastructure. GPU clusters. Massive data centers. Power substations. Cooling systems. The whole stack.
But what happens if revenue expectations don’t materialize? The BIS isn’t saying AI is useless. It’s saying the scale of the bet might be wildly out of proportion to the speed at which enterprises actually adopt and monetize AI services.
And here’s where the historical analogies bite. Railroads transformed the 19th century, but they also triggered one of the worst investment busts in modern history. Thousands of miles of track got built on speculation — some of it never turned a profit. Dot-coms revolutionized commerce, but the late-1990s fiber build-out left the industry drowning in overcapacity and debt. Capacity overshoot led to years of write-downs and consolidation.
I can’t help but think of it like this: imagine five neighbors all deciding to build Olympic-sized swimming pools in their backyards at the same time, each convinced they’ll run the neighborhood’s hottest pool party business. Maybe one or two succeed. The others? They’re stuck with expensive holes in the ground and monthly maintenance bills they can’t cover.
The BIS is essentially asking: are hyperscalers building pools everyone will use, or are they building expensive holes?
Tech Investors Push Back Hard on the Railroad Comparison
Not everyone buys the doom-and-gloom narrative. Tech investors and some industry analysts argue that comparing AI infrastructure to railroads and dot-coms is fundamentally misleading. Their counterargument? Modern cloud AI services already have substantial demand and monetization.
Railroads were speculative. Many dot-coms had zero revenue. But OpenAI‘s ChatGPT reportedly hit $2 billion in annual recurring revenue within two years. Microsoft’s Azure AI services are growing fast. Enterprises are signing contracts. Developers are building on top of foundation models.
The bulls say the BIS is fighting the last war. This isn’t a speculative land grab — it’s infrastructure for a platform shift that’s already happening. The capex isn’t ahead of demand; it’s racing to keep up with it.
But here’s the rub. Current AI valuations and capex plans embed highly optimistic assumptions on future model performance, regulation, and enterprise adoption. If the next generation of models doesn’t deliver step-function improvements, or if regulatory crackdowns slow deployment, or if enterprises decide they’d rather wait for cheaper, more efficient alternatives — well, then you’ve got a trillion-dollar bet that doesn’t pay off.
And that’s the scenario the BIS is worried about.
What This Means for the Hyperscaler Arms Race
Over the last 18 months, hyperscalers have dramatically increased AI capex. They’ve built new data centers, acquired GPUs at unprecedented scale, and raced to ship larger and more capable foundation models. Investors have largely treated this as a structural growth story — a one-way bet on the future of computing.
The BIS intervention introduces a different narrative. It frames the AI infrastructure boom as a classic boom-bust pattern, similar to fiber build-outs during the dot-com era. Overcapacity. Write-downs. Consolidation. Pain.
If that narrative takes hold, it could cool investor enthusiasm for aggressive AI expansion. Shareholders might start asking harder questions about return on invested capital. Boards might pump the brakes on the next wave of data center spending. Analysts might start modeling downside scenarios instead of assuming hockey-stick growth.
That shift would ripple through the entire ecosystem. GPU vendors like NVIDIA would face softer demand. Power-equipment suppliers would see order books thin out. Cloud customers might negotiate harder on pricing, knowing hyperscalers are sitting on excess capacity.
And it could give an opening to more capital-efficient challengers. Open-source model builders don’t need to spend billions on infrastructure. Smaller cloud providers can lease capacity instead of building it. If the hyperscaler playbook starts looking like a liability instead of an asset, the competitive landscape shifts fast.
Three Things to Watch as the AI Capex Story Unfolds
First, watch how hyperscalers talk about AI capex on their next earnings calls. If they start hedging — emphasizing flexibility, scalability, or optionality instead of raw spending numbers — that’s a signal they’re feeling pressure. Any pullback in guidance would confirm the BIS thesis is gaining traction.
Second, monitor enterprise AI adoption metrics. The bull case depends on businesses actually deploying AI at scale and paying for it. If adoption stalls or if enterprises start complaining about ROI, the revenue side of the equation breaks down. That’s when overcapacity becomes a real problem.
Third, keep an eye on regulatory moves. Governments are still figuring out how to handle AI — everything from data privacy to energy consumption to model safety. If regulators crack down hard, they could kneecap demand just as supply is peaking. That’s the worst-case scenario for hyperscalers: you’ve built the infrastructure, but you can’t fully monetize it because the rules changed.
FAQ
Which companies is the BIS warning about in its AI capex report?
The BIS report targets the five largest hyperscale tech firms, which likely include Alphabet, Amazon, Meta, Microsoft, and possibly Apple or NVIDIA. These companies are collectively on track to spend over $1 trillion on AI infrastructure between 2025 and 2026, according to the BIS analysis.
Why does the BIS compare AI infrastructure spending to railroads and dot-coms?
The BIS draws parallels to 19th-century railroads and the late-1990s dot-com bubble because both involved massive capital investment that outpaced actual demand, leading to overcapacity, write-downs, and painful corrections. The institution warns that if AI revenue expectations don’t materialize, hyperscalers could face similar outcomes despite the transformative potential of the technology.
Do tech investors agree with the BIS’s AI capex warning?
Many tech investors and industry analysts disagree with the BIS comparison, arguing that modern cloud AI services already have substantial demand and monetization, unlike speculative railroads and unproven dot-coms. They contend that AI infrastructure spending is responding to real customer demand, not speculative future bets, making the historical analogies misleading.
What could trigger an AI infrastructure bust according to the BIS?
An AI infrastructure bust could occur if revenue expectations fail to materialize due to slower-than-expected enterprise adoption, regulatory crackdowns that limit deployment, or if next-generation models don’t deliver anticipated performance improvements. The BIS warns that current capex plans embed highly optimistic assumptions that may not hold, potentially leaving hyperscalers with expensive overcapacity.
Source: Reuters
