Button TextButton Text
Article

When Exponential AI Demand Meets Exponential Supply

What the telecom boom can teach us about scarcity, overbuild, and the falling cost of intelligence

Jason Kowal
•
November 19, 2026

AI infrastructure is genuinely scarce today, and the demand for compute is growing quickly. But the current shortage is also triggering an extraordinary investment response just as the cost of useful intelligence is falling at an equally extraordinary rate. I spent more than a decade at TeleGeography during the Internet infrastructure boom, when rapidly growing demand for bandwidth and data centers produced massive investment—followed by overcapacity and collapsing prices. While Internet traffic did continue to grow remarkably through the early 2000s, many of the companies that built its infrastructure did not survive to see it. 

At the height of the telecom boom 25 years ago, I was studying the Internet’s rapid growth and the infrastructure being built to support it. At my research firm, TeleGeography, we tracked not only international networks, submarine cables, traffic and prices, but also the growing data center business. At the time, the basic story seemed clear enough: Internet traffic was growing quickly, businesses and consumers were moving online, and the world needed far more bandwidth and computing infrastructure to support what was coming.

The broad thesis turned out to be right. The Internet grew very quickly and became incredibly important to commerce and culture. The problem was what happened on the supply side.

Telecom companies laid enormous amounts of new fiber, including submarine cable systems connecting the world’s major markets. At the same time, advances in dense wavelength division multiplexing (DWDM) increased the number of wavelengths each fiber could carry, while faster transmission rates increased the capacity of each wavelength. During the second half of the 1990s, the capacity of a single fiber pair increased by roughly two orders of magnitude. The industry was therefore adding more physical infrastructure just as technology was making each new strand vastly more productive. Fiber that had looked like a scarce resource could, in just a few years, carry 100 times as much traffic.

Capital was abundant as well. Investors wanted exposure to the Internet boom, growth forecasts were rewarded, and there was plenty of incentive to believe the most aggressive scenarios. In some cases, optimism crossed into fraud. WorldCom executives manipulated the company’s accounts to meet Wall Street expectations, while executives at Enron Broadband Services made false claims about the capabilities of the technology the company was developing. These extreme cases weren’t representative of the industry as a whole, but they signaled an environment in which enthusiasm was abundant and skepticism was not always rewarded. (SEC, “Bernard J. Ebbers”; U.S. Department of Justice, “Former Enron Broadband Co-Chief Executive Officer Sentenced for Wire Fraud”)

Eventually, supply grew faster than the market’s ability to absorb it at the prices investors had expected. Network and data center capacity became cheaper, asset values collapsed, and many companies failed even as Internet usage kept rising.

That history keeps coming to mind as I watch the current race to build AI infrastructure. Today the scarce resources are GPUs, data centers and, increasingly, electricity. Demand is real and capacity is tight. But those things were also true during the early stages of the Internet infrastructure boom.

Today’s shortage is real

Any comparison with the telecom bust needs to begin by acknowledging that the AI infrastructure market does not look overbuilt today.

CBRE reports that supply in the major North American data center markets increased 33.7 percent year over year in the first half of 2026, yet vacancy still fell to a record-low 1.4 percent. More than 80 percent of capacity currently under construction was already preleased. In Northern Virginia, vacancy fell to 0.2 percent. (CBRE, “North America Data Center Trends H1 2026”)

TeleGeography’s recent work reaches much the same conclusion from another direction. Its analysis of AI infrastructure describes energy availability as the “ultimate gatekeeper” for expansion. Data centers designed for AI require far more power than conventional cloud facilities, and obtaining that power can involve years of permitting, grid upgrades and construction. (TeleGeography, “Infrastructure, Not Compute, is the Real AI Bottleneck”)

There is little reason to question the shortage. The more interesting question is what that shortage is causing companies to build.

The revenue hurdle is getting very large

A new analysis from Azeem Azhar and the Exponential View team puts some useful numbers around the investment cycle.

They estimate that hyperscalers and neocloud providers will spend about $880 billion in total capital expenditures in 2026. Importantly, they do not treat all of that as AI spending. These companies were already investing heavily in infrastructure before ChatGPT arrived. After adjusting for that baseline, Exponential View estimates that approximately $560 billion of this year’s spending is AI-related.

From there, they model three possible investment paths through 2030. Their low case assumes that growth peaks in 2028 and AI capital spending falls to about $600 billion a year by 2030. Their medium case reaches more than $800 billion. In their high case, annual AI capital spending approaches $1.9 trillion.

Servicing the investment on the high-growth scenario—including depreciation, operating costs, financing and a 15 percent return on invested capital—would require approximately $1.3 trillion in annual infrastructure revenue by 2030.

Their current estimate of annualized AI infrastructure revenue is about $175 billion. Reaching $1.3 trillion would require a 7.4-fold increase from current figures in just four years, or approximately 59 percent annual growth. (Exponential View, “How much revenue does the AI buildout need in 2030?”, AI Investment Brief, Sept. 24, 2026)

Subsequent earnings reports have reinforced the scale of that estimate. Amazon, Alphabet, Microsoft and Meta alone are now guiding toward roughly $720 billion to $740 billion of capital spending in 2026, although their accounting definitions differ and not all of the spending is AI-related. All four continue to identify AI infrastructure as a major driver of investment, and several are already signaling further increases in 2027.

Anthropic’s IPO prospectus makes both sides of this equation visible. The company disclosed approximately $901 billion of purchase commitments and leases, substantially all related to compute and infrastructure, including about $452 billion expected before the end of 2030. At the same time, Anthropic reported nearly $17 billion of third-quarter revenue, more than twelve times the year-earlier level, and a 15 percent adjusted operating margin. Much of its revenue, however, is consumption-based and not secured by long-term contracts, while its compute commitments extend years into the future. Anthropic itself warns that forecasting errors could leave it committed to excess capacity that remains underutilized.

Anthropic’s growth shows that revenue can rise quickly enough to support extraordinary levels of investment, at least for a time. Infrastructure revenues across the sector have recently been growing much faster than 59 percent, and hyperscalers may accept lower returns for years if they believe the spending strengthens their competitive position. But the underlying issue remains the same: the infrastructure being built has to generate enough revenue, at sufficient prices, to produce an acceptable return on trillions of dollars of investment.

That was the same problem telecom investors eventually faced.

A lot of tomorrow’s supply is not operating yet

Another feature of the current cycle is also familiar: the lag between the investment decision and the arrival of usable capacity.

Large data center projects take years to plan, finance, connect to power and build. Exponential View has previously documented a growing backlog of hyperscaler assets under construction or otherwise not yet in service. In the new analysis, the team again emphasizes that large amounts of announced and allocated spending remain tied up in unfinished facilities that cannot yet generate revenue. Meta, for example, has to wait 1.7 years for capex it spends today to come into service — a year more than FY2024 (Exponential View, “Making sense of the AI CapEx logjam”)

That creates a timing problem. Today’s scarcity is being measured before much of the capacity financed in response to that scarcity has entered service.

The supply response is also extending beyond data centers. TeleGeography says submarine cable construction is returning to levels not seen since 2000–2001, with more than $16 billion of new systems planned to enter service between 2026 and 2029. The dollars are small compared with data center investment, but the pattern is familiar: companies are adding physical network infrastructure while improvements in transmission technology continue to increase the capacity of that infrastructure.

This was one of the defining characteristics of the millennial telecom boom. High prices and constrained capacity sent a perfectly rational signal to individual companies to build more. The problem became apparent when many companies acted on the same signal and several years of investment arrived in the market at roughly the same time.

‍

The price of intelligence is falling even faster

The most interesting part of the AI infrastructure story may be the cost of what all that infrastructure produces.

Epoch AI recently examined the cost of achieving a fixed level of AI performance across benchmarks in mathematics, science and games. Since 2023, it estimates this cost has fallen approximately 47 percent per quarter, or about 13-fold per year. (Epoch AI, “The Plunging Price of Thought”)

One example makes the scale easier to appreciate. In January 2025, Epoch estimated that achieving a 75 percent score on a graduate-level science benchmark cost about 30 cents per question. Less than 18 months later, comparable performance cost approximately $0.0004 per question, a 725-fold decline. (Epoch AI, “The Plunging Price of Thought”)

Anthropic’s results show the same dynamic in the market: the company says its latest Opus model reduced customer cost per task by about 40 percent even as usage and revenue continued to grow rapidly.

Operating efficiency is improving quickly as well. Microsoft researchers estimate that optimized frontier-model inference currently consumes a median of about 0.31 watt-hours per query under realistic large-scale operating conditions. They estimate that improvements in model design, serving systems and hardware could reduce energy consumption per query another 8 to 20 times. (Microsoft Research, “Energy Use of AI Inference, Efficiency Pathways, and Test-Time Scaling”)

None of this means that total demand for compute will fall. Azeem Azhar and his team have made a strong argument for the opposite. Their earlier analysis found that every 10 percent reduction in token prices was associated with a 12 to 18 percent increase in token consumption. Cheaper intelligence encourages people to use more intelligence, a version of the Jevons paradox. Exponential View’s model of the AI economy also estimates that AI revenues are already growing faster than those of earlier Internet and mobile technology waves. (Exponential View, “The State of the AI Economy”)

I find that argument persuasive. But I do not think it resolves the infrastructure question.

Bandwidth went through a similar phenomenon. As communications costs fell, people invented new ways to consume bandwidth. Streaming video, cloud computing, social media and videoconferencing all became practical in part because networks became faster, cheaper and more abundant. Usage exploded.

That did not prevent network and data center capacity from becoming overbuilt. Rising consumption and deteriorating infrastructure economics can happen simultaneously.

The same could occur with AI. We may consume vastly more intelligence precisely because it becomes cheap, while some of the infrastructure built when intelligence was scarce becomes less valuable than its investors expected.

When scarcity sends compute underwater and into orbit

One indication of how strongly the market believes in today’s constraints is the range of alternatives being explored.

Microsoft’s Project Natick demonstrated that a full-scale data center module could operate on the floor of the North Sea. Its servers experienced one-eighth the failure rate of a comparable land-based control group. The experiment also showed some obvious attractions of subsea infrastructure, including access to natural cooling and offshore renewable power. (Microsoft Research, “Project Natick”)

Now some of the same thinking is moving into space. NVIDIA has launched computing technology intended for orbital use, while companies including Aetherflux, Axiom Space and Starcloud are pursuing different approaches to space-based computing. (NVIDIA, “NVIDIA Launches Space Computing, Rocketing AI Into Orbit”)

These ideas may become important or remain niche. Their importance to this argument is more modest. Serious companies are considering radically different places to locate compute because power, cooling, land and permitting appear likely to remain binding constraints.

That may turn out to be exactly right. History still gives us reason to ask what happens if some of those constraints prove less durable than they look today.

This is not the telecom boom all over again

There are important differences between the AI boom and the telecom boom.

The largest AI infrastructure investors today include Microsoft, Amazon, Alphabet and Meta. These companies have enormous existing businesses and much stronger balance sheets than the heavily leveraged telecom carriers that depended on continued access to capital markets during the late 1990s.

But the financing picture is becoming more complicated. In November, the Federal Reserve estimated that debt issuance related to advanced computing infrastructure had exceeded $680 billion through October, roughly five times the level of 2023. Public debt outstanding exceeded $1.1 trillion, while future leases, guarantees and other contractual commitments added trillions more in potential obligations. Most strikingly, the Fed compared the current hyperscaler investment cycle directly with the telecom boom, concluding that the pace of AI capital spending may soon surpass it. The Fed also noted an important difference: strong hyperscaler cash flows have so far allowed them to support much more investment with less debt than telecom companies required.

Power may also prove to be a more serious constraint than bandwidth. Optical technology could multiply the capacity of existing fiber without laying another cable. Electricity still has to be generated, transmitted and delivered.

McKinsey recently concluded that the greater near-term risk to the U.S. power system may be underbuilding rather than overbuilding. At the same time, its analysis shows why announced demand deserves scrutiny. Large-load interconnection requests are roughly nine times larger than McKinsey’s projected 2030 demand, in part because companies submit speculative and duplicative requests. Even contracted and high-confidence projects represent about twice its projected demand. (McKinsey, “Powering AI: How Real Is the Risk of Overbuilding?”)

The infrastructure itself is different too. AI chips have much shorter economic lives than fiber, although Exponential View notes that productive chip lives have recently lengthened to roughly five or six years. Training frontier models may continue to favor enormous centralized facilities, while inference may become much more distributed. Capacity can also be excessive in one geography and scarce in another.

A familiar Singularity idea

One of the frameworks popularized during Singularity’s early years was Peter Diamandis’s “6 Ds” of exponential technologies: digitization, deception, disruption, demonetization, dematerialization and democratization. Once technologies become digital, the framework argues, exponential improvement can eventually make capabilities dramatically cheaper and more widely available. (Singularity Hub, “Digitized to Democratized: These Are the 6 Ds of Exponential Technologies”)

AI appears to be moving through that process remarkably quickly. Epoch AI’s estimate that the cost of a fixed level of performance is falling roughly 13-fold a year is close to a textbook example of demonetization.

For users, that is good news. Cheaper intelligence should make increasingly powerful capabilities available to more people and organizations. For investors in infrastructure, however, demonetization presents a different challenge. Assets being financed today are being valued in a market where compute, power and data center capacity remain scarce. Their future returns depend in part on how long that scarcity lasts.

Where this leads

My expectation is that AI demand will continue to grow rapidly, perhaps faster than we expect today. But I also expect the cost of compute and useful intelligence to keep falling as new capacity comes online, chips improve and models become more efficient. Those forces will not affect every market equally. Power may remain scarce in some places for years, while particular types of compute or facilities become abundant elsewhere. But some of the infrastructure being financed on today’s scarcity assumptions may earn significantly lower returns than its investors expect.

The telecom boom offers a useful precedent. Internet demand exceeded most early forecasts, while bandwidth prices collapsed and large amounts of network and data center capacity became economically impaired. The technology succeeded beyond expectations; that did not guarantee attractive returns for the companies that financed its infrastructure.

I expect AI to produce a similar tension. It can become vastly more important, consume vastly more compute and require trillions of dollars of infrastructure—while still leaving some investors with assets built at the wrong price, in the wrong place or around assumptions that did not survive technological change.

If AI follows the pattern of other exponential technologies, the riskiest assumption is that scarcity will last.

‍

Get access

Jason Kowal

Jason Kowal is a consultant at Singularity and the former CEO of TeleGeography, where he spent more than a decade analyzing global telecom, Internet infrastructure and data center markets through the boom and crash of the late 1990s and early 2000s.