The Capex Cycle:
Telecom 2000 vs. AI 2026

Executive Summary

Two build-outs, set side by side

The fiber thesis of 2000 turned out to be right, and the cables were eventually used. Even so, nine of the ten largest companies in the world at that peak went on to trail the index for the next 25 years, and the Nasdaq did not pass its March 2000 high again until 2015. That gap between a correct idea and a profitable investment is the reason this comparison is worth drawing carefully.

On structure, the two cycles resemble each other. Both pair a major new technology with capital spending measured in the hundreds of billions a year. Both rest, at least in part, on demand forecasts that run ahead of current revenue. Both involve financing that loops between suppliers and customers. And in both, the assumptions about how fast the equipment wears out have a real effect on reported profit. Judged on structure alone, the present cycle can look like a repeat of the last one.

On substance, several differences stand out, and they point the other way. The telecom carriers funded their networks with debt raised against earnings that were, in the most serious cases, fabricated. Today's large platform companies fund most of their spending from established, cash generative businesses and hold net cash rather than net debt. AI usage is visible in reported revenue and in metrics like tokens processed, where the 2000 demand projections never showed up in the numbers. And the depreciation question today concerns disclosed estimates, not the fraud that defined WorldCom, Qwest, and Global Crossing.

The newer questions sit at the edges of the system rather than at its center. Electricity has become the practical limit on how fast data centers can be built, a constraint that barely existed in 2000. A growing share of the financing is moving into private credit and off balance sheet vehicles, where it is harder to see. And one structural fact applies to passive and active investors alike: the ten largest companies now make up a record share of the S&P 500. None of this is presented as a prediction. The aim is to lay out the evidence on both sides, identify the specific things worth monitoring, and let you weigh them.

At A Glance

Same, similar, or different, line by line

Nine points of comparison, each marked for how closely this cycle tracks the last. The shorthand is simple: same the pattern repeats, similar it broadly resembles 2000, different it has genuinely changed from the last cycle, and new it has no real precedent in 2000.

The pattern that follows: funding and demand are differences that lower systemic risk relative to 2000. Asset life is also a difference, and it runs the other way, since the hardware must be earned back far faster than fiber did. Circular financing is a genuine parallel to 2000, while capital intensity, off balance sheet leverage, and concentration are broadly similar. Power is genuinely new. Each is taken in turn below.

Part I  ·  The Precedent

The telecom build-out and the downturn that followed

To judge the present cycle, it helps to recall in detail what the last one actually looked like, since it is now often remembered only in summary.

The spend, and the capital behind it

Through the second half of the 1990s, US investment in communications equipment rose from roughly $62 billion a year in 1996 to more than $135 billion by 2000 in constant dollars, a pace of nearly 18% a year.¹ Total telecom capital spending peaked near $120 billion in 2000. The Telecommunications Act of 1996 had opened local markets to competition, and capital flowed in to build the networks that competition was expected to need. Much of that money was borrowed. The share of US high yield bond issuance coming from telecom climbed from about 15% in 1997 toward half of the market by 2000, and a generation of new long distance and competitive local carriers funded fiber routes with debt raised against business plans rather than customers.

How capital spending inflated reported earnings

The mechanism that turned heavy network investment into a solvency problem is worth stating plainly, because a version of it recurs today. When a company builds a network, the cash goes out at once but the cost is spread across the asset's expected life through depreciation. The longer the assumed life, the smaller each year's charge, and the larger reported earnings look in the meantime. Pair generous asset lives with heavy borrowing and a company can report growing operating earnings while consuming cash and adding debt. That worked while capital was available. It stopped working the moment funding tightened.

The demand forecast that did not arrive

The build-out was justified by a widely repeated claim that internet traffic was doubling every hundred days, an annual rate near 1,000%. The figure originated with a single carrier's marketing in the late 1990s and was repeated in industry and government reports until it was treated as fact. Independent measurement by the researcher Andrew Odlyzko found that traffic was in truth roughly doubling each year, fast, but closer to 70 to 150% than to 1,000%.⁴ Networks had been sized for the larger number. When demand grew at the smaller one, the result was a vast surplus of unused capacity.

The surplus, the failures, and the losses

By the early 2000s more than 80 million miles of fiber had been laid in the United States, and industry estimates suggest the large majority of it sat unlit for years, since precise figures were never knowable.⁵ Prices for long haul capacity fell sharply, by some measures around 90%. As funding markets closed, the carriers that had borrowed against projected traffic could not service their debt. The failures that followed were among the largest of the era, and several involved accounting fraud rather than only optimism.

Vendor financing, the demand the sellers paid for

One feature of the period has a direct parallel now. The equipment makers helped their customers buy the equipment. Lucent extended about $8.1 billion in financing commitments to buyers, Nortel about $3.1 billion, and Cisco about $2.4 billion, with industry commitments reaching perhaps $25 to $33 billion by late 2000.⁴ When a supplier lends a customer the money to buy its product, the resulting sale is real on the income statement but the demand behind it is partly the supplier's own capital returning as revenue. That made end demand look stronger than it was, and it is the cleanest precedent for the cross investment in today's market.

Part II  ·  The Aftermath

A right idea, and the wrong stocks

The most useful lesson from 2000 is not that the technology disappointed. It is that being correct about the technology did not protect investors who owned the leaders at the peak. This is the part of the history most worth holding in mind.

The index, a fifteen year round trip

The Nasdaq Composite closed at 5,048.62 on March 10, 2000. It bottomed at 1,114.11 on October 9, 2002, a decline of 78%, and did not close above its 2000 high again until April 23, 2015, roughly fifteen years later.⁶ Adjusted for inflation, an investor who bought the index at the peak was still behind well into the following decade. The broad market fared better than the Nasdaq but still produced what is often called a lost decade, with the S&P 500 returning close to zero on a price basis over the ten years from 2000.¹¹

The individual outcomes

The index figures understate what happened to the companies most associated with the build-out. Cisco, the most valuable company in the world for a time in 2000, fell about 89% from its high and did not reclaim that price until December 2025, more than a quarter century later. Intel did not exceed its 2000 high until 2026. Nortel, worth roughly $398 billion at its peak and about 38% of the value of the entire Toronto exchange, filed for bankruptcy in 2009. JDS Uniphase fell more than 99%. The equipment that these companies built proved its worth. The equity did not follow.

The largest companies, twenty five years on

Consider the ten largest US companies at the 2000 peak and how their stock performed against the S&P 500 over the following 25 years. By common measures only one, Microsoft, outpaced the index, and even Microsoft spent more than a decade below its 2000 level before its later software and cloud businesses carried it past.⁷

Constituents and outcomes compiled from multiple public sources; per name 25 year returns are estimates and the ranking varies slightly by date and method. The pattern, not any single figure, is the point.⁷

Part III  ·  The Present Cycle

The AI build-out, larger spending on stronger balance sheets

The present cycle is bigger than 2000 in dollar terms and is being funded very differently. Both facts matter, and they pull in opposite directions, which is why a single label does not fit.

The scale

Combined capital spending by the largest platform companies rose from about $162 billion in 2022 to roughly $448 billion in 2025, a figure that includes the cloud build-out at Amazon, Microsoft, Alphabet, Meta, and Oracle.⁸ Estimates for 2026 are higher still, generally in the six hundred billion dollar range, though these are forecasts and depend on definitions. Measured against the roughly $120 billion that the entire US telecom industry spent at its 2000 peak, several of these companies are now individually approaching or exceeding that figure. The absolute numbers are without precedent in corporate history.

Who is paying, the decisive difference

The most important contrast with 2000 is not the size of the spending but the balance sheets behind it. The telecom carriers borrowed to build, against earnings that in the worst cases were not real. The large platforms today fund the majority of their capital spending out of the cash flow of established, highly profitable businesses, and most hold more cash than debt. That single difference is the strongest reason a synchronized, debt driven downturn of the kind seen in 2000 is less likely now. It does not remove every risk, as later sections show, but it changes the base case materially.

Spending relative to revenue

A useful way to size the commitment is capital spending as a share of revenue. On 2026 estimates this ranges from about a quarter of revenue at Amazon to the mid forties for Microsoft and Alphabet, to more than half at Meta, to the mid eighties at Oracle.¹⁰ These are high figures by any historical standard for companies of this size, and they are the clearest sign that the firms regard the opportunity as large. They are also the reason free cash flow has come under pressure even at very profitable businesses.

The gap between spending and revenue

In 2024 the investor David Cahn of Sequoia framed the central financial question of the build-out as a gap between the capital being deployed and the revenue required to justify it, an exercise widely known as the six hundred billion dollar question.¹² The gap is real and worth watching. It is also narrowing, because AI revenue is growing quickly from a small base, in contrast to 2000, when the projected demand never showed up in the numbers at all.

The revenue is concentrated in a few providers and is growing fast, though the figures are self reported annualized run rates rather than audited results and should be read as such. OpenAI's run rate rose from roughly $200 million in early 2023 to about $13 billion in 2025 and toward an estimated $24 billion in early 2026. Anthropic reported a run rate of about $9 billion at the end of 2025 and roughly $30 billion by April 2026, a figure that is measured on a gross basis and is disputed by competitors.¹³

Depreciation, a shared question about earnings quality

The depreciation mechanism described in the telecom section applies here too, with one important difference. The useful life a company assigns to its servers and chips determines how quickly their cost runs through the income statement. Most large platforms depreciate this equipment over five to six years. Some investors, most prominently Michael Burry, have argued that the economic life of a leading chip is closer to two or three years, and estimated that the gap could understate depreciation across the group by around $176 billion between 2026 and 2028, overstating profit at some companies by twenty percent or more by 2028.¹⁴ Others counter that older chips remain useful for less demanding work and continue to earn, so the economic life is longer than the bears assume. The important points for a reader are that these are disclosed estimates rather than hidden facts, that reputable analysts disagree about the size of the effect, and that depreciation is a non cash charge, so it changes reported profit but not the cash already spent.

Chip economics, a much shorter asset life

This is where the comparison turns against the present cycle. Fiber, once laid, stayed useful for decades, which is why the surplus of 2000 was eventually absorbed at a profit by later users. The expensive hardware at the center of the AI build-out is different. Leading chips are refreshed on a roughly two year cadence, and the rental price of a prior generation chip falls quickly as a newer one arrives, by some estimates around 70% over the period since 2023. Because this hardware is the majority of the cost of a modern data center, while the building itself lasts far longer, the economics require continual reinvestment to stay current. That is a structurally different proposition from laying a cable that will still be valuable in twenty years.

Circular financing, the same structure at larger scale

The vendor financing of 2000 has a direct descendant. Capital is again moving in loops among the companies that supply the build-out and the companies that buy from it. A chip maker invests in a model developer, which buys the chip maker's chips. A cloud provider funds a model developer, which runs on that provider's cloud. None of this is improper, and reputable reporting is careful to separate ordinary circular investment from the fraudulent round trip transactions of the past. The reason to watch it is the same as in 2000: when investment from a supplier returns as that supplier's revenue, the strength of underlying demand becomes harder to read from the headline numbers.

• $100BNvidia's pledged investment in OpenAI, a ceiling tied to staged deployment rather than cash committed, described by Nvidia's chief executive as unlikely to be drawn in full¹⁶
• ~$13BMicrosoft's cumulative investment in OpenAI, which runs on Microsoft's Azure cloud
• $8BAmazon's investment in Anthropic, which runs on Amazon Web Services
• $6.3BNvidia's commitment to purchase unused CoreWeave capacity through 2032

The specialized providers, the closest parallel to telecom debt

If a 2000 style strain were to appear, the most likely place is not the large platforms but the specialized providers that borrow to buy chips and rent them out, sometimes called neoclouds. CoreWeave is the clearest example, and it deserves an even handed reading because the outcome is genuinely unknown. On the cautious side, the company carried roughly $11 billion of debt in mid 2025, alongside about $34 billion of scheduled lease payments through 2028, its interest costs are large and rising, and in the second quarter of 2025 a single customer, Microsoft, accounted for about 71% of revenue, a concentration that turns one contract into an existential question.¹⁶ On the constructive side, revenue grew about 134% from a year earlier to roughly $1.4 billion in the third quarter of 2025, the company reported a contracted backlog above $55 billion, and it has the backing of Nvidia and offtake commitments from large, creditworthy customers. The structure, borrowing against fast depreciating assets with concentrated demand, is the one that failed in 2000. Whether the contracted backlog and strong sponsors make this time different is a real and open question, and no one can answer it with confidence today.

Part IV  ·  The New Constraint

Power and the grid, a limit that did not exist in 2000

Fiber used very little electricity. AI data centers use a great deal. The clearest difference between the two cycles is that the scarce input is no longer capital or even chips, but power and the equipment that delivers it. This is the part of the comparison with no real precedent, and it works both for and against the case that demand is durable.

US data centers used about 4.4% of the country's electricity in 2023, roughly 176 terawatt hours, up from under 2% only five years earlier. The Department of Energy and Lawrence Berkeley National Laboratory project that share rising to between 6.7 and 12% by 2028.²⁰ Globally, the International Energy Agency estimates data center demand at about 415 terawatt hours in 2024 and projects it roughly doubling to around 945 terawatt hours by 2030, slightly more than Japan uses in a year.²⁰ By the end of the decade the United States is on track to use more electricity for data centers than for the production of aluminum, steel, and cement combined.

The grid cannot keep pace

Demand is only half the story. The physical system cannot be expanded fast enough to meet it. At the end of 2023, about 2,600 gigawatts of proposed generation and storage were waiting in US interconnection queues, a figure that has since eased to roughly 2,060 gigawatts by the end of 2025, with typical waits of five years or more, and historically only around a tenth of queued capacity is ever built. The IEA estimates that roughly a fifth of planned data center projects worldwide face delays from grid connection alone. The equipment is its own bottleneck. The lead time for a large transformer has roughly doubled to about five years, and gas turbines, the usual fallback when the grid falls short, now carry delivery windows of five to seven years, with the leading manufacturer selling slots for 2030. Companies have responded by building their own generation on site rather than waiting in line.

The turn to nuclear power

The clearest evidence that demand is real is what the companies are willing to commit to in order to secure power. After years of buying wind and solar credits, several have turned to nuclear for steady, around the clock supply, signing agreements that only make sense if the underlying compute demand is durable. Microsoft agreed a twenty year contract with Constellation to restart a unit at Three Mile Island, renamed the Crane Clean Energy Center, for 835 megawatts of dedicated power, with first power expected around 2027 to 2028 and a project cost of about $1.6 billion supported by a federal loan.¹⁷ Amazon paid $650 million for a campus beside the Susquehanna nuclear plant and invested a further $700 million in a small reactor developer. Google contracted for 500 megawatts from a small modular reactor developer. Across the industry, large technology companies have committed to roughly ten gigawatts of nuclear power in little more than a year.

A selection of roughly thirteen announced nuclear agreements totaling close to ten gigawatts. Sources. Company and Constellation announcements, Utility Dive, DCD.¹⁷

Where this becomes a risk

The same facts that prove demand is real also create pressures that telecom never faced. In the regional grid known as PJM, which serves about 65 million people, the annual cost of the auction that keeps reserve power available rose from about $2.2 billion to $14.7 billion in a single year, and the grid operator's market monitor attributes roughly 40 to 45% of those costs to data center demand.¹⁸ Households across the region are being told to expect electricity bills 1.5 to 5% higher, and elected officials are looking for ways to shift the cost onto the companies driving it. A build-out that raises consumers' utility bills faces a political limit as well as a physical one. There is also the matter of idle capacity: a chip bought today begins losing value immediately, but if the power to run it is years away, much of that value is spent waiting.

Part V  ·  The Financing

Private credit, special purpose vehicles, and leverage off the balance sheet

The statement that the large platforms fund their spending from cash flow is true, and a growing share of the build-out is nonetheless being financed in ways that do not appear on their balance sheets. That combination is the most precise parallel to 2000 in the whole comparison, and the one that is hardest to observe from the outside.

The template was set in October 2025, when Meta and Blue Owl Capital arranged about $30 billion of financing for a data center in Louisiana, the largest private capital transaction on record. Roughly $27 billion of debt and about $3 billion of equity sit in a separate vehicle. Meta owns 20% of that vehicle and outside investors own the rest, with large insurers and asset managers providing the debt. The bonds mature in 2049 and carry an investment grade rating. Meta develops, operates, and uses the site, but it is not the borrower, so the debt does not appear on Meta's balance sheet and its credit rating is unaffected.²³ Oracle has used similar structures, building data centers through vehicles it then leases back.

This is a structural change in how the build-out is paid for, not a handful of deals. Data center related borrowing reached about $182 billion in 2025, roughly double the prior year, and large technology companies issued around $121 billion of corporate bonds, well above their recent average.²¹ Beneath the visible bond market, private credit lent to AI related companies has grown from almost nothing to more than $200 billion in a few years, and reputable estimates see it heading higher still over the rest of the decade.²²

The layers, from steady to fragile

It helps to picture the build-out as layers of capital, from the cheapest and most transparent at the top to the most leveraged and least visible at the bottom. The risk is not at the top. It builds as you move down, and the lower layers increasingly resemble the structure that failed in 2000.

Ordered from steadiest and most visible at the top to most leveraged and least visible at the bottom. The system's true leverage is the sum of all four layers, but only the top two are easy to read from published balance sheets.

Why it is reasonable, and why it is worth watching

There is a sound case that this is sensible financing rather than a warning sign. Long lived infrastructure is being funded by long duration capital, the insurers and pension funds behind the private credit managers, against multi year contracts with creditworthy tenants. Spreading the financing across the system is arguably healthier than concentrating it on a few balance sheets. The case for watching it closely is equally clear. Moving debt off the balance sheet means the true leverage of the build-out is understated by the headline figures, much as structured and vendor financing understated telecom's leverage a generation ago. The collateral inside these structures is not a long lived asset but hardware that ages in a few years. The arrangements are less liquid and less transparent than public markets, so strain can build quietly. And because insurers and pension funds hold much of the paper, the exposure reaches into the retirement and insurance systems. None of this is improper, and none of it is a problem today. It is the place where strain would be hardest to see in advance.

Part VI  ·  The Balanced Read

No single label fits

A careful comparison resists a one word answer. Sorted honestly, the evidence falls into four groups, and a complete picture holds all four at once.

Part VII  ·  Things To Monitor

Signals worth watching

What follows is a set of observable indicators, not advice and not a forecast. Each is something a careful observer can track over time to see whether the dynamics described above are holding steady or shifting. The final item is a structural fact rather than a signal.

1   Earnings quality and cash flow

Because reported profit depends on depreciation assumptions, the cleaner figure to follow is cash. The spread between reported operating income and operating cash flow is a direct read on earnings quality, and persistent negative free cash flow funded by rising debt would mark a meaningful change from the current cash funded model.

2   The revenue gap

The distance between capital spending and AI revenue is the central financial question of the cycle. Whether that revenue keeps compounding toward the spending, or levels off, is observable each quarter, and it is the single clearest indicator of whether the build-out is being justified by demand.

3   Depreciation disclosures

The useful lives that companies assign to servers and chips appear in their filings. Further extensions of those lives would warrant attention, since they flatter near term profit, while shortenings would indicate that the faster replacement cycle is being recognized.

4   The financing layers

The growth and disclosure of off balance sheet vehicles and private credit financing can be tracked over time. The specific things to watch are how much leverage sits below the visible balance sheets, the customer concentration and debt levels at the specialized providers, and whether any vehicle is forced to mark down its hardware collateral.

5   Power and the grid

Power is now the difference between a data center that earns and one that sits idle. The observable signals are the gap between announced capacity and capacity actually energized, the clustering of project delays around grid constraints, and whether rising consumer electricity bills harden into cost shifting or limits on new connections.

6   Index concentration, a structural fact

This last point is not a signal to monitor but a present condition to understand. The ten largest companies are a record 40.7% of the S&P 500, which means a broad index fund now carries an unusually large exposure to the outcome of a handful of AI linked businesses.¹⁹ The base rate from 2000 is worth weighing: owning the largest, most celebrated names at peak valuations produced many years of underperformance even as the underlying technology reshaped the economy. Owning the technology and owning its current leaders remain two different things.