Is AI Capex Already Bigger Than the Dot-Com Telecom Boom? Three Investment Cycles, Measured by Share of GDP

Data-center and AI-related investment accounted for an estimated 74% of all U.S. GDP growth in the first quarter of 2026, according to the Bureau of Economic Analysis's own sub-component data. That is not a typo, and it is not the only time in the last thirty years that one investment cycle has carried the whole economy. Here is how today's AI buildout compares with the late-1990s telecom boom and the mid-2000s housing boom, measured the same way: as a share of GDP.

Map of data center infrastructure locations across the United States, published by the U.S. Department of Energy's National Renewable Energy Laboratory.
Data center infrastructure in the United States, 2025 map. Source: U.S. Department of Energy / National Renewable Energy Laboratory (NREL), a U.S. federal government work in the public domain. Wikimedia Commons file page.

Key takeaways

  • The BEA's third estimate of Q1 2026 GDP shows information-processing equipment and intellectual-property investment — the categories that capture AI data-center spending — contributed an estimated 1.55 of the 2.1% annualized growth rate, or roughly 74%.
  • Hyperscaler capital spending has grown from 0.3% of GDP in 2019 to 1.4% in 2025, and is on a consensus path toward roughly 3% by 2027, per Apollo Global Management's chief economist.
  • That already exceeds the late-1990s telecom and fiber buildout's peak of 1.2% of GDP in 2000, and the pace of increase is roughly 5.7 times faster than telecom's and 1.7 times faster than the fastest years of the 2002–2005 housing boom.
  • In level, the data-center buildout still sits below the housing boom's 2005 peak of 6.6% of GDP — this is a faster cycle, not yet a bigger one.

Figures checked: November 6, 2026.

What the Official Growth Data Actually Shows

On June 25, 2026, the BEA released its third and final estimate of first-quarter 2026 GDP: real GDP grew at a 2.1% annualized rate, revised up half a percentage point from the second estimate of 1.6%. The BEA does not publish a single "AI contribution" line item. But its NIPA Table 1.5.2 breaks nonresidential fixed investment into sub-components, and summing information-processing equipment (servers, GPUs, networking gear) with intellectual-property products (software) gives an estimated 1.55 percentage points of that 2.1% headline — about 74% of the quarter's entire growth. Personal consumption, which normally makes up roughly 70% of GDP, contributed only a small fraction of the growth rate after a sharp downward revision, and residential investment subtracted from growth for a consecutive quarter.

The arithmetic is worth sitting with for a moment. If information-processing and IP investment had contributed nothing to Q1 2026, the headline growth rate would have been closer to 2.1% − 1.55% = 0.55 percentage points annualized — not a contraction, but close to Q4 2025's already-weak 0.5% pace. The 2.1% headline is not describing a broad-based acceleration; it is describing one category of spending that happened to accelerate sharply while most of the rest of the economy did not.

Three Investment Cycles, Measured by Share of GDP

Comparing today's AI buildout with past boom-bust cycles only works if the comparison uses the same yardstick every time. Capital spending as a percentage of GDP is that yardstick, because it strips out inflation and the economy's overall size and asks a narrower question: how much of the nation's total output is this one category of investment absorbing?

CycleStarting sharePeak share of GDPFastest pace of increase
Telecom / fiber buildout (late 1990s)n/a1.2% in 2000~0.15 percentage points per year
Housing (2002–2005)5.1% (2002)6.6% in 20050.5 percentage points per year
Hyperscaler AI data-center capex0.3% (2019)1.4% in 2025; ~3% projected by 2027–2029~0.85 percentage points per year (2025→2027)

Three conclusions follow directly from this table. First, in sheer level, today's AI buildout (1.4% of GDP in 2025) is already bigger than the telecom peak (1.2% in 2000) and still less than half the housing peak (6.6% in 2005) — it sits between the two previous cycles. Second, in cumulative change, the data-center buildout's rise from 0.6% of GDP in 2023 to a projected 3.1% in 2027 is a 2.5-percentage-point move, larger than either prior cycle's own rise. Third, and most strikingly, in speed: going from 1.4% to 3.1% of GDP in just two years works out to about 0.85 percentage points a year, which is roughly 1.7 times the housing boom's fastest two-to-three-year stretch (0.5 points a year) and about 5.7 times the telecom boom's pace (0.15 points a year) — my own division of Apollo's reported figures, shown so the comparison can be checked.

What Happened When Each Previous Cycle Unwound

The level and the speed both matter because of what happened when these two earlier cycles reversed. Telecom's share of GDP fell by a relatively small amount after 2000, and the recession that followed — beginning in March 2001 — is remembered by the National Bureau of Economic Research as the mildest post-war U.S. recession, lasting eight months. Housing's unwind was far larger: its share of GDP fell from 6.2% in early 2006 to 3.0% by the end of 2008, a swing more than four times the size of telecom's reversal, and the recession that followed it was the most severe since the 1930s.

The pattern across both cycles is consistent: a capex cycle that climbs at a given pace tends to be able to unwind at a similar pace, and a bigger, faster buildout creates a bigger hole if the spending stops. Applied to the AI data-center cycle, that is the specific risk, not a vague one: a buildout rising at roughly 0.85 percentage points of GDP a year could, mechanically, subtract a similar amount from growth in a downturn — more than four times telecom's drag, even if it never reaches housing's absolute 2005 peak.

The Counter-Argument: Why This Buildout Might Not Behave Like 2000

It would be a mistake to treat the GDP-share comparison as a verdict that a crash is coming on any particular timetable, and the two prior cycles differ from this one in a way that matters for how an unwind would actually play out. Much of the late-1990s telecom buildout was financed by competitive carriers that borrowed heavily in high-yield debt markets to build fiber networks ahead of demand that never fully materialized; several, including WorldCom and Global Crossing, filed for bankruptcy within two years of the peak. Today's largest AI capital spenders — Microsoft, Alphabet, Amazon, Meta, and Oracle — are, for the most part, funding this buildout from operating cash flow and investment-grade debt rather than speculative junk financing, which is a meaningfully different starting balance sheet than the companies that drove the telecom bust.

That distinction is a reasonable basis for thinking this cycle's unwind, if one comes, would look less like the 2001 telecom bust and more like a slowdown than a wave of corporate bankruptcies — this is the writer's own reading of the balance-sheet differences, not a forecast. It does not remove the mechanical point above: cash-rich companies can still cut capital spending quickly if the revenue they expect from AI products does not show up, and a 74%-of-growth concentration in one investment category is, by definition, a narrow base for an economy to stand on. Both things can be true — the financing is sturdier than 2000's, and the growth composition is still unusually narrow — and a reader deciding what to do with that information should weigh both rather than picking the one that fits a preferred conclusion.

Related reading

FAQ

Is AI capex really bigger than the dot-com telecom boom?

By share of GDP, yes, already: hyperscaler data-center capex reached 1.4% of GDP in 2025, above telecom's 1.2% peak in 2000, and consensus estimates point toward roughly 3% by 2027–2029. In absolute size relative to the economy, it is not yet bigger than the 2002–2005 housing boom's 6.6% of GDP peak.

Did AI spending really drive 74% of GDP growth in a quarter?

That figure comes from an analyst calculation, not an official BEA line item: summing the information-processing-equipment and intellectual-property-products sub-components of the BEA's third estimate of Q1 2026 GDP gives about 1.55 of the 2.1% annualized growth rate, or about 74%. The BEA does not publish a single "AI contribution" figure.

Why does the speed of the buildout matter more than its size?

Because a cycle's unwind tends to be roughly as fast as its buildup. The AI capex share of GDP is rising at about 0.85 percentage points a year, versus roughly 0.5 points a year at the fastest point of the housing boom and about 0.15 points a year for telecom — so a reversal, if one happens, has mechanically more room to subtract from growth than telecom's did in 2001.

Does this mean a recession is coming?

Not on any specific date. This is a measure of how concentrated recent growth is in one investment category, not a prediction of when or whether that spending stops. The two historical cases differ in outcome — a mild recession after telecom, a severe one after housing — partly because of how each buildout was financed, which is also different again for today's hyperscalers.

Where do these numbers come from?

The GDP growth contribution figures come from the BEA's third estimate of Q1 2026 GDP (released June 25, 2026) and its NIPA Table 1.5.2 sub-components. The capex-share-of-GDP comparison across the three cycles comes from Apollo Global Management chief economist Torsten Slok's August 6, 2026 analysis, which itself cites BEA, FactSet, and Haver Analytics data.

Tip: Comparing one capex cycle to another is easier when you have read how the earlier ones ended — the telecom and railway booms are standard case studies in a good stock market investing book. Writing down what you expected each cycle to do, and checking back later, is what an investment journal is for. (These are Amazon Associate links — we may earn a small commission on qualifying purchases.)

Sources

This article is for general investor education only and is not personalized investment, tax, or legal advice. Figures are sourced and dated as noted above; consult a licensed financial professional about your own situation.

Get new market analysis by email. Free, unsubscribe anytime.


Start Here: every analysis by topic · About · Privacy Policy

Comments

Post a Comment

Popular posts from this blog

Did High-Yield Credit Spreads Really Widen 29%? What the FRED Series Actually Shows

Image
The New York Stock Exchange -- credit spreads, traded in the bond market, often move before equity prices do. Photo: Carol M. Highsmith/Library of Congress, Wikimedia Commons , public domain. Key takeaways The ICE BofA US High-Yield Option-Adjusted Spread closed at 3.44% on October 30, 2026 and 3.43% on November 2. That is not a 2026 high: the series printed 3.46% on March 30, 2026. The "29% in five weeks" framing depends entirely on which day you start from. Measured over the last 25 trading days to October 30 the rise is 17.4%, from 2.93% on September 25. Measured from the actual 2026 low - 2.60% on August 28, not 2.66% on September 21 - it is 32.3%, over two months rather than five weeks. Most of the move is two days old. The spread was 3.14% on October 28 and 3.44% on October 30, a 9.6% jump in two sessions after four straight sessions of narrowing. The 2008, 2016 and 2020 crisis peaks cannot be checked against this series any more. FRED's own series ...
Read more

Gold Is Down 30.7% From Its January 2026 Record. Here's How That Compares With Gold's Five Worst Bear Markets on Record

Image
Gold's record-breaking rally ended in the dead of winter, and the reversal hasn't let up since. Spot gold peaked at $5,589.38 per troy ounce on January 28, 2026, then fell roughly 30.7% to $3,875.84 by October 30, 2026 — a nine-month round trip from a fresh all-time high back to levels last seen before the final leg of its 2025–2026 run. The real question for anyone holding gold, a gold ETF, or shares in a gold miner isn't whether this counts as a correction; a decline past 20% from a peak already clears that bar on its own. It's whether this looks like any of the five genuine gold bear markets the World Gold Council has documented since 1974 — and on the one measure that matters most for sizing a position, speed, this one is already moving faster than every single one of them. That speed is the real headline here, not the size of the drop. A 30.7% decline is, so far, smaller than every one of the five historical bear markets detailed below. But reaching it in roughl...
Read more

S&P 500 Midterm-Year Pattern Since 1946: Does the 20-for-20 Record Hold for October 2026?

Image
Update November 6, 2026: October 2026 is now complete. The S&P 500 closed the month at 7,580.07 on October 30, down 0.93% from its September 30 close of 7,651.54 and up 10.73% for the year. Its highest daily close of 2026 was 7,818.93 on October 6, so the 7,844.52 “52-week high” cited below is not a closing level. On Election Day, November 3, the index closed at 7,568.92, 3.11% below the October 9 reference close used in this article; that close is the starting level for the 12-month post-election window Truist and UBS track (S&P Dow Jones Indices data via FRED, series SP500). One line about midterm years circulates every two years: since 1946, the S&P 500 has been positive in the twelve months after every single midterm election — twenty for twenty. It is the kind of statistic that gets repeated without anyone opening the spreadsheet. What follows takes the claim apart in three checks: what the counting window really is, what the raw annual returns say wh...
Read more

The Stock Market's "Best Six Months" Start November 1. Here's What 76 Years of Data Actually Show

Image
Every November 1, the same statistic circulates: since 1950 the Dow Jones Industrial Average has averaged a 7.3% gain from November through April against just 0.8% from May through October. The claim is easy to repeat and hard to check, because the two most-cited sources publish different numbers, neither publishes the hit rates that get quoted alongside them, and the arithmetic example that usually follows does not reconcile. So this version is built out of comparison tables: one table per claim, with a column for what we could verify ourselves. The New York Stock Exchange on Wall Street. Photo: Arnoldius, CC BY-SA 3.0 , via Wikimedia Commons . File photo, not tied to any specific trading session discussed in this article. Key takeaways Fidelity, citing FactSet data since 1945, puts the S&P 500 at roughly 7% for November-April against about 2% for May-October. Those are price returns, dividends excluded. Our own count on 81 November-April windows since 1945 (Shiller monthl...
Read more

Is a Sudden VIX Spike a Warning Sign? What October 30, 2026 Looks Like Next to 2018 and 2024

Image
The VIX closed at 19.42 on October 30, 2026 , up 2.22 points or 12.91% from 17.20 the session before, as an AI-spending sell-off took the S&P 500 down 1.82% and the Nasdaq Composite down 2.86%. Two sessions in the modern record are the ones people reach for when a volatility index jumps: February 5, 2018 and August 5, 2024. Put all three side by side and the 2026 move is the smallest of them by a wide margin - and it is smallest for a reason that has nothing to do with how bad the day felt. New York Stock Exchange trading floor, 1963. Photo by Thomas J. O'Halloran, U.S. News & World Report collection, Library of Congress , public domain. Key takeaways February 5, 2018: VIX 17.31 to 37.32 , a 115.6% single-session jump, off a base that had touched an all-time intraday low of 8.56 ten weeks earlier. Dow -1,175 points (-4.6%). August 5, 2024: VIX 23.39 to 38.57 , up 64.9% , off an already elevated base, on a yen carry-trade unwind. S&P 500 -3%. October 30, 202...
Read more

Are S&P 500 Buybacks Flashing the Same Warning Sign as 2007 and 2018?

Image
Correction November 10, 2026: the 1% buyback excise tax was described as cutting Q3 2025 operating earnings by "roughly 0.36%–0.41%" and as already reflected in the reported buyback totals. S&P Dow Jones Indices reports 0.36% for operating earnings and 0.41% for As Reported (GAAP) earnings as two separate figures, and the tax reduces earnings rather than the buyback dollars; the buyback-yield passages now state that the 1.4% figure is this article's own calculation rather than a third-party tracker's. Source: S&P Dow Jones Indices, "S&P 500 Q3 2025 Buybacks" press release, December 18, 2025. Berkshire Hathaway told investors on November 7, 2026 that it spent roughly $9.6 billion buying back its own stock in the third quarter, its biggest buyback quarter ever. That announcement landed three weeks after Ned Davis Research reported that net S&P 500 buybacks had slipped back below the $1 trillion mark (on a trailing 12-month basis) for the first...
Read more

The S&P 500's Dividend Yield Just Hit an All-Time Low of 1.05%. Here's Why It's Not Quite 2000 Again

Image
The S&P 500's dividend yield is the lowest it has ever been. The record itself is not in dispute — the series that produces it labels its own minimum as 1.05%, set in October 2026. What is worth an hour of your attention is that almost every word in the sentence "the dividend yield just hit an all-time low" is doing more work than it looks like, and three of them are doing something different from what most write-ups assume. Photo: Bruno Sanchez-Andrade Nuño (Flickr) · original file page · CC BY 2.0 . Key takeaways The record is real: the series' stated minimum is 1.05%, October 2026 , against a 155-year mean of 4.20% and a maximum of 13.84% in June 1932. The most recent daily reading we could verify is 1.07% at the close on Friday, October 30, 2026 . Readings after June 2026 are estimates , not reported figures: the numerator is fixed at twelve months of dividends through June 2026. Every move since then is a price move by construction. The prio...
Read more

The Personal Saving Rate Just Fell to 3.8% — the Lowest Since October 2022, and the 7th Percentile of 67 Years of Data

Image
Correction November 11, 2026: The original headline and opening paragraph said that only the mid-2000s pre-crisis years were consistently lower than September 2026's 3.8%. FRED's PSAVERT series shows eight consecutive months from March to October 2022 at 2.4%-3.6%, so the accurate statement is that 3.8% is the lowest reading since October 2022. Also corrected: the 2006-2007 band (1.9%-3.3%, not 1.9%-2.4%), the 1998-2001 range (3.7%-4.8%, which applies to September 1999-December 2000), the count of monthly declines in the past year (six of twelve, not nine) and the recession-pattern claim (two of three episodes, not two of two). Source: Bureau of Economic Analysis Personal Saving Rate (PSAVERT) via FRED, Federal Reserve Bank of St. Louis. Correction November 12, 2026: The episode table still showed the pre-correction 2006-2007 band of 1.9%-2.4%; it now shows 1.9%-3.3%, matching the text. The 1998-2001 row now states that the 3.7%-4.8% range covers September 1999 through Decembe...
Read more