© Deagreez / iStock via Getty Images
Although Wall Street has rewarded every hyperscaler willing to pledge more silicon, more concrete, and more megawatts over the past 24 months, the math behind the AI data center boom is finally drawing scrutiny. Combined 2026 capital expenditure guidance from Alphabet (NASDAQ:GOOGL | GOOGL Price Prediction), Meta Platforms (NASDAQ:META), Amazon (NASDAQ:AMZN), and Microsoft (NASDAQ:MSFT) now sums to roughly $800 billion. But Janus Henderson analysts cited on the latest This Week in Tech podcast project that only 84 of the 157 gigawatts of announced AI data center capacity will actually be built by 2030. That gap is roughly half of what has been promised to the market.
What’s particularly troubling is that the historical mirror Wall Street keeps reaching for, the late-1990s fiber-optic overbuild, may understate the risk rather than overstate it.
The Buildout, in Numbers Nobody Quite Believes
On the TWiT panel, host Leo Laporte noted that Google is raising $80 billion via a stock issue to fund data center construction, with $10 billion coming from Berkshire Hathaway, while Meta’s parallel capex announcement hurt its stock. Father Robert Balassare tallied the cumulative AI spend: Amazon close to $291 billion, Alphabet around $262 billion, and Meta roughly $227 billion. Those are running totals on balance sheets that, until very recently, were the cleanest in corporate America.
The Q1 earnings reports confirm the pace. Alphabet raised 2026 capex guidance to $180 billion to $190 billion, with CFO Anat Ashkenazi adding that “we expect our 2027 CapEx to significantly increase compared to 2026.” Microsoft now expects roughly $190 billion in calendar 2026 capital expenditures. Meta lifted its full-year guide to $125 billion to $145 billion. Amazon’s Q1 capex alone hit $44.2 billion, which helped drag trailing twelve-month free cash flow down 95% to $1.2 billion.
The Long Memory: 1999 Fiber, Plus a Twist
The closest precedent is the telecom and fiber overbuild of 1999 to 2001, when carriers laid roughly $500 billion in fiber capex and an estimated 85% to 95% of the strands went dark for years. Cisco, Sun Microsystems, and Lucent rode the same enthusiasm, then watched their order books evaporate. Equity values fell more than 80%. The fiber, however, eventually became the substrate for the modern internet.
I’ve been watching hyperscaler capex disclosures for the better part of three years, and Balassare’s argument is the cleanest articulation of why this cycle might rhyme without repeating. When these facilities stop being used, “that’s coming up very, very fast, they’re useless,” he said, pointing to xAI’s $20 billion facility already running NVIDIA chips that newer models have surpassed. Dark fiber sat in the ground and waited. A rack of last-generation GPUs depreciates on a schedule closer to a smartphone than a railroad tie. His bottom line: “Yes, they will build them. Yes, they will use them. But whether or not they’re going to make money off of them, that’s a huge unknown.”
What (Short Term) Results Are Already Saying
The stocks have started to hear the question. Since its April 29 earnings report, Meta has fallen 15% to $570.98. Amazon is down 10% over the same window. Microsoft has slid 6%. NVIDIA (NASDAQ:NVDA), the supposed beneficiary of every dollar in this cycle, is off 10% since its May 20 earnings report, even after disclosing $119 billion in total supply-related commitments. Polymarket’s June 2026 contract pins NVDA at a modal $192 with 59% probability. Traders are not pricing a melt-up.
Financing is not getting cheaper. The 10-year Treasury yield sits at 4.53%, in the 96th percentile of its historical range. Every basis point makes a fifteen-year data center investment harder to pencil.
What to Watch
Long term, Wall Street still heads higher in the decades to come. NVIDIA’s data center revenue grew 92% year over year, Microsoft’s AI business run rate hit $37 billion, up 123%, and Google Cloud backlog nearly doubled quarter on quarter to $462 billion. The demand is real. So was the demand for internet bandwidth in 2000.
The Long Memory’s lesson is that overbuilds historically produce overcapacity, write-downs, and a long digestion period before the survivors are sorted from the casualties. If Balassare is right that the obsolescence clock ticks faster on a GPU than on a glass fiber, the digestion will be uglier than the bulls expect. The companies will still be standing. The open question is what kind of returns the shareholders financing this cycle actually take home.
For me? I’m long every one of these companies. I’m willing to hold them through the unknown. If you’re long them as well, make sure you have the same mindset.
