Three months of AI capex coverage predicted the spending correctly and missed what mattered. The telecom dark fiber precedent was available from the start.
Between February and May 2026, this journal published a series of entries tracking the AI capital expenditure cycle. The Industrial Turn argued the Magnificent Seven were spending like industrials but priced like software. The Capex Verdict went long the proof against the promise. The Cost Basis found component inflation hiding inside headline numbers. The Scorecard graded The Industrial Turn after one week of earnings.
Q2 2026 earnings are now complete. All four hyperscalers revised capex guidance upward. Three months of predictions can be evaluated against outcomes.
What We Got Right
The spending prediction was correct beyond our own estimates. Microsoft guided $190 billion for the full year. Amazon committed $200 billion. Alphabet raised to $180 to $190 billion. Meta increased its range to $125 to $145 billion. Combined hyperscaler capex is on track to reach approximately $725 billion for the year.
NVIDIA reported $81.6 billion in Q1 revenue, beating the $79.2 billion consensus, and guided $91 billion for Q2. The supplier thesis from The Industrial Turn held: component cost inflation transferred billions from hyperscalers to the semiconductor supply chain.
The capex acceleration was real and larger than consensus expected.
What We Missed
The journal asked the right question too late. We identified that the Magnificent Seven were spending at industrial scale. We did not weight a more immediate question: when would markets demand evidence that the spending was producing returns?
The Scorecard caught a version of this after one week. Alphabet rose 7 percent because Google Cloud grew 63 percent with operating margins that nearly doubled. Meta fell 7 percent despite beating revenue because it disclosed no quantified AI revenue metric. Revenue visibility, not capex magnitude, was the market's sorting function.
By May, the pattern had deepened. NVIDIA posted the largest beat in semiconductor history and the stock fell on sell-the-news dynamics. The Russell 2000 surged 2.56 percent the same day as capital rotated from growth into value. Year to date, the rotation had widened to 13 percentage points: value up 8.5 percent, growth down 4.7 percent.
No hyperscaler demonstrated positive AI infrastructure ROI at scale. Microsoft disclosed $37 billion in AI annualized revenue against $190 billion in capex. Meta's AI improved ad targeting but generated no separately reportable revenue. The spending was real. The returns were projected.
The Dark Fiber Precedent
The telecom boom of 1999 to 2001 is the sharpest historical analog for the current AI capex cycle. Telecom companies invested more than $500 billion in fiber optic infrastructure in the five years after the Telecommunications Act of 1996. Peak annual spending reached $120 billion in 2000 dollars.
By 2001, only 5 percent of installed fiber carried traffic. By 2002, utilization had fallen to 2.7 percent. By 2005, 85 percent of fiber remained dark. Meaningful demand recovery did not begin until 2006 to 2008, five to seven years after the spending peak.
The infrastructure eventually became essential. Fiber became the backbone of the internet economy. But equity holders were destroyed during the gap. WorldCom reached a peak market value of $150 billion in January 2000 and filed for bankruptcy in July 2002, destroying approximately $175 billion in shareholder value. Twenty-three telecom companies failed in the initial cascade. Breakeven on fiber infrastructure required 10 to 15 years.
The current AI capex cycle is 1.45 times larger than the telecom peak in inflation-adjusted terms. The demand assumption may prove correct. The equity assumption has no historical support.
The Finding
This journal confused two questions. The first: would hyperscalers continue spending at accelerating rates? The answer was yes, and we called it accurately from The Industrial Turn onward.
The second: would markets reward that spending before demonstrated returns arrived? The answer was no. We were late to frame it.
The Industrial Turn and The Capex Verdict treated revenue visibility as a differentiator within the group. They did not ask whether the entire group faced the same durational risk that destroyed telecom equity holders: infrastructure built five to seven years ahead of demand, where the infrastructure becomes essential but the equity does not.
The supplier trade also carries an expiration date. Component vendors capture value during the build phase. When the build phase ends or decelerates, the trade reverses. The Industrial Turn recognized the build phase. It did not ask when.
The telecom precedent was available from the beginning. We should have used it earlier. Three months of coverage produced a correct call on spending, a partially correct call on relative value, and a missed call on the market's timeline for demanding proof of returns.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)