"History doesn't repeat… but it rhymes." — Mark Twain
◉ THE PRESENT
Microsoft told the Street last night it is cutting fiscal-2027 AI capital spending from a planned $124 billion down to $62 billion, the largest single-year capex reset by a hyperscaler on record. The cut lands six months after the AI cohort lost a quarter of its market cap and four weeks after Meta and Oracle quietly canceled data-center leases in Texas and Arizona. By the time the press release crossed the wire at 4:11 p.m., Nvidia was already down nine percent in the after-hours tape and the futures had given back the week.
MSFT −13.4% AFTER HOURS | AI CAPEX CUT $62B | NVDA −9.1% | SOX −7.8%
When the biggest buyer in a market stops buying, the bubble was already gone. We saw the same letter, almost word for word, in a press release sent out twenty-four years ago this spring.
◉ THE ECHO — APRIL 16, 2002
A San Jose warehouse, full of routers nobody wanted.
It was just after seven in the morning when John Chambers walked into Building 10 on Tasman Drive. He had been at Cisco for ten years and CEO for seven. He had told the analysts in 2000 that the network would grow forever, that bandwidth demand was doubling every hundred days, that Cisco's customers needed every switch and router the company could ship. He believed it. So did the customers, who placed orders eighteen months out and paid for them with debt.
Now it was April 2002 and the warehouse was full. Pallets stacked twelve high. Forty thousand square feet of Catalyst switches that no one was going to buy because the customers were dead. Worldcom was burning. Global Crossing had filed in January. Williams Communications was about to. The dark fiber those companies had laid across the country, more than eighty million miles of it, sat unlit in the ground. Industry surveys would later show that ninety-two percent of it was never turned on.
That evening Cisco filed an 8-K cutting capital spending guidance by thirty-one percent and writing down $2.25 billion in inventory. Sun Microsystems followed two weeks later with a $2 billion writedown of its own. JDS Uniphase had already taken the biggest charge in corporate history, $44.8 billion, the previous July. Nortel was on its way to laying off sixty thousand people. The story everyone had been telling — that the internet was a one-way demand ramp — was over. The story they hadn't been telling, that demand had been pulled forward by speculative buildout financed with junk bonds, was the one that mattered.
The Nasdaq had closed at 5,048 on March 10, 2000. The day Chambers cut capex it was 1,816. Most of the people who owned it on the way up did not own it on the way down. They had sold by then, or been sold out, or stopped opening their statements. The S&P had lost a quarter of its value and was about to lose another quarter. Alan Greenspan had cut the funds rate eleven times in eighteen months, from 6.50 to 1.75, and the market did not care.
That is the day Microsoft is rhyming with tonight.
◉ THE RHYME — WHAT'S IDENTICAL
Capex cycles do not end with a quiet glide path. They end when the biggest spender admits the demand they were building for never showed up. Cisco said it on April 16, 2002. Microsoft said it last night.
◉ THE DIVERGENCE — WHAT'S DIFFERENT THIS TIME
The pattern is the same. The plumbing is not. Four things separate this from the spring of 2002, and each one cuts differently.
The hyperscalers are not Cisco. Microsoft has $84 billion in cash on the balance sheet and no debt to speak of. The telecom buildout in 2000 and 2001 was financed by junk bonds at Worldcom, Global Crossing, Williams, and a dozen others that were already insolvent when Cisco cut. There will be no equivalent of the high-yield wave of defaults from June 2002 through 2003. The pain this time concentrates in the equity, not the credit market.
AI has actual revenue. Copilots, agents, code generation, and inference workloads inside enterprises pulled in roughly $180 billion in real software revenue last year. Compare that to 2001, when the internet was producing banner ads and a few subscription businesses that lost money on every customer. The bubble was demand pulled forward, not demand that never existed. That changes the recovery shape.
The buyer concentration is worse, not better. Five companies are responsible for roughly seventy-eight percent of AI infrastructure spend. In 2001 the telecom buildout had hundreds of buyers, most of them small and badly capitalized. Today's overhang is more cleanly removable, but when those five decide together to slow, the supply chain feels it as a step function, not a slope.
The sovereign overlay is new. Saudi Arabia, the UAE, India, France, and a half-dozen others have committed to AI data-center buildouts driven by national-strategy money, not return on invested capital. That demand does not turn off when Microsoft's CFO turns hers off. It is smaller in dollars than the hyperscaler complex, but it does not respond to the same signals, and it puts a floor under Nvidia's customer mix that Cisco never had.
◉ THE RECKONING — WHAT HAPPENS NEXT
From April 16, 2002, the Nasdaq did not bottom for another six months. It fell another thirty-eight percent to 1,114 on October 9, then ground sideways through 2003 before the slow climb back. Cisco, the company that announced the cut, never reclaimed its March 2000 peak. Not in 2007, not in 2021, not yet. It still trades below where it was the day Chambers signed that 8-K.
The damage in the supply chain was worse than at the headline names. JDS Uniphase finished the year down ninety-eight percent from peak. Sun Microsystems lost ninety-six percent. Lucent, Nortel, Ciena, Corning — every name that had supplied the buildout — bottomed somewhere between minus eighty-five and minus ninety-nine. Most of them were acquired, broken up, or quietly delisted by 2008. The companies that survived survived because they had cash, owned distribution, and could buy the pieces of their own ecosystem for ten cents on the dollar.
That is where the smart money went. Not into shorting Cisco, which had already done most of its falling by April. Not back into the broad index, which would lose another twenty-five percent. Into the survivors with cash who could buy the wreckage. Microsoft itself, then trading at $52, was one of them. Oracle was another. Both spent the next three years acquiring fragments of the dot-com supply chain at fractions of replacement cost. By 2008 they had emerged with infrastructure positions that defined the next fifteen years.
The capex cut is not the bottom. It is the announcement that the cycle has changed. The next eighteen to thirty months in the echo were spent watching the supply chain unwind, the survivors hoard cash, and a new leadership group form in places nobody was looking — energy, financials, industrials, emerging markets. The NDX did not lead again until 2009.
The pattern: when the biggest spender cuts, the supply chain has not yet priced it in, the index has more to give back, and the next leadership cohort is not in the headlines. The smart move in 2002 was not buying the dip in the names that just fell. It was watching who had cash, what they were buying, and where they were buying it. The same map applies tonight.
◉ TOMORROW’S WATCH
Watch the high-yield AI debt issued by the smaller cloud players over the last eighteen months — CoreWeave, Lambda, Nebius, Crusoe. Spreads on that paper have been quiet. They were quiet at Worldcom too, right up until February 2002.
