Microsoft Cut 3.5GW of AI. Big Tech Bet $660B.

Key Takeaways

• Microsoft pared back roughly 3.5GW of near-term AI capacity over the past year. In a separate move, seven hyperscalers, Microsoft included, signed the White House Ratepayer Protection Pledge committing to take-or-pay treatment for new data center power.
• Aggregate 2026 hyperscaler capex hit roughly $660 billion, with Alphabet alone committing $175–185B, nearly double its 2025 spend.
• Gas turbines are sold out through 2030. GE Vernova’s Q1 backlog grew to $163B with new data center orders booked through the end of the decade.
• Earnings drop Wednesday April 29. Microsoft, Alphabet, Meta, and Amazon all report after the close on the same day. The contradiction between near-term cuts and long-term commitments lands on the calls.

Two Numbers That Don’t Add Up

Over the last twelve months, Microsoft has walked away from more than 2 gigawatts of non-binding data center lease letters of intent and frozen 1.5 gigawatts of self-built projects that had been scheduled to come online in 2025 and 2026. That sums to roughly the continuous power draw of 3 million U.S. homes, capacity that will not be coming online on the original schedule.

The same company spent $29.9 billion on property and equipment in fiscal Q2 2026 alone (calendar Q4 2025), bringing first-half FY2026 capex to roughly $72 billion. Microsoft has said roughly two-thirds of recent quarterly capex is going to short-lived AI assets like GPUs and CPUs, with the balance funding data center shells and grid build. Meanwhile, Alphabet has committed $175 to $185 billion in 2026 capex, almost double its 2025 spend of $91.4 billion. Meta, fresh off cutting 8,000 employees, guided to $115–$135 billion. Add Amazon and the rest, and the aggregate hyperscaler bet for 2026 lands around $660 billion.

The industry can freeze near-term builds while raising long-term capex commitments only if the deployment date is sliding. A large share of this year’s announced spending will not light up gigawatts this year. It is purchasing slots, equipment, and grid access for capacity that comes online several years later.

The Pledge That Took Away the Off-Ramp

On March 4, 2026, seven companies (Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI) signed the Ratepayer Protection Pledge at the White House. The political framing was consumer protection: hyperscalers would not pass the cost of AI to ordinary utility ratepayers.

The fine print is the story. Signatories committed to pay rates for the power and related infrastructure brought online to service their data centers, whether they use the electricity or not. That mirrors the structure of a take-or-pay arrangement, codified as a high-profile policy commitment by seven of the most prominent AI and cloud companies as named signatories.

Take-or-pay is a structure familiar to oil and gas, not software. You pay for the gas pipeline whether or not you ship a molecule through it. The pipeline operator gets paid; the user takes the volume risk. Hyperscalers just put themselves on the operator side of that bet for electricity.

The implication: if AI revenue softens, hyperscalers cannot quietly walk away from the power they have lined up. The pledge itself is a policy commitment rather than a courtroom-enforceable contract, but the underlying utility deals it commits signatories to negotiate are real legal contracts with breakup costs. The political cost of being seen to abandon the pledge is high; the contractual cost of unwinding any signed take-or-pay deals is higher. After March 4, the off-ramp on long-dated commitments narrowed. Pulling back is no longer free.

The 2028 Bet

So why are CFOs spending now for power they will not use until 2028?

Because the physical infrastructure for AI has multi-year lead times that cannot be compressed with money alone. GE Vernova, which has the leading market share for heavy-duty H-class gas turbines installed in the United States, reported that its gas power equipment backlog plus slot reservations grew from 83 to 100 gigawatts by the end of Q1 2026. New turbines ordered in early 2026 will not arrive at a customer site until late 2028 at the earliest, and slots are tightening through 2030. Gas turbine lead times have stretched to four and five years across the industry.

Generator step-up transformers (the iron boxes that step new generation up to grid voltage) average around 143-week lead times, with broader large power transformer averages near 128 weeks. Copper, which costs about 27 tons per megawatt of new data center capacity, is running a 150,000- to 400,000-tonne refined deficit in 2026, with prices that hit record highs in January.

The math is brutal. Even if every dollar of the $660 billion in announced 2026 capex is real, the gigawatts cannot be deployed in 2026. They will be deployed in 2028, 2029, and 2030, when turbines, transformers, and copper actually arrive. So the 2026 capex line item is, mostly, the down payment on physical infrastructure that lands two to four years later.

That helps explain Microsoft’s portfolio reshuffle. Microsoft has publicly said it is short on capacity, yet it is also walking away from specific 2025 and 2026 builds. The two facts are reconcilable if Microsoft is reweighting its near-term queue away from sites that no longer fit the GPU-optimized profile, while reserving long-dated slots (2028 and later) where giving up a turbine reservation means losing it for years.

$$\text{Useful 2026 GW} = \frac{\text{Capex Spent}}{\text{Cost per GW}} \times \text{Deployable Fraction}$$

The deployable fraction is what kills the optimism. Industry build estimates put the all-in cost of new AI-class data center capacity in a range of several billion dollars per gigawatt, depending on power source, cooling, and grid interconnect. Naively, the announced capex translates to tens of gigawatts. But only a fraction lands as live capacity in calendar 2026, because turbines, transformers, and grid interconnects are the binding constraints, not money. The rest of the spend sits as work-in-progress, options on slots, and cash committed to deals that complete in 2028 or later.

The Bottleneck Is Already Cashing Checks

If you want to know whether the AI capex is real, do not look at hyperscaler revenue. Look at GE Vernova.

The Q1 2026 results, released April 22, were extraordinary. Quarterly orders hit $18.3 billion, up 71 percent organically year-over-year. Backlog grew to $163 billion. The Electrification segment alone booked $2.4 billion in equipment orders to support data centers in just three months, more than the segment’s data center orders for all of 2025. The company raised 2026 financial guidance.

This is the picks-and-shovels trade made visible. Big Tech announced roughly $660 billion in 2026 capex. GE Vernova, Vertiv, Eaton, and a small list of grid equipment suppliers are the entities physically receiving that money in the form of turbine deposits, electrical equipment orders, and slot reservations. The “AI revenue” the public market keeps demanding from Microsoft and Alphabet has already arrived. It just landed on the income statements of industrials.

For more on the specific FERC ruling that forced hyperscalers to start paying for grid load, see The Co-Location Cap: FERC Ends Big Tech’s Grid Free Ride. For the bear case on what happens if AI demand never catches the hardware, see Why 2026 is AI’s Valley of Death.

A Pattern From the 1990s

The closest historical analogy is not the 2000 dot-com bust. It is the 1996–2000 telecom buildout that preceded it.

In the late 1990s, long-haul carriers like WorldCom, Global Crossing, and Qwest committed to building national fiber backbones, with traffic-growth claims circulating in the industry that internet usage was doubling every three to four months. Cisco, the picks-and-shovels equivalent of GE Vernova, hit roughly a $555 billion market cap on March 27, 2000, briefly the most valuable publicly traded company in the world. Lucent and Nortel were also peaking at large but lower market caps in the same window selling optical equipment.

The carriers paired the build-out with extensive long-dated capacity-swap and indefeasible-right-of-use deals, some with take-or-pay terms, to lock in supply and revenue for capacity that would not be lit until 2001 or 2002. When the traffic forecasts and accounting both unraveled, the financial pressure rolled through the sector quickly. Global Crossing filed for Chapter 11 on January 28, 2002. WorldCom filed in July 2002 with $103.9 billion in assets, the largest U.S. bankruptcy at the time, after an accounting fraud that overstated assets by over $11 billion. Qwest avoided bankruptcy but settled a separate multi-billion-dollar accounting-fraud case with the SEC over the same period. Cisco’s stock fell from a March 2000 peak of $80.06 to a low of $10.32 in October 2002, an 87 percent drop.

Notice the order. Cisco printed money before the carriers proved demand; the carriers had to validate their use case retroactively, and several could not. The shovel-sellers booked real revenue from physical equipment shipments while the buyers were still pitching forecasts.

The 2026 setup looks structurally similar. GE Vernova is having its Cisco moment. Whether Microsoft, Alphabet, and Meta have their WorldCom moment, or build a real durable revenue line that justifies the take-or-pay, is the question Wednesday’s earnings will start to answer.

What to Watch Wednesday

Microsoft, Alphabet, Meta, and Amazon all report after the close on Wednesday April 29. Microsoft’s quarter is fiscal Q3 2026, ending March 31, while Alphabet, Meta, and Amazon report calendar Q1 2026. Three things matter on the calls:

1. Capex run-rate vs. trajectory. Microsoft’s first-half fiscal 2026 capex tracked at roughly $72 billion, with the second-half run-rate the central question for the call. A pace that holds suggests near-term cancellations only trimmed the speculative buffer; a pace that lags suggests the cuts are biting into actual deployment.
2. Cloud backlog conversion. Google Cloud’s backlog grew by 55 percent quarter-over-quarter to $240 billion as of Q4 2025, framed by Alphabet leadership as driven by AI-product demand. Investors will want to know how much of that is signed, how renewable it is, and how fast it converts to revenue. Slow conversion means the AI revenue is theoretical; fast conversion validates the capex.
3. Take-or-pay disclosure. None of the four companies has cleanly disclosed how much of their announced capex is now contractually non-cancelable. Signatories of the Ratepayer Protection Pledge committed to pay for power and related infrastructure whether or not they use the electricity. Q&A is likely to push for specifics on how those obligations show up on the balance sheet.

The Bottom Line

The cleanest read of the data: hyperscalers appear to have concluded that meaningful AI revenue is years away rather than imminent. The response has been to pull back on near-term capacity, the part of the queue that would have been built speculatively, while locking in long-dated capacity that cannot easily be replaced once a slot is forfeited.

That is a defensible strategy. It is also a non-cancelable one. The Ratepayer Protection Pledge, the four-year turbine waits, and the take-or-pay utility contracts mean the bear case is no longer “Big Tech pulls back.” The bear case is “Big Tech cannot pull back.”

GE Vernova is not waiting for that question to resolve. Its $163 billion backlog is already a concrete piece of AI revenue showing up in industrial earnings. The companies whose names are in the headlines (Microsoft, Alphabet, Meta, and Amazon) get to find out after Wednesday’s close whether their version of the trade also pays.

For the broader picture on grid bottlenecks and the physical limits of the AI buildout, see The Transformer Crisis: Why the Grid Can’t Handle AI and The Shadow Utility: Meta’s 6.6GW Nuclear Maneuver.