- InfoWorld columnist David Linthicum reports that five US providers plan $660 billion to $690 billion of capital spending in 2026, nearly double the year before.
- He argues that fear-driven buying feeds the shortages buyers face now, and that overbuilt supply could lower prices in the next few years.
- His advice is to commit only to what current demand justifies. Ask your team which commitments reflect need and which reflect fear.
The providers carry the build risk. Buyers choose their commitments.
Enterprises are rushing to secure cloud capacity while providers commit hundreds of billions to capacity that will not be fully used for three to four years. David Linthicum, who writes the Cloud Insider column for InfoWorld, argues that this leaves buyers with a choice the providers lack. The providers are betting hundreds of billions that demand keeps climbing. Buyers do not have to make the same bet.
That is his opinion, and it deserves a careful read. The facts underneath it are worth setting out first.
What the column reports
Linthicum writes that Microsoft, Alphabet, Amazon, Meta and Oracle have committed between $660 billion and $690 billion in capital spending for 2026. He describes that total as nearly double the year before.
He breaks the total down by company. Amazon's plan is the largest, at about $200 billion. Alphabet's range is $175 billion to $185 billion, set against $91 billion in 2025. Meta's range is $115 billion to $135 billion. Microsoft is on pace for $120 billion or more. Oracle aims for roughly $50 billion.
Two figures in the column cover different groups. The 62% rise below is for four providers: about $630 billion planned, against the $388 billion they spent in 2025. "Nearly double" is the columnist's description of the five-provider total.
The column does not name the filings or reports behind these figures. Treat them as the columnist's account, not as audited numbers.
What the money actually buys
Linthicum says roughly two-thirds goes to compute. That means graphics processors and custom accelerators, which are chips built for AI workloads. He names Google's Tensor Processing Units and Amazon's Trainium.
The remaining share pays for buildings and what keeps them running: electricity, cooling and network links. He says power takes a larger slice each quarter. Some sites use liquid cooling and high-density floors. Where the grid cannot keep up, some use on-site gas-fired generation.
The column says the build-out runs three to four years from commitment to full utilization. It does not say why. One likely reason, in WebPulse's reading, is the power constraint he cites: a data center is of little use until it has the electricity to run. He also notes that capacity committed now is for delivery in 2027.
Power also limits supply today. Microsoft, he says, holds Azure orders worth $80 billion that it has not been able to deliver. He attributes most of the gap to limits on available power.
Why buyers feel squeezed
For the near term, Linthicum says a company buying processors, RAM or cloud capacity should expect to pay more, face allocation limits and wait longer. In his view, companies are buying early because they believe those who hold back will fall behind in the AI boom. In his words, "Nobody wants to explain to their board in 2028 why they spent 2026 waiting in line."
That fear is human and reasonable. The IT leader who waits and then misses a window pays a visible price. Linthicum's point is that the fear itself feeds the shortage. He adds that some demand is real and some is fear-driven, but "it all spends the same" for vendors.
He does not claim demand is weak. He calls demand over the next two years "already firm" and says every hyperscaler reports being limited by supply, not by demand. The caution is about what comes after.
The skeptic's case
Linthicum says he has watched earlier overbuilds, including fiber in the late 1990s and the first cloud buildout. He lists three ways the current spending could disappoint: models may need less compute per task, AI adoption may be slower, and price competition may squeeze returns.
He also raises the possibility of circular demand. Enterprises and vendors may be buying capacity from each other in commitments that never become steady workloads.
From this he forecasts that many data centers now being committed will be used at lower rates than assumed for 2029. He says this is not a prediction of collapse. It is a forecast, and the column offers no data to test it. WebPulse reports it as his view, not as a finding.
For customers, he sees an upside. He expects overbuilt supply to lower prices in the next few years, which would make cloud rates soften and data center space cheaper.
What leaders should ask their teams
The useful part of the column is the discipline it asks for. Linthicum advises buyers to lock in what they genuinely need and avoid long commitments that current demand cannot justify. Five questions follow from that.
First, which workloads need capacity today, and which are being bought because competitors are buying? Second, how long are our commitments, and what are the minimum-spend and exit terms? Third, if prices fall, can the contract be repriced or shortened? Fourth, can the workload move between providers without a rewrite? Fifth, who inside the company owns the risk that the capacity goes unused?
The lesson here is about clocks. A shortage runs on the seller's clock. A contract should run on yours.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: InfoWorld.
