In October 2026, a colocation buyer in Northern Virginia received a renewal quote 19% higher than the expiring rate, and the provider’s explanation was not about chips. It was about power, land, and a construction queue full of hyperscalers. US data center prices are up 19% on AI demand and limited capacity, according to Data Center Dynamics’ reporting of the market, and record construction is failing to ease the crunch, according to CBRE via Data Center Frontier. The buyer’s problem was not that GPUs got expensive. It was that four companies are buying the entire industrial base out from under the rest of the market.
The scale of that buying is the number that sets every rack price in 2026. The four largest hyperscalers — Amazon, Alphabet, Microsoft, and Meta — guide to roughly $730 billion of 2026 capital expenditure, up about 78% from roughly $410 billion in 2025. The split: Amazon about $220 billion, Alphabet $195–205 billion, Microsoft about $170 billion, and Meta $135–145 billion. All four reported third-quarter 2026 results between October 21 and 29, and all four held or narrowed their guidance rather than cutting it. Value Add VC’s AI spending tracker, updated November 4 with all four Q3 reports, carries the company-by-company figures and the 78% year-on-year increase.
The quarterly run-rate is still accelerating. Combined cash capital expenditure across the four reached about $202 billion in the third quarter of 2026, up from $165 billion in the second quarter. Alphabet spent $51.4 billion in the quarter, up 114% year on year. Amazon spent $75.9 billion net of incentives. Microsoft spent $45.2 billion including finance leases, and Meta $34.9 billion including lease principal. Made by Agents’ AI CapEx Tracker, sourced directly from SEC EDGAR filings, shows the same step-change in the quarterly series. Goldman Sachs projects the combined number nearing $1.14 trillion in 2027.
What they are buying explains why the money lands on colocation pricing rather than just on chip prices. GPUs and servers take 40% to 50% of the spend; data center construction takes 25% to 35%. Microsoft said roughly three-quarters of its quarterly capex went to short-lived assets, mainly CPUs and GPUs, and Azure revenue grew 48%. Alphabet said about 65% of its technical-infrastructure spend went to servers and networking equipment, while Google Cloud revenue more than doubled to $30.7 billion in the quarter, up 103%. Meta broke ground on a 2GW-plus facility in Louisiana. Useluminix’s Q3 2026 hyperscaler capex review records the category splits and the Louisiana build.
There is an accounting twist that makes the headline number understate the build. Microsoft lengthened the accounting life of its data-center buildings from 15 to 25 years, which moved more new site leases into operating leases — excluded from capex — and those additions jumped from $1.4 billion to $6.5 billion in a year. Contracted cloud backlog across Microsoft, Alphabet, and Amazon approaches $1.8 trillion. NVIDIA’s data center revenue grew from $15 billion in fiscal 2023 to $115 billion in fiscal 2025, largely on the back of this spending. The buildout is partly lease-financed, which means the $730 billion is a floor on the commitment, not a ceiling.
Every rack quote in 2026 is a small slice of a $730 billion construction program. Price it like one.
The honest counterargument
The build is debt- and lease-funded, and three of the four builders ran negative free cash flow in the third quarter: Alphabet at minus $9.1 billion, Meta at minus $2.1 billion, Amazon at minus $42.4 billion over the trailing twelve months. Microsoft was the only one that stayed cash-positive. If AI demand misses expectations, the assets strand fast — servers depreciate over five to six years, buildings over ten to forty — and the industry has stranded capacity before. Microsoft’s own trim, from about $175 billion to about $170 billion on October 28, shows the constraint is real, even if the cause was delayed third-party capacity deliveries rather than softening demand: the spending slipped because the buildings were not ready, and those deliveries now land in the first half of 2027.
The supply response is also slow by design. Power, not concrete, is the constraint — CBRE’s reporting shows record construction failing to ease the capacity crunch, because the queue for grid connections stretches years, not quarters. That slowness cuts both ways: it protects today’s pricing, and it guarantees that today’s pricing will look cheap if the demand forecasts are right, and expensive if they are wrong.
What this means for buyers
Three practical consequences follow. First, lock power early: the cheapest rack is the one with power attached, and the power queue is the real negotiation. Second, multi-year colocation terms now price the construction queue, so a renewal quote is a bet on hyperscaler capex as much as on local supply. Third, watch the delivery calendar — the delayed third-party capacity that trimmed Microsoft’s 2026 number lands in the first half of 2027, and the first sign of relief in the market will show up there, not in chip prices.
Data center pricing stopped being a chip story two years ago. It is a construction story now, and the construction story is a $730 billion program growing 78% a year with a $1.8 trillion backlog behind it. The buyers who negotiate on dollars per kilowatt without watching hyperscaler capex are negotiating blind. The ones who watch it know exactly why the renewal quote went up 19% — and whether to sign it or wait for the capacity that is already funded, already permitted, and already late.
The definitive source for data center pricing intelligence is still up for grabs. DataCenterPrices.com is available for acquisition by the company that will own this category — inquiries to info@primedomaincapital.com.