Big Tech AI Capex Inflated by 20% Due to Cost Inflation
Claire Weston
Google, Amazon, Microsoft and Meta plan over $700 billion in combined capex this year, but Morgan Stanley estimates per-gigawatt AI build costs have risen roughly 20% — meaning a sizable share of the bigger number is just paying for higher prices, not more compute.
$700 billion in capex — how much is just paying for inflation?
The four giants' combined 2026 capex plans top $700 billion, yet Morgan Stanley calculates the per-gigawatt cost of building AI compute has climbed about 20%.
This means → the headline number grew, but the actual compute purchased hasn't grown by nearly the same proportion. Part of the money is simply absorbed by higher prices.
Specifics: a standard Nvidia GPU configuration now costs roughly $35 billion per gigawatt, up from $29 billion; a newer-generation setup rose from about $41 billion to $49 billion.
The drivers go beyond chips — memory prices surged, and power equipment, construction materials, skilled labor and grid-connection capacity are all tight.
What's the real split — 70% or 80% genuine expansion?
Brad Gastwirth, head of research at Circular Technology, estimates 20–30% of the next wave of AI capex growth reflects cost inflation, with 70–80% representing real capacity additions.
In plain terms = for every extra $100 spent, roughly $20–30 just covers higher prices; the rest actually buys new equipment.
A separate data point sharpens the picture: one study found memory price increases alone explain about 45% of this year's hyperscaler capex growth.
This means → if memory prices retreat, capex growth could slow markedly — but actual compute deployment may not shrink in step. The number gets smaller; the hardware doesn't necessarily follow.
Which metrics separate real expansion from capex bloat?
Gastwirth's advice: don't stop at the capex total — check whether companion metrics are rising in lockstep.
The verification checklist: power capacity, GPU deployment volumes, memory procurement, networking equipment and new data-center campus count — only when these move up together can you call it genuine expansion.
In plain terms = looking at the bill alone tells you nothing; you need to see how much gear actually arrived.
Why might 2027 guidance be the bigger variable?
Cantor Fitzgerald analysts expect 2026 capex plans to hold roughly steady this earnings season, but 2027 forecasts to be revised sharply higher.
The numbers: Google's 2027 capex estimate sits at $283 billion, Amazon at $271 billion, Meta at $200 billion.
This reflects a self-reinforcing loop: no company wants to blink first in the AI race → orders keep flowing → supply tightens → prices rise further → spending forecasts ratchet up → creating even more demand.
This means → the central question for this earnings season is: do bigger numbers represent more compute, or just a more expensive bill?
Content is for reference only, not financial advice.