Is the AI build-out running ahead of itself?
Artificial intelligence runs on data centers packed with specialized chips called GPUs. Building them has become one of the largest investment waves in modern history. The five biggest tech companies alone spent $566 billion on buildings and equipment in the past year, much of it for AI.
Railways in the 1800s and fiber-optic networks in the 1990s changed the world. They also bankrupted many of the companies that built them: too much was built, with borrowed money, on equipment that aged fast. AI could change the world and still follow the same path.
This page tracks the early warning signs. Every figure links to its public source, and nothing is estimated to fill gaps.
Research, not investment adviceUpdated 22 Oct 2026
Borrowing cost, riskiest companies13.27%Oct 20, 2026 · +479 bpsworseningHigher borrowing costs increase refinancing pressure.
+56%1 year
Borrowing cost, all junk-rated3.33%Oct 20, 2026 · +34 bpsworseningHigher borrowing costs increase refinancing pressure.
+11%1 year
Big tech capex, last 12 months$566Bto Jun 30, 2026mixedSpending is accelerating; cash generation still funds it.
+81%1 year
Nvidia H100 on AWS$6.88per GPU-hour, list price · Oct 19, 2026mixedCheaper hardware may reflect competition or newer chips.
0%1 year
Empty data-center space1.4%Main North American markets · June 2026reassuringLow vacancy indicates current supply is still tight.
−13%1 year
Borrowing costs
Most AI data centers are built at least partly with borrowed money. The key number is the "spread": the extra interest a company pays over the US government. When lenders get nervous, spreads widen, and the most indebted companies feel it first.
Lending to the riskiest companies costs 479 basis points more than a year ago
Extra interest over US Treasuries, percentage points
Read: Worsening: the weakest borrowers are paying sharply more while the broad junk market is near average.
Companies rated CCC, the lowest rung of junk debt, now pay 13.27 percentage points more interest than the US government. A year ago they paid 8.48 more. (A basis point is a hundredth of a percentage point.)
Evidence and context (3 notes)
Across all junk-rated companies, the gap is 3.33 points, up 34 basis points over the same year.
The weakest borrowers are being charged much more while the wider market barely moves. Stress usually shows up there first.
Since October 2023, the junk-bond gap has averaged 3.08 points, so today's level is close to that average.
The spending
Data centers and the chips inside them are paid for out of "capital expenditure", or capex. The big tech companies fund it mostly from their own cash. Newer specialist AI clouds, or neoclouds, fund it mostly with debt. Both are worth watching, for different reasons.
Big tech spent $566 billion on data centers and equipment in a year, up 81%
Microsoft, Alphabet, Amazon, Meta, Oracle · 12-month totals
Read: Mixed: spending is climbing rapidly, but the group still produces positive free cash flow.
Microsoft, Alphabet, Amazon, Meta and Oracle spent $566 billion on capital projects in the 12 months to Jun 30, 2026. A year earlier it was $313 billion.
Evidence and context (3 notes)
The companies say a growing share goes to AI data centers and chips. Their filings don't say how much.
What's left after that spending, their free cash flow, is $126 billion. That is 49% below its peak of $247 billion in March 2024.
These companies can afford it: their businesses still produce more cash than they spend. The question is how long that stays true if spending keeps climbing.
CoreWeave is spending 3.0× the cash its business brings in
CoreWeave · 12-month totals
Read: Worsening: investment exceeds operating cash flow, increasing dependence on lenders.
In the 12 months to Jun 30, 2026, CoreWeave spent $21 billion on capital projects. Its operations brought in $7 billion in cash.
Evidence and context (2 notes)
Borrowing and leases cover the gap. Its long-term debt was $25 billion on Mar 31, 2026. The Credit page lists each loan and bond.
Neoclouds like this depend on lenders staying willing. That is why borrowing costs and used-GPU prices matter so much here.
The price of computing
If too much capacity gets built, the price of renting a GPU, and of reselling a used one, should fall. Both matter: rental prices are what these data centers earn, and used-chip values are what lenders can recover if a borrower fails.
Renting an Nvidia H100 on AWS costs 44% less than in November 2024
Nvidia H100, list price per GPU per hour
Read: Mixed: price cuts can point to weaker economics, but newer hardware can produce the same pattern.
Amazon's list price is $6.88 per GPU per hour, compared with $12.29 in November 2024. The last change came in June 2025.
Evidence and context (2 notes)
Microsoft lists the same chip at $12.29 per GPU-hour.
Falling prices can mean more capacity than demand. But newer chips make older ones cheaper by design, so a price cut alone doesn't prove a glut.
Used data-center GPUs resell for 23% less than six months ago
Resale index of used data-center GPUs · early 2025 = 100
Read: Worsening: falling resale values reduce the collateral supporting GPU-backed borrowing.
An index of completed sales stood at 75 in August 2026, where 100 is the level of early 2025.
Evidence and context (1 notes)
Many GPU loans use the chips themselves as collateral. If resale values fall faster than loans are repaid, lenders are less protected.
The buildings
The final test is whether finished data centers fill up. Real-estate brokers report how much space is empty, how much is being built, and how much of it is already leased.
Data centers are nearly full: only 1.4% of space is empty
Main North American markets · each line has its own scale
Read: Reassuring now: vacancy remains low and most new capacity is already leased.
Different scales: compare direction, not distance.
In North America's main data-center markets, vacancy was 1.4% at the end of June 2026.
Evidence and context (3 notes)
Another 7,481 megawatts was under construction, up 43% from a year earlier.
80% of that new space is already leased before it is finished. Most of it is spoken for, which is reassuring.
Vacancy usually rises late in a downturn, so it is a slow signal.
Warning signs
One chart alone rarely tells you much. The pattern to watch is two things pulling apart: spending still rising while prices fall, or debt growing while what backs it loses value. Each pair below shows two lines, each on its own scale. They show what happened, not why.
GPU rental prices vs hyperscaler capex
Watch for: Rental prices falling while capex keeps rising.
Different scales: compare direction, not distance.
Borrowing costs vs CoreWeave's order book
Watch for: Borrowing costs rising while the order book keeps growing.
Different scales: compare direction, not distance.
Used-GPU prices vs CoreWeave's debt
Watch for: Used-GPU prices falling while CoreWeave's debt stays high.
Different scales: compare direction, not distance.
New building vs space already rented
Watch for: Building speeding up while less of it is rented in advance.
Different scales: compare direction, not distance.
Assumed server lifespan vs spare cash
Watch for: Servers assumed to last longer (which flatters profits) while spare cash shrinks.
Different scales: compare direction, not distance.