Shevaan Jayasinghe
← Home

Research · Draft · Updated 3 November 2026

Spending is growing.
The funding is stretched.

A dial for AI bubble risk, in the spirit of a fear and greed index. It reads four things: whether businesses keep paying, whether margins hold, whether data centres stay full, and whether the plans can be paid for without heavy debt.

Backed by fundamentals Healthy boom Stretching Bubble risk Bubble territory 0 25 45 60 80 100

57.5

Stretching

One number from four topics, weighted toward the weakest

72.8

Bubble risk

Funding, the weakest topic

0 is backed by fundamentals. 100 is bubble-like strain. Draft, not yet validated.

Updated 3 November 2026. 12 of 71 inputs refresh automatically; the rest change when companies and publishers report. 5 are overdue.

The four topics

01

Businesses keep paying

25.2
  • Who's paying8.3
  • Whether they stay and grow21.0
  • Whether it pays off71.6

What it says55.9% of US businesses on Ramp now pay for AI, and median spend per employee is up 214% in a year. But most firms report little payoff so far: 83.4% in one large survey see no productivity impact.

What could change thisNo public source tracks renewals across vendors, and the retention figures come from the fastest growers.

02

Margins hold as prices fall

46.7
  • AI labs31.0
  • Cloud segments35.0
  • Neoclouds86.7
  • Prices37.5
  • Earnings quality43.1

What it saysAWS and Google Cloud widened their margins to 38.7% and 36.2%. Microsoft Cloud slipped to 66% from 68%, Oracle's fell about 5 points, and the typical neocloud still loses money.

What could change thisMargins can look best near the top of a build-out. Enterprise token prices barely fell in early 2026, so the falling-price test hasn't really come yet.

03

Data centres stay busy

24.0
  • Running32.3
  • Power66.2
  • Supply coming21.4

What it saysVacancy is 1.4% on CBRE's count, and one-year H100 contracts cost about 60% more than a year ago. Capacity is scarce, and power is the tightest part: Goldman Sachs expects about 60% of next year's scheduled capacity to arrive on time.

What could change thisSold isn't the same as running. No one publishes how busy these chips are, rental-market readings need 30 days of data, and power prices also move with fuel costs.

04

Plans funded without heavy debt

72.8
  • Internal cash51.3
  • Debt build-up74.5
  • Hidden commitments91.0
  • Neoclouds95.9
  • Credit stress56.8
  • Circular funding67.1

What it saysThe five biggest builders put 95.5% of their operating cash flow into capex and have signed about $1.3 trillion of leases that haven't started. Seven neoclouds spend about 3 times their operating cash flow.

What could change thisThe big builders still earn a lot, and many neocloud contracts are take-or-pay with strong customers. Private credit and partners' debt aren't counted, so the real strain could be higher.

Who is carrying it

Past 12 months, from SEC filings, in US$ billions. Long-term debt compares the latest filing with a year earlier; Oracle shows total debt at its May year end. Leases signed but not started are future payments from each company's latest quarterly filing.

CompanyCapexOperating cash flowCapex, % of cash flowLong-term debt, thenLong-term debt, nowLeases signed, not started
Microsoft135.9195.969%35.4Sep 202531.1Sep 2026364.6
Alphabet159.8179.689%21.6Sep 2025125.1Sep 2026130.5
Amazon215.2168.3128%50.7Sep 2025173.1Sep 2026164.9
Meta104.4133.178%28.8Sep 202580.9Sep 2026355.7
Oracle75.746.9161%92.6May 2025129.5May 2026288.0

How it works

Read the direction before the decimal.

Each input gets a stress score. Inputs roll up into subcategories, subcategories into four topics, and topics into one number that leans toward the weakest topic, because one failing part can break a build-out.

I tested this against the 1999 to 2001 telecom bust. It flagged the builders early, but it also raised a false alarm on healthy hyperscalers in 2015, and only two of the four topics have usable history. Treat it as a measure of fragility, not a crash date.

Anthropic's figures carry real weight in some topics, and this page was built with Claude, an Anthropic model, so check those numbers yourself. This is research, not investment advice.

Sources include SEC filings, Ramp, the US Census Bureau, the EIA, CBRE, JLL, Silicon Data, PJM and company reports.

Live products

Open what is already shipped.

Open the live demo

Chappie is on the App Store. The AI Bubble Index is live on this site. This page is a record of earlier work.