An interactive reading guide

Running Dry,
in numbers

Ed Zitron argues the AI boom ends when the money runs out. His Nov 3, 2026 piece says the money is now getting harder to raise. This page turns his figures into charts. Click any bar, card or node to drill down.

Source: “Running Dry”, Where’s Your Ed At, Nov 3, 2026. It is an opinion piece. Figures are as the article reports them. Tags mark which numbers I worked out myself.

☝ Click any bar, card or node to open its detail panel
01 ¡ OpenAI

The round was smaller and cheaper than planned

OpenAI hoped to raise $30B at about $1.4T. It closed $11B at $1.25T. The article says that cash covers less than one quarter of its costs.

OpenAI’s fundraise, Sept–Oct 2026

Switch the view. Click a row to see where the number comes from.

$11B vs. one quarter of costs

Dollars, billions. Click a bar.

Cash coming in, 2026

Known inflows named in the article.

02 ¡ Valuations

Everyone’s price target slid

Anthropic’s expected IPO price fell by hundreds of billions in under a month. OpenAI needs a much better exit than Anthropic to give its newest investors a gain.

Valuation, $ trillions

Solid bars are single figures. Wider spans show a reported range.

Click a row for the source and what it implies.
03 · Anthropic, Q2 → Q3 2026

Costs grew faster than revenue

Revenue rose about 47%. Compute costs rose about 77%. So each compute dollar brought in less revenue: $2.08 → about $1.72.

Revenue vs. compute spend

Dollars, billions. Click any column.

Q3 compute: who gets the money

$9.9B total. Click a segment.

Discounts are growing

Revenue lost to price cuts and contract discounts, $ millions.

04 · Anthropic’s commitments

Close to a trillion in promises, and more in the footnotes

The five named items add up to $895.3B. Two footnoted ranges could add $275B–$325B more. The article says the filing’s “per year” table hides how soon the money is due.

What Anthropic owes or may owe, $B

Click a segment or a legend item. Hatched segments are estimates with no fixed dollar amount.

When is it due?

Dollars, billions. Click a bar.

Cash on hand vs. the next 4 years

Dollars, billions. Click a bar.

05 ¡ Try it

A toy model of the gap

Move the sliders. See how big the funding gap gets if revenue does not pay for the commitments.

Illustrative only

This is my simple model, not the article’s and not Anthropic’s. It ignores timing, interest income, and cancellable deals.

Funding gap by end of 2030
Gap per year (4.25 years left)
Yearly interest on gap, once borrowed

With the default sliders the gap is about $50B a year. That is close to the article’s own estimate of “around $50 billion a year” in new debt.

06 ¡ Debt markets

The borrowing window is narrowing

Data center construction starts fell 64% in one month. Neocloud bonds are shut out. A junk rating would push Anthropic’s borrowing cost to 8%–11%.

US data center construction starts

$ billions per month (Westside, via article). Click a column.

Cost of money

Interest rates named in the article, %.

CoreWeave: debt now vs. debt needed

Dollars, billions.

07 ¡ Hyperscalers

The big clouds now lean on two labs

OpenAI was nearly 11% of Microsoft’s latest quarterly revenue. About 44% of Anthropic’s revenue passes through Google and Amazon.

Azure revenue, quarterly

$ billions. Click a column.

OpenAI’s weight at Microsoft

Share of revenue, %.

The loop the article describes

Click each step.

08 ¡ Subsidies

Cheap subscriptions may not last

The article says many AI plans lose money on heavy users. When funding tightens, prices or limits change. It calls this a rug pull.

Replit gross margin, 2025

Per The Information, Aug 2025. Hover or tap a point.

Where lab revenue comes from

Share of OpenAI + Anthropic revenue, %.

09 ¡ The argument

“It ends when the money runs out”

Two money streams hold the boom up. The article says both are weakening.

Stream 1: venture capital

Stream 2: debt

10 ¡ Read this too

Caveats and plain-language terms

This is one critic’s argument. Ed Zitron is a long-time skeptic of the AI industry. The article mixes SEC filing figures, press reports and his own opinion. I did not check the filings. Treat each number as “as reported in the article”.
Some numbers are loose. Q3 revenue is “a little under $17B”, so the page uses 17. The article’s headline says “$895bn to $1.07tn”. Adding both footnoted ranges gives more than $1.07T. Compare the “Named items” and “+ footnoted ranges” views.
S-1 / S-1/A
The sign-up form a company files with the SEC before an IPO. “/A” means amended.
Compute
Rented computer power (GPUs, TPUs) used to train and run AI models.
Non-cancellable
A contract the company must pay even if it stops using the service.
Neocloud
A newer cloud firm that rents AI chips, such as CoreWeave, Nebius or Lambda.
Hyperscaler
A giant cloud provider: Amazon, Google, Microsoft.
Junk rating
A credit rating below investment grade. Lenders charge more. Pension funds and insurers often cannot buy it.
Token-maxxing
Staff who burn as many AI tokens as they can, because the company encourages it.
Rug pull
A sudden price rise or limit cut after a service has won users with low prices.
Negative EBITDA
Operating costs are higher than revenue, before some accounting items.