How Froth Watch works

What it is

Companies are spending hundreds of billions of dollars a year on AI chips, data centres and power. Froth Watch checks each week whether that spending is being paid back, or whether the strain is building up somewhere along the supply chain.

Each issue covers the whole chain, from the AI labs and cloud companies down to data centre builders, power, cooling, networking, chips, foundries, materials and shipping. It rates each part, sets the odds on seven questions, and lists the week's main developments with their sources. Good news and bad news get the same scrutiny.

The seven questions

Every issue sets a probability for each of these, over several time horizons.

• H1Will the big AI spenders cut planned capex by 15% or more from its peak?
• H2Will an AI-infrastructure company default, restructure its debt or be forced to sell assets, at $5B or more?
• H3Will Nvidia's data-center revenue fall two quarters in a row?
• H4Will OpenAI, Anthropic or xAI raise money at a flat or lower valuation, or cut compute commitments?
• H5Will a shortage of power, packaging, memory or materials delay 5 GW or more of AI capacity by two quarters?
• H6Will AI revenue grow faster than AI capex over the next year?
• H7Will AI capability keep improving at its recent pace?

How an issue is made

01 Collect

Collectors gather company filings, exchange releases, market prices, credit spreads, supplier revenue, GPU rental and token price indices, and news. They run every day, and the data is stored with the address it came from.

02 Analyse

Analyst models check the week's events against the original sources and rate twelve layers of the supply chain on five measures: financing, demand, physical supply, valuation and technology. Two more models argue the bull case and the bear case from the same evidence.

03 Check

Code checks every figure in the draft against the stored source it cites, a separate model checks the facts, and a reviewer model from another AI lab reads the whole report. No figure comes from a model's memory.

How to read the probabilities

Each probability is the median of five separate model runs. Three runs answer the question as written and two answer the opposite statement, which is converted back. The runs never see anyone's opinion on whether AI is a bubble. The words come from the US intelligence community's ICD 203 standard:

Short forecasts that resolve from collected data are scored when their dates arrive, and the running score is published, so you can judge how well calibrated the system is.

How sure each item is

confirmed means an official source such as a filing or exchange release. corroborated means two or more independent reports. reported means one credible report.

What it is not

• It is not investment advice, and it does not tell you what to buy or sell.
• It does not predict the date of a crash. It tracks where the strain is and how it changes.
• It is written with AI models, under rules that tie every claim to a stored source. Mistakes are still possible. If you spot one, tell us and the next issue will correct it.

Who makes it

Froth Watch is published by sender name (and ABN if sending as a business). Write to contact address with corrections, questions or ideas for a deep dive.