How to Read This Site
In plain language: what this site tracks and what each score means.
What is the “AI bubble”?
A “bubble” is when the price of and spending on some asset climbs far faster than the money it currently earns — propped up mostly by expectations and herd behavior. For AI, that means the money poured into chips, datacenters, and model companies (capital expenditure) is growing astronomically, while the revenue AI actually generates lags far behind. The one core question: will all that spending ever pay off? If yes, it's a historic build-out; if no, it's a bubble.
This site doesn't tell you the answer. It continuously collects news and data from the US and China, lays the “bubble” and “not-a-bubble” evidence side by side, and lets you judge.
Bubble signal (−2 to +2)
This is a direction score: does this news make the bubble worse, or cool it down?
So: negative = the bubble is cooling, positive = the bubble is inflating, 0 = neutral (related but no clear direction).
Relevance (0 to 100)
This is a strength score — a different axis from the direction score above: how much does this news bear on the question “is AI a bubble?”
Scoring is not a free-form number. The rubric has five bands: pick the band first, then a value inside it — so two items both scoring 70 are comparable:
Speaks directly to the question of whether AI is a bubble.
e.g. Capex, valuation or funding totals with real figures; AI revenue measured against AI spending; a named institution or investor explicitly arguing about an AI bubble.
A concrete flow of money that adds to the numerator or denominator of the bubble question.
e.g. A specific AI funding round or acquisition; a large chip order; a datacenter started or cancelled; a sharp move in AI-linked share prices.
Shifts expectations about AI investment without itself carrying a figure.
e.g. Policy or export controls that redirect AI investment; compute supply and demand; qualitative AI commentary in an earnings call.
AI-industry news, but several steps removed from the investment cycle.
e.g. A model release or research result; an executive hire or departure; a conference or partnership announcement.
The word “AI” merely happens to appear.
e.g. Product reviews and how-tos; articles where AI is incidental background.
Relevance measures only how much an item informs the bubble question — not how dramatic the headline is, nor which way it points. A calm article carrying real capex figures outranks a heated opinion piece carrying none.
The current display threshold is 30: items below it don't reach the feed. That's why an item with a 0 direction but relevance ≥ 30 still shows up — it genuinely bears on the bubble, it just doesn't lean either way right now.
Where the news comes from & how it's filtered
- Fetch: pull articles on a schedule from US and China media feeds and keyword news search.
- Dedupe: when the same story is syndicated across outlets, only one copy is kept.
- AI reading: a model writes a bilingual summary, a bubble explanation, category and region for each item, and assigns relevance and bubble-signal scores using the rubric above.
- Filter: items scoring below 30 don't reach the feed.
A caveat: the summaries, the bubble explanations and both scores are model-generated, and can be wrong or biased. Every signal links to the original article — for anything that matters, read the source.
Categories
Each signal is filed under one theme: valuation, funding, capex, chips & compute, policy, models, markets, or other. The US-vs-China page lays both countries' activity side by side along these dimensions.
How to read the dashboard numbers
Under every chart is the organisation the data came from; the link goes to the original, so you can check it yourself. We don't compute these numbers — the organisations publish them.
Three things affect how you should use them:
- Definitions differ, so don't mix them. Institutions genuinely disagree on what counts as an AI company or as capital expenditure, which is why each indicator states whose definition it follows. Don't add or compare figures across sources unless they cite the same one. The US and China private-investment series are directly comparable because they come from the same table of the same report.
- The numbers lag. Annual figures update once a year and quarterly ones wait on filings or surveys, so the latest point is usually months behind reality. Trends are more reliable than any single point.
- Sparse charts are deliberate. If we can't point to a public source, we don't plot it, even if that leaves gaps. Figures appearing inside news articles don't automatically become points either — those are often misquoted or stripped of context.
Copyright & attribution
For news we keep only the headline, a short summary, and a link to the original — never the full text. Indicator data cites its source organisation and link; the original source is authoritative.