Real companies · filing evidence · no ratings · data pulled 2026-10-30
What the gauge reads in real filings
The demo shows a rating on fictional companies. These pages show the other half: the evidence the gauge works from, taken from real companies’ SEC filings, quoted in the filing’s own words and linked to the source.
There is no rating, score or verdict here. Each page shows what a company disclosed: its quarterly numbers, how it talks about AI in earnings releases, and, where it applies, changes to how long it depreciates servers, customers that pay for a large share of revenue or that it helps finance, and how its annual report describes the AI in its products.
Earnings-release language
Nvidia NVDA
Its last eight earnings releases average about 20 mentions of AI per 1,000 words, and about 1 in 15 of those AI sentences carries a figure. Two customers accounted for 22% and 14% of revenue.
See the evidence →
Server useful life
Microsoft MSFT
Extended the estimated life of its servers and network equipment from 4 to 6 years in fiscal 2023, which lowers depreciation and raises reported profit. The 10-K gave two reasons: investments in software that made the equipment more efficient to operate, and advances in technology. Its fiscal 2026 10-K states a life of two to six years for servers and network equipment.
See the evidence →
A useful-life reversal
Amazon AMZN
Extended server lives from 4 to 5 years, then to 6, then shortened a subset of the fleet back to 5 years from January 2025, after a useful-life study. The only shortening among the infrastructure companies read.
See the evidence →
Customer financing
AMD AMD
Issued warrants for AMD shares to OpenAI and Meta, tied to their purchases of AMD chips. No single customer reported at 10% of revenue.
See the evidence →
Where AI sits in the product
CrowdStrike CRWD
Of the 55 AI sentences in the business section of its latest 10-K, 15 describe the AI carrying out a step itself (detecting and blocking threats) and 11 describe it helping a person (speeding up analysts’ investigations); the other 29 name AI without saying what it does. Labels by Claude Fable 5.1, with a reason for each.
See the evidence →
Where AI sits in the product
Adobe ADBE
Of the 66 AI sentences in the business section of its latest 10-K, 3 describe the AI carrying out a step itself (agents orchestrating customer journeys and multi-step workflows) and 20 describe it helping a person (generating, suggesting, assisting); the other 43 name AI without saying what it does. Labels by Claude Fable 5.1, with a reason for each.
See the evidence →
How to read the evidence
The three filings behind the evidence
The 10-K is the annual report. It holds the audited financial statements, the notes that explain the accounting choices behind them (including how long equipment is depreciated and which customers matter most), the risk factors, and Item 1, the company’s own description of its business and products. The 10-Q is the quarterly report: reviewed rather than audited financial statements for each of the first three quarters, with shorter notes.
The earnings release is filed as Exhibit 99.1 to a Form 8-K on the day a company reports its results. It is not audited and is written to present the quarter in the company’s preferred terms, which is why its language is worth reading next to the financial statements that follow it.
All three are free on the SEC’s EDGAR system. The financial figures are taken from the XBRL data companies attach to their 10-Q and 10-K, which tags each number with a standard accounting concept so that companies can be read the same way.
Reading the quarterly results
Every figure in the quarterly table comes from the company’s own XBRL filings with the SEC: the structured data attached to each 10-Q and 10-K. Companies report the fourth quarter only as part of the full year, so the fourth-quarter figures are the annual 10-K total less the first nine months reported in the third-quarter 10-Q. Each quarter links to the filing it came from.
For software and services companies the table shows revenue and gross margin: revenue less the direct cost of delivering the product, as a share of revenue. AI features that run on rented computing show up in cost of revenue, so a product that leans harder on AI with no change in price tends to show in the gross margin before it shows anywhere else.
For companies that build or sell the hardware and capacity AI runs on, the table shows revenue, operating cash flow, capital expenditure and free cash flow (operating cash flow less capital expenditure and finance-lease principal payments). Data-centre spending appears as capital expenditure long before it appears as revenue, so cash flow during a build-out often looks very different from reported profit.
Reading the earnings-release language
Each quarter, US companies file their earnings press release with the SEC as an exhibit to a Form 8-K. The release is the company’s own account of the quarter, written for investors and the press. These pages count how often it mentions artificial intelligence, after removing the financial tables, the forward-looking-statements boilerplate, the “About the company” paragraph and the company’s own name.
Two counts are shown for each of the last eight releases. The first is AI mentions per 1,000 words of narrative text, which makes long and short releases comparable. The second is how many of the sentences that mention AI also carry a figure: a dollar amount, a percentage, a count of customers. “AI demand drove record data-centre revenue of $X” and “we continue to lead in AI” are both AI sentences; only the first gives a reader something to check against the financial statements.
The counts use generic AI vocabulary only, not product names, so companies that brand their AI differently are counted the same way. Example sentences from the latest release are quoted with and without a figure, so the counting can be checked by reading them.
Reading the server useful-life evidence
When a company buys servers and network equipment, it spreads the cost over the years it expects to use them. That estimate, the useful life, is a judgment the company makes and discloses in the notes to its financial statements. Lengthening it spreads the same cost over more years, which lowers depreciation expense and raises reported operating income without any change in cash spent.
Since 2020 most of the largest cloud operators have lengthened their server lives, typically from three or four years to five or six, citing software and operating improvements that kept equipment in service longer. AI accelerators are a newer and faster-moving class of hardware, which is why changes to these estimates are watched closely.
These pages read each 10-K and 10-Q filed since 2023 for statements about the useful life of servers and network equipment: the change, the years before and after, when it took effect, whether it applies to the whole fleet or part of it, and the effect on income where the company discloses one. Each reading quotes the sentence it came from and links to the filing.
Reading the customer evidence
US accounting rules require a company to disclose when a single customer accounts for 10% or more of its revenue. For companies that sell AI hardware and capacity, those disclosures show how much of the business depends on a handful of very large buyers, often the same cloud operators and AI labs.
Some suppliers also help finance their customers: an investment in a customer, revenue from a related party, or warrants for the supplier’s own shares that vest as the customer buys more. These arrangements are disclosed in the notes and in 8-K filings. They are not unusual on their own, but they mean part of the revenue is connected to money or equity flowing the other way.
These pages read the latest 10-K and the 10-Qs filed after it for both kinds of statement, quote each one and link to the filing. Sentences in the risk factors (“may”, “could”, “if”) are skipped, so only arrangements the filing states as fact are shown.
Reading the workflow evidence
A software company’s annual report opens with Item 1, a description of the business and its products. How Item 1 describes AI says something about where the AI sits in the customer’s work: whether the software carries out a step itself (detects and blocks the threat, routes the claim, closes the books) or helps a person do it faster (suggests, drafts, summarises).
These pages take every sentence in Item 1 of the latest 10-K that mentions AI and sort it into one of three groups: the AI carries out a step, the AI helps a person, or the sentence names AI without saying what it does. Each label comes with a short reason, and up to three sentences from each group are quoted so the sorting can be checked against the filing’s own words.
The sorting can be wrong, and one company’s count depends on how much of its business section it gives to AI. The quoted sentences are the evidence; the counts are a summary of them.
Questions about the evidence
Why are there no ratings on these pages?
The AI Signal Gauge is a prototype, and its method is still being tested against history. Until that work is finished, ratings for real companies are kept to a small review group. These pages show the evidence the gauge reads, which stands on its own: each item is a company’s own disclosure, quoted and linked.
Why these six companies?
Each one shows a single kind of evidence clearly: Nvidia (earnings-release language), Microsoft (server useful life), Amazon (a useful-life reversal), AMD (customer financing), CrowdStrike (where ai sits in the product), Adobe (where ai sits in the product). They were chosen for what their filings illustrate, not as a view on the stocks.
Where does the data come from?
Only from SEC filings on EDGAR: the structured XBRL data in 10-Q and 10-K reports, the text of those reports, and the earnings releases companies file as Form 8-K exhibits. No analyst estimates, surveys or company-supplied data are used.
How are the quotes found?
Rule-based readers pull out the sentences for each kind of evidence: useful-life statements, customer concentration and financing, AI mentions in earnings releases. The sentences in a 10-K business section are sorted by a language model, which gives a reason for each label. Any of these can misread a sentence, which is why every quote is shown in full and links to its filing.
What is XBRL, and why use it?
XBRL is the structured data format the SEC requires alongside the financial statements in every 10-Q and 10-K. Each number is tagged with a standard accounting concept, such as revenue, cost of revenue or payments for property and equipment, so the same figure can be read the same way across companies and quarters without retyping it from a PDF. Where companies use different tags for the same line, the readers pool the equivalent tags and check that they agree.
How often is the evidence updated?
The data is pulled from EDGAR in batches, most recently on 2026-10-30. New quarterly filings appear after the next pull, and the notes on each page are checked against the data when it is refreshed.
Prototype · not investment research. Companies were chosen because each shows one kind of evidence clearly, not as a view on the stock. Ratings for real companies exist only in a review sandbox while the method is calibrated.