Mining the AI Capex Trade from Earnings Call Transcripts

The AI capex cycle is the largest capital deployment story in public markets right now. Per public reporting this summer, hyperscalers are expected to nearly double their capex in 2026, aggregate AI infrastructure spend is heading toward $1 trillion in 2027, and a growing share of it is debt-financed. That is the macro headline. The tradeable question is more granular: which of the 12,000+ listed companies in the corpus are spending, which are collecting, and when does the talk start to fade?

Earnings calls are the highest-frequency primary source for that question. Every 90 days, on the record, management teams tell you whether AI infrastructure is showing up in their orders, their backlog, or their budget — usually two or three quarters before it shows up in segment reporting. This post walks through ten quarters of mention data pulled from the EarningsAPI corpus, then gives you the exact queries to rebuild and extend the analysis yourself.

Ten quarters of "AI capex"

Start with one full-text query:

curl 'https://earningsapi.io/api/v1/search?q="AI capex"&type=transcripts&date_from=2026-04-01' \
  -H "X-API-Key: $KEY"

The q parameter supports exact phrases in quotes, AND/OR operators, and - negation. Run it per calendar quarter, count distinct companies, and you get this series:

One methodology note before reading anything into this: these counts are calls matching a full-text search. That makes them an upper bound — the phrase can appear in an analyst's throwaway framing or an incidental co-occurrence, not only in substantive discussion. Treat the level as noisy and the trend as the signal.

The shape is clean. All of 2024 sits in a flat band of 765–826 companies per quarter — a baseline hum, drifting up single digits while the phrase was still analyst jargon. The first break comes in 2025 Q1 at 1,097 (+43% YoY), and the growth rate has accelerated in every year-over-year comparison since: +59% in Q4 2025, +60% in Q1 2026, and +85% in Q2 2026, the fastest print in the series. In absolute terms the count more than doubled from the 2024 baseline to Q1 2026's 1,757 companies.

Two denominator caveats. Q2 quarters carry less corpus volume than the rest of the year — 2025 Q2 had 8,270 calls against a typical ~10,000 — which is why that quarter's flat YoY print overstates the lull. And Q2 2026's corpus was back up to 10,053 calls, so its 1,447 stands on a full denominator: roughly one in seven earnings calls last quarter touched AI capex.

Industrials lead — and no GICS screen finds them

Here is the part that surprises most people. The Q2 2026 sector split of companies matching the query:

• Industrials: 461 companies
• Information Technology: 378
• Consumer Discretionary: 281
• Communication Services: 183
• Health Care: 156
• Financials: 155

Industrials lead. The companies building, powering and wiring data centers talk about AI capex more than the tech sector does. Caterpillar's August 4 call is the archetype:

"In Power and Energy, sales to users grew a robust 33%. Power generation grew 72%, driven by very strong demand for large gen sets and turbines used in data center applications."

Power generation up 72% at a heavy-equipment maker — driven by data centers. That sentence never appears in a semiconductor screen.

This is the case for doing universe discovery from text rather than classification codes. A GICS filter for "AI exposure" hands you semis and software — a universe everyone already owns at consensus multiples. The full-text query surfaces the Industrials tail — genset and turbine makers, electrical equipment, engineering and construction, cooling — companies where AI infrastructure demand hits the P&L before any index provider reclassifies them. One query returns the whole universe, and the sector= filter on the same endpoint slices it without a second data source.

Spenders vs. collectors

With the universe in hand, split it. Every AI capex dollar sits on two sides of the same invoice: someone deploys the capital, someone collects it. Three search refinements do most of the classification work:

# Collectors: capex talk co-occurring with order books
curl 'https://earningsapi.io/api/v1/search?q="AI capex" AND backlog&type=transcripts&date_from=2026-04-01' \
  -H "X-API-Key: $KEY"

# Debt-funded spenders: who is stretching the balance sheet
curl 'https://earningsapi.io/api/v1/search?q="AI capex" AND financing&type=transcripts&date_from=2026-04-01' \
  -H "X-API-Key: $KEY"

# Supply-side pricing power
curl 'https://earningsapi.io/api/v1/search?q="capacity constrained"&type=transcripts&date_from=2026-04-01' \
  -H "X-API-Key: $KEY"

On the spender side, the language itself has changed. On AMD's August 11 call, management framed deployment as "a matter of us executing and making sure we have land and power and shell and capital commitments," including "the first full gigawatt with Anthropic in 2027." Gigawatts, land, shells — utility and real-estate vocabulary coming from a chip company. When spenders start talking like developers, the "AI capex" AND financing cut earns its own watchlist, especially with the buildout increasingly debt-financed.

The collector side has a subtler third archetype: the beneficiary whose own capex stays tiny. Western Digital's August 5 call:

"Operating cash flow was $1.4 billion and CapEx was $108 million... free cash flow generation of $1.3 billion for the quarter and a strong free cash flow margin of 34%."

EPS of $3.56, up 109% year over year, per the call — selling into the buildout at a 34% free cash flow margin while spending $108 million on itself. Spenders burn cash to build capacity; collectors convert the burn into backlog; and a third group converts it into free cash flow without joining the arms race. The AND backlog and AND financing refinements separate the first two; screening the match set for capex-light income statements finds the third.

Ask who raised it first

Mentions have an author, and the author matters. When management volunteers AI capex in prepared remarks, that is positioning. When analysts push on it in Q&A before management brings it up, that is the sell side telling you where the model sensitivity lives. The speaker index makes the split one parameter:

curl 'https://earningsapi.io/api/v1/search?q="AI capex"&type=speakers&speaker_type=analyst&date_from=2026-04-01' \
  -H "X-API-Key: $KEY"

Compare the analyst-only match set against the full set per company. Names where the phrase lives almost entirely in analyst questions are ones where the market is demanding an answer management has not scripted — a useful pre-positioning screen ahead of the next call. The inverse cut works too: companies where management leads with AI capex quarter after quarter while analysts never ask are worth a skeptical second read, because unsolicited theme attachment is free and analyst attention is not. It is the same speaker-attribution trick behind several of the five text signals quants pull from earnings calls.

Put it on a cron

The table above is not a research artifact; it is the output of a scheduled job. Run the query weekly during earnings season, append one row per period to a CSV, and chart it:

#!/bin/sh
# weekly during earnings season; one row per run
curl -s 'https://earningsapi.io/api/v1/search?q="AI capex"&type=transcripts&date_from=2026-07-01&date_to=2026-09-30' \
  -H "X-API-Key: $KEY" \
  # count distinct companies in the response, append date + count
  >> ai_capex_by_quarter.csv

The reason to automate rather than revisit: mention counts flatten before headlines do. Nobody publishes a story when a growth rate merely stops accelerating — but that is exactly the print you want to catch, and the 2024 plateau in the table shows what "quiet" looks like in this series. Keep the refinement queries on the same schedule so the spender, collector, and financing sub-series accumulate alongside the headline count; the divergences between them are usually more interesting than the total. The same cron pattern extends to any thematic phrase — on-shoring, GLP-1s, tariffs — and the mechanics are covered in our practical guide to analyzing earnings calls automatically.

Segments, not documents

Keyword search gets you universe and trend. Classification beyond keywords — "is this segment about AI infrastructure demand, supply, or financing risk?" — needs embeddings, and the corpus structure does the chunking for you. EarningsAPI exposes 11.9 million speaker-tagged segments: one speaker, one turn, with role metadata attached. They drop straight into an embedding pipeline as pre-chunked RAG units, no splitter heuristics required, and the MCP server serves the same segments directly to agents. The full setup is in our guide to feeding earnings call transcripts into a RAG pipeline via MCP.

That layer is where the two-sided classification stops being three keyword refinements and becomes a proper classifier: embed the segments from your match set, label a few hundred, and let the model separate a genset maker's demand commentary from a hyperscaler's financing disclosure — including the calls that discuss the buildout without ever saying "AI capex."

What to watch from here

As of August 17, 2026, the series still points one way: Q2's +85% YoY is the fastest growth in ten quarters, and the sector split says the physical layer — the Industrials actually pouring the concrete and shipping the turbines — is now the loudest part of the trade. Two things are worth watching in the data rather than the news flow. First, the first flat YoY print: this series went vertical five quarters before capex hit the front pages, and it will likely go quiet early too. Second, composition: if Industrials mentions roll over while tech's hold, the buildout is decelerating at the site level first.

Everything in this post runs on full-text search over 253,000+ transcripts from 12,000+ companies across 175 exchanges — browse the raw material in the transcript library. The whole pipeline is a handful of curl calls and a cron entry.

Ready to run these queries yourself? Get an API key at earningsapi.io — full-text search, speaker-tagged segments, and MCP access on quarterly plans from $145.