August was the peak of the reporting quarter. 4,809 earnings calls from 4,508 companies landed in the corpus — up from 2,861 in July and 1,613 in June. If you run NLP over earnings calls, this was the month your pipeline earned its keep: an NVIDIA print that added roughly $440B of market cap in a session, a Jackson Hole speech that pushed September rate-cut odds above 90%, and a tariff-refund wave that showed up in 313 call transcripts.
This post is the recap — but framed the way a developer should read it: every event below maps to a query, and the whole summary table can be reproduced by a monthly cron job against one endpoint.
One methodology note, which applies to every count in this post: "X companies" means companies whose August calls matched a full-text search for that phrase. Keyword matching is an upper bound — it counts a CFO denying a trend the same as one confirming it. Treat the numbers as a screen, not a conclusion.
What 4,508 companies talked about
Each row below is one full-text search over August's calls, counting distinct companies with at least one match between 2026-08-01 and 2026-08-31.
| Search phrase | Companies matching (Aug 2026) |
|---|---|
| "data center power" | 1,180 |
| "rate cut" | 985 |
| "AI capex" | 852 |
| "power constraints" | 520 |
| "tariff refund" | 313 |
| "inference demand" | 167 |
| "agentic commerce" | 164 |
| "sovereign AI" | 123 |
| "humanoid robot" | 56 |
| "stablecoin" | 29 |
Before you read momentum into any of these: August had roughly 66% more companies reporting than July. A theme that merely rides the reporting calendar should grow about 66% month-over-month with no change in underlying attention. That's the baseline to beat.
Three themes beat it decisively — all of them more than doubled:
- "agentic commerce": 78 companies in July → 164 in August (+110%)
- "humanoid robot": 26 → 56 (+115%)
- "stablecoin": 14 → 29 (+107%)
Small bases, real acceleration. Agentic commerce in particular has crossed from payment-processor earnings calls into general retail and software vocabulary — we traced the earlier phase of that shift in what payment giants told investors about agentic commerce and stablecoins. The August counts say the language is no longer confined to that sector.
The middle of the table rewards a different workflow. At 167 matching companies ("inference demand") or 123 ("sovereign AI"), the result set is small enough to read exhaustively — an afternoon of skimming matched segments tells you whether the phrase means committed budgets or conference-season vocabulary. At 1,180 matches it isn't, and you fall back on filters: restrict to a speaker role, exclude crowded sectors with - negation, or narrow the date window to a single week.
Meanwhile the biggest rows in the table — "data center power" at 1,180 companies, "AI capex" at 852 — are no longer signals in the momentum sense. They are the ambient condition of the market. The signal now lives in who stops mentioning them.
NVIDIA, August 26: the print the whole table orbits
NVIDIA's Q2 FY2027 call on August 26 was the month's center of gravity. The numbers: revenue of $96.2B, up 106% year over year. Data center revenue of $89.0B, up 117%, against expectations around $85.7B. Gross margin of 75.0%. EPS of $2.22. The stock rose roughly 9% the next session — about $440B in added market cap.
Two derived facts from those numbers: data center is now roughly 92% of total revenue, and the beat against the ~$85.7B data center expectation came to more than $3B — the surprise alone was the size of a large company's quarter.
The line your pipeline should have flagged wasn't the beat. It was the guidance framing:
"We expect to grow revenue by approximately 70% in fiscal 2028. This is a supply-constrained outlook."
"Supply-constrained" is doing enormous work in that sentence. Management is saying the forecast is bounded by what can be built, not by what can be sold — and repeated the mechanism explicitly: "more compute drives more revenue as new GPU capacity comes online." For anyone modeling the AI capex chain, that phrase converts NVIDIA's guidance from a demand estimate into a capacity schedule.
The second flaggable line was about what's not in the outlook:
"there is no China data center compute revenue in our forward outlook."
A zero baked into guidance is an asymmetry: any China resolution is pure upside to the printed number. Both quotes are exactly the kind of forward-looking, management-voiced sentence that full-text search retrieves and headline numbers miss — the category of signal that makes transcript text, not the press release, the primary source for the AI capex trade.
The power constraint went industrial
1,180 companies said "data center power" in August. 520 said "power constraints." What makes this a supply chain story rather than a tech story is who the matches are.
Caterpillar's August 4 call reported power generation revenue up 72% on data center demand — a heavy-machinery company posting hyperscaler-shaped growth. One week later, AMD's August 11 call framed the binding constraint as "land and power" — not chips — and put a date on the buildout with its "first full gigawatt with Anthropic in 2027."
That's the pattern a cross-corpus search surfaces and single-ticker monitoring never will: the same constraint narrated from opposite ends of the supply chain, eleven days apart, in a machinery transcript and a semiconductor transcript. With 12,000+ companies across 175 exchanges in the corpus, the query "data center power" -semiconductor style refinements (the q parameter supports - negation) let you isolate exactly the non-obvious sectors where the theme is spreading.
Jackson Hole moved 985 conference calls
Powell's Jackson Hole speech signaled openness to a September cut. Markets rallied to record highs, and September-cut odds moved above 90%. On the corporate side, 985 companies discussed rate cuts on August calls — banks and REITs, obviously, but the long tail is where the texture is: homebuilders talking buyer psychology, staffing firms talking client budgets unlocking.
Here's where speaker-role filtering matters. A "rate cut" match in a management prepared statement is narrative; the same phrase in an analyst question is pressure. The search endpoint separates them — type=speakers&speaker_type=analyst restricts matches to analyst turns, which is the cleanest way to measure what the sell side was forcing management to address versus what management volunteered. The corpus's 11.9M speaker-tagged segments make that a query parameter, not a preprocessing project — the same role separation quants lean on when they build analyst-pressure features.
The date window is the other lever. Run the query for the full month, then again with date_from narrowed to the days after the speech, and the difference measures how quickly a September cut moved from a possibility executives hedged around to a base case they planned aloud.
Tariff refunds: 313 companies and counting
The aftermath of the Supreme Court's IEEPA ruling kept flowing through income statements: 313 companies matched "tariff refund" in August alone. The most instructive disclosure came from Walmart's August 20 call, which described deploying refund proceeds into price — refunds arriving as margin and immediately leaving as price investment.
That's the competitive dynamic worth querying for: once the largest retailer converts refunds into lower prices, every peer's call becomes a forced response. A standing monthly search on "tariff refund" is now effectively a tracker for how a one-time legal windfall is being competed away in real time. The AND operator sharpens it further — require the refund phrase and pricing language in the same call, and the matches narrow to companies actually describing the Walmart pattern rather than just acknowledging the ruling.
The cron job that writes this post
Everything above reduces to one endpoint. Each row of the table is a single request:
curl 'https://earningsapi.io/api/v1/search?q="agentic commerce"&type=transcripts&date_from=2026-08-01&date_to=2026-08-31' \
-H "X-API-Key: $KEY"
The q parameter supports "exact phrase" matching, AND/OR operators, and - negation. Loop over your phrase list, count distinct companies per phrase, diff against last month's file, and apply the reporting-calendar baseline before calling anything a trend. That's the entire methodology of this post — a monthly cron and a dozen queries. Swap type=transcripts for type=speakers&speaker_type=analyst and the same loop produces the analyst-pressure view.
For ingestion rather than search, the path is /api/v1/transcripts/recent with cursor pagination — since for a date floor, after_id to resume a cursor. The pattern for a new pipeline: backfill by paging after_id until the cursor is exhausted, then switch the same job to a daily since sync. Segments arrive already tagged by speaker and role, so the chunking work for an embedding store is done before you fetch — the full recipe is in feeding earnings call transcripts into a RAG pipeline.
curl 'https://earningsapi.io/api/v1/transcripts/recent?since=2026-08-01' \
-H "X-API-Key: $KEY"
Transcripts land in the corpus within minutes of calls ending, so an August 26 NVIDIA alert fires on August 26, not when a vendor batch arrives. And if your research surface is an agent rather than a script: any MCP client pointed at the EarningsAPI MCP server can run every search in this post conversationally — "which companies mentioned tariff refunds in August, excluding retailers?" — with the same date and speaker filters. Setup docs are at /mcp, and we've written up the full agent pattern in building an earnings research agent with MCP and Claude.
September's watchlist
The queries to keep warm as the calendar thins out:
- "agentic commerce", "humanoid robot", "stablecoin" — the three themes that outran the baseline. Small denominators, so one more doubling is meaningful.
- "supply-constrained" — if NVIDIA's framing propagates down the AI supply chain, it shows up here first.
- "tariff refund" — watch for the Walmart response pattern: refund language co-occurring with price-investment language at competing retailers.
- "rate cut" — after the September FOMC meeting, the interesting split flips from anticipation to reaction, and the analyst-turn filter tells you which managements get pressed on what they do with cheaper money.
September's calendar is quieter, which is exactly when theme counts get informative again — fewer companies reporting means every match carries more weight, once you normalize for the base.
All 4,809 August calls are in the archive alongside 253,000+ transcripts back through 2020, browsable at /transcripts and queryable through the same search endpoint used throughout this post.
Run this month's table yourself: the API docs cover search syntax, speaker filters, and cursor pagination — quarterly plans start at $145 on the homepage.