September 2026 Earnings Preview: What to Watch

The headline earnings season wrapped in mid-August, and the financial calendar pretends nothing happens until October. The database on my side of the screen disagrees. Roughly 10,000 earnings calls flow through it every quarter — 253,000+ transcripts and 11.99 million speaker segments going back to 2020 — and the flow never actually stops. It shifts to the off-quarter reporters: companies with fiscal quarters ending in July and August. In September, that cohort happens to contain most of what the market is currently arguing about — enterprise software, the memory cycle, and the American consumer.

One piece of housekeeping first. I'm writing this on August 25, and NVIDIA reports tomorrow, August 26. It's the market's main event this week, and since the call hasn't happened, you'll get no predictions from me. What matters for this post is simpler: every AI-adjacent name on the September calendar will be answering questions in whatever mood that call leaves behind.

Here's the month, then the six things I'll actually be listening for.

The calendar, week by week

These are scheduled report dates from our earnings calendar, and scheduled dates drift until companies confirm them — that's why Darden currently sits on both Sep 17 and Sep 24. Treat the week as firm and the day as provisional.

One structural point before the themes. These companies punch far above their headcount because of when their books close. A fiscal quarter ending in July or August covers months that no June-quarter reporter has said a word about. Between the big seasons, these calls are the only fresh corporate data the market gets — which is why a single Oracle or Micron call routinely moves an entire sector that reported six weeks earlier. About twenty companies get treated as proxies for a few thousand.

What the last complete quarter told us

Q2 2026 — calls held April through June, the last complete quarter in the corpus — was dominated by four themes, measured in companies matched: "rate cut" (2,183), "data center power" (2,174), "AI capex" (1,447), and "tariff refund" (381) — the biggest completed quarter on record for that phrase, after the Supreme Court's IEEPA ruling turned two years of tariff costs into refunds.

Methodology note, once for the whole post: every count here is companies matched by full-text search for the exact phrase in a quarter's transcripts. Keyword matching is an upper bound — a passing mention in a question counts the same as a ten-minute discussion — but the noise floor is roughly constant quarter to quarter, so the trends are real.

Two of those themes just got a catalyst. Jackson Hole is behind us and markets are pricing a likely September Fed cut, which makes September's reporters the first companies to field "what does a cut actually change for you" as a live question rather than a hypothetical. And the refund wave is still moving through P&Ls — the consumer names on this calendar will show us who's next.

Oracle and Broadcom: the AI capex handoff

"AI capex" matched 1,447 companies in Q2 — hyperscalers describing what they intend to spend, and everyone else describing what that spending does to them. September puts two companies squarely on the receiving end of the money: Broadcom (Sep 2) and Oracle (Sep 8–9).

What I'll listen for is not the revenue print — it's the forward-commitment language. Backlog and RPO — remaining performance obligations — is where hyperscaler spending becomes visible before it becomes revenue, and the phrasing around it often tells you more than the number. "Supply constrained" is the tell I care about most: a supplier that says it out loud is telling you demand exceeds everything it can ship, and that the pricing pen is in its hand, not the buyer's.

The other thing to expect: "data center power" matched 2,174 companies in Q2 — more than "AI capex" itself. Both companies will get power questions in Q&A whether they raise the topic or not. I traced the whole spending chain from hyperscaler to supplier in the AI capex supercycle post if you want the multi-quarter view.

Micron: the memory read

Micron (Sep 22–23) is the closest thing this calendar has to a commodity read on AI demand. GPUs are supply-managed and story-rich; memory is cyclical and honest. When AI demand is real, it shows up as HBM — high-bandwidth memory — commitments. When it wobbles, memory pricing wobbles first.

So HBM commentary is the tell. Specifically: how far out management says HBM supply is spoken for, and what the wafer trade-off does to conventional DRAM. Every wafer allocated to HBM is a wafer that doesn't become commodity memory, which is how an AI demand story turns into a pricing story for everything else. If prepared remarks quantify that allocation, I'll take the number. If they retreat to adjectives, that's information too.

Timing helps here. Micron's fiscal quarter runs through late August, so by the time management speaks on Sep 22–23, it's describing weeks of demand that nobody else on the calendar has reported on yet. That's the freshest AI-demand data point the market will get before the October season opens.

Salesforce, Snowflake, MongoDB, UiPath, C3.ai: numbers or adjectives

The first week of the month stacks the enterprise software cohort: MongoDB (Sep 1), Salesforce, Snowflake, C3.ai (Sep 2), UiPath (Sep 3). These are companies that have spent quarters telling us agents will change their business. September is the next checkpoint on whether agent products show up in the numbers or still just in the adjectives.

My test is mechanical. First, count "agentic" in prepared remarks — density there is a management choice. Then read the analyst pushback: are analysts asking what the product does, or how much of it got sold? The moment Q&A shifts from roadmap to attach rates, customer counts, and revenue contribution, the theme is monetizing. Deflections — "early innings," "unprecedented interest" — mean another quarter of adjectives.

This is also why the clustering matters: five calls in three days is a cohort, and a cohort answers questions any single company can dodge. If four of five get the same monetization question and only one answers with a number, that one call reprices the other four.

The quarter-over-quarter comparison is the other half of the test. Prepared remarks are written, rehearsed, and chosen — so when a company's own phrasing shifts from "opportunity" language to booked-business language between two of its calls, that shift is deliberate. I keep last quarter's transcript open next to the live one for exactly this reason. Same speaker, same product, ninety days apart: the delta is the signal.

Nike, Lululemon, Costco, Kroger, Darden: the tariff-refund tail

The consumer names report with the refund wave still rolling. 381 companies matched "tariff refund" in Q2 — a record for a completed quarter — and August handed us the playbook to compare against: Walmart told investors it received refunds in the quarter and deployed much of the money straight back into price.

That makes September's consumer calls a three-way sort. Who quantifies the refund — an actual figure in prepared remarks, not "a benefit"? Who invests it in price, following Walmart — margin-neutral for them, margin poison for competitors who don't match? And who lets it flow through and raises the guide? Lululemon (Sep 3), Kroger (Sep 10), Darden, Costco (Sep 24), and Nike (Sep 29) each land somewhere on that map. Nike — which sources nearly everything abroad and reports last — will get the question either way, with a month of competitor answers already on the record.

The rest of the consumer slate — Chewy (Sep 9), General Mills (Sep 16), CarMax (Sep 24), GameStop (Sep 8) — gets the same sort applied further down the list. By now the refund question is standard analyst equipment, which means silence is also an answer: a consumer company that takes no refund question and volunteers no refund number in September is telling you something about its sourcing, its analysts, or both.

The full history of how a Supreme Court ruling turned a cost line into a capital-allocation debate is in the tariff refund quarter post.

Lennar: the first rate-cut read

"Rate cut" matched 2,183 companies in Q2 — the most-matched theme in the corpus last quarter — and essentially all of it was anticipation. With Jackson Hole behind us and a September cut priced as likely, Lennar (Sep 17) is the first genuinely rate-sensitive reporter to hold a call with that cut delivered or imminent.

The near-term mechanics — incentives, mortgage buydowns — will get the headlines. I'll be listening for the 2027 planning language instead: land acquisition pace, community count, and how management talks about demand if financing costs actually fall. A homebuilder that starts committing capital to 2027 volume on a September call is telling you what it thinks the cutting cycle is, not just the cut.

How I'll work the month

Two workflows, both cheap.

Per call: 15 minutes of prep. Before each call, I pull the company's prior transcripts through the MCP server and have Claude surface the open loops from last quarter — what management promised, what analysts pushed on, what never got answered. The full workflow is in how I prep any earnings call in 15 minutes. Every call above lands in the transcript archive shortly after it ends.

Per month: theme tracking across the cohort. One search endpoint answers "who said it" across every September transcript. This is the September slice of the tracker I built in the theme tracker post:

curl -s "https://earningscalls.dev/api/v1/search?q=%22agentic%22&type=transcripts&date_from=2026-09-01&date_to=2026-09-30&limit=100" \
  -H "X-API-Key: $EARNINGSCALLS_API_KEY"

Swap the query for %22supply constrained%22, %22tariff refund%22, or %22rate cut%22 and the whole watchlist above is four requests. Run it weekly, diff the tickers against last week's file, and September stops being a quiet month — it becomes the first read on every theme Q4 will be priced on. Full parameter reference is in the docs.

Want to run this on your own watchlist? earningscalls.dev has 253,000+ earnings call transcripts from 12,799 companies — web, REST API, and MCP server, with full-text search across every call since 2020.