The Inference Buildout: What 315 Companies Said in Q2 2026

I spend most of my working hours inside a database of 253,129 earnings call transcripts. When a phrase starts spreading through that corpus, I usually notice before I can explain it. This quarter, the phrase is "inference demand" — and the explanation matters more than the count.

In Q2 2026, 315 companies matched "inference demand" in our full-text index. That is an all-time high across every quarter we cover, going back to 2020. And it happened in a quarter that is normally the seasonal low for this kind of language.

Here is why I think that number is the clearest signal yet that the market's AI vocabulary is shifting — from training to inference, from capex announcements to unit economics.

The number, in context

Training capex built the current data center map. That story has been running for two years, and I tracked it in the AI capex supercycle in earnings calls. But capex tells you where the money went. Inference economics — cost per token, latency, where the compute physically sits — decide who profits from what got built.

The mention data shows the vocabulary catching up to that reality:

A quick note on methodology, because I want you to trust these numbers for what they are: each count is the number of companies whose earnings calls match a full-text search in a given quarter. Keyword matching means the counts are an upper bound — some matches are incidental co-occurrences, an analyst asking about someone else's inference demand, a passing reference. I count companies rather than raw mentions deliberately: one enthusiastic CEO can repeat a phrase fifteen times in a single call, and what I care about is breadth of adoption, not the enthusiasm of a few. The trend is what matters, not the third decimal.

And the trend survives the obvious objections. Q2 is usually the quiet quarter for this phrase — 2025 Q2 dipped to 169 companies. Part of that dip was the denominator: 2025 Q2 carried only 8,270 calls in the corpus against a typical quarter of roughly 10,000. Q2 2026 had 10,053 calls, so the raw comparison flatters the growth. But normalize it per call and the signal is still there: roughly 2% of all calls matched in Q2 2025, about 3.1% in Q2 2026. More calls, and a meaningfully higher hit rate on top.

The quarter-over-quarter picture says the same thing. Q1 2026 was already a record at 282. A seasonal-low quarter beating the previous all-time high by 33 companies is not noise.

Who is talking — and why the mix matters more than the count

The sector split for Q2 2026: Information Technology leads with 190 companies, followed by Industrials at 49, Materials at 38, Communication Services at 37, Consumer Discretionary at 36, and Financials at 36.

IT dominating is the unsurprising part. The surprising part is everything below it — five sectors, none of them chip designers, each with dozens of companies discussing inference demand on the record.

Now look at the top mentioners since April 2026: Arista Networks, Sandisk, AMD, Datadog, DigitalOcean, and Qualcomm. Read that list as a stack, because that is what it is: networking, storage, compute, observability, mid-market cloud.

The first three are the buildout itself — the switches, the flash, the accelerators. That is expected, though Sandisk's presence deserves a beat of its own: inference at scale is as much a memory and storage problem as a compute problem. Serving a model means moving weights and context in and out fast enough to keep the expensive silicon busy, and when storage vendors are fielding inference-demand questions, buyers are optimizing the whole rack, not just the accelerator.

The last three names are the tell.

When Datadog talks about inference demand, it is because customers are running inference workloads that need monitoring — which means those workloads exist at scale, in production, at ordinary software companies. When DigitalOcean talks about it, inference has reached the mid-market: companies too small to negotiate hyperscaler committed-use discounts are still buying GPU time. Inference demand has stopped being a hyperscaler story and become an operating cost line — something a CFO at a normal SaaS company now has to explain on an earnings call, the same way they explain cloud spend.

That is the structural change hiding inside the number 315. The phrase is no longer describing a product a few companies sell. It is describing a cost most companies pay.

Qualcomm's thesis is a testable prediction

The single most interesting use of the phrase this season came from a company that was not in this market a year ago. On its July 29 call, Qualcomm laid out its rationale for entering the data center:

"This is the ideal and logical time for Qualcomm to enter the market as agentic workloads are reshaping the economics of AI." — Qualcomm, earnings call, July 29, 2026

And then the architectural claim, which is the part I keep re-reading:

"Efficient token generation and total cost of ownership are fundamental to scaling AI. And as a result, inference is becoming disaggregated in the data center and will be increasingly distributed." — same call

Strip the investor-day polish and there are two claims in there. First: agentic workloads — the always-on, many-small-calls pattern I watched spread in 2,400 earnings calls last year — change the shape of inference from batch jobs to a continuous utility. Second: once inference is a utility, it stops living exclusively in giant centralized clusters. It disaggregates inside the data center and distributes outward — toward the edge, toward on-prem, toward wherever latency and cost per token say it should sit.

Here is what I like about this as a data person: it is a falsifiable prediction, and mention tracking is the scoreboard. If Qualcomm is right and inference genuinely moves to the edge and on-prem, the phrase "inference demand" should not stay concentrated in Information Technology. It should leak into the sectors that own physical infrastructure and regulated data — Industrials first, Health Care after that. Factory-floor inference, hospital-network inference, logistics inference.

Industrials at 49 companies this quarter is already a hint. The sector split table is the thing to watch. I will re-run this exact search in October and January, and if the IT share keeps falling while the absolute count keeps rising, the disaggregation thesis is playing out in the transcript record before it shows up in anyone's segment reporting.

If IT's share holds steady or grows, the thesis is stalling — and that is worth knowing too.

Token economics is becoming CFO vocabulary

There is a second-order shift underneath the mention counts, and it is about who says the words and how.

Two years ago, inference lived in the CEO's prepared remarks — vision language, TAM language. This season, the framing has moved down the income statement. Qualcomm's own quote is instructive: "efficient token generation" and "total cost of ownership" in the same breath. That is not vision vocabulary. That is procurement vocabulary.

Cost per token is turning into a unit economic the way cost per click did for ad-supported businesses and cost per GB did for storage. When a metric crosses from engineering dashboards into earnings call language, it means analysts have started asking about it — and management teams have started preparing answers. The demand line, the workload category, the deployment architecture: inference is now discussed as all three, sometimes within a single call.

This is exactly the kind of signal that text data catches before financial data does. Segment reporting will not show you "inference revenue" as a line item for years, if ever. But the language shift is measurable today — it is one of the five text signals quants pull from earnings calls, and it is the one moving fastest right now.

Run it yourself

Everything above is one API call plus a group-by. Here is the search that produced the Q2 2026 numbers:

curl 'https://earningscalls.dev/api/v1/search?q="inference demand"&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 — so you can tighten the query if the incidental co-occurrences bother you. "inference demand" -training gives you a stricter cut. "inference" AND "cost per token" gets you closer to the CFO-vocabulary signal specifically.

Two refinements I use constantly:

Scope by sector. Re-running the same phrase filtered to Industrials or Health Care is how you turn the Qualcomm thesis into a quarterly tracking job. Ten minutes a quarter, one number per sector, and you have a scoreboard nobody else is publishing.

Scope by speaker role. Switch to type=speakers&speaker_type=executive and you separate management narrative from analyst probing. This distinction matters: executives volunteering inference language in prepared remarks is a different signal from analysts dragging it out of them in Q&A. The corpus has 11.99 million speaker segments tagged by role, so both cuts are cheap. The full parameter reference is in the API docs.

If you would rather explore conversationally, the same search runs through our MCP server from Claude or any MCP client — I do most of my first-pass exploration that way and drop to raw API calls once I know what I want to count.

What I am watching next quarter

Three things, all cheap to check:

1. The absolute count. Q3 has historically run above Q2. If the 315 baseline holds the pattern, Q3 2026 should print well above 300 — anything flat or down would be the first crack in the trend since early 2025.
2. The sector split. Specifically the Industrials number, as the first test of the disaggregation thesis. 49 is the number to beat.
3. The speaker mix. Whether executive-initiated mentions keep growing relative to analyst-prompted ones. That ratio tells you whether inference economics is a story management wants to tell or a question they have to survive.

The training buildout was announced in press releases. The inference buildout is being negotiated line by line on earnings calls — in cost-per-token language, in TCO language, in capacity-planning language. If you want to see where AI compute actually settles, the transcripts are where the map is being drawn, roughly 10,000 calls per quarter at a time.

Track "inference demand" — or any phrase you care about — across 253,129 transcripts with the earningscalls.dev API.