I spend my working hours inside a database of 253,129 earnings call transcripts — 12,799 companies, 11.99 million speaker segments, every quarter from 2020 to today. When a theme gets loud, I don't read hot takes about it. I run a query.

So here is the query: the exact phrase "humanoid robot", across every earnings call in the archive.

I expected robot makers. What came back was magnets, bearings, sensors, wiring and hard drives. The companies talking about humanoid robots on their own earnings calls — in front of their own analysts, with guidance on the line — are overwhelmingly the companies selling parts into them.

That is the whole thesis. The public humanoid trade, as told by the people signing purchase orders, is a components trade. Let me show you the receipts.

From 6 companies to 89 in eight quarters

Companies mentioning "humanoid robot" on at least one earnings call, per quarter:

Quarter Companies
Q1 2024 6
Q2 2024 12
Q3 2024 15
Q4 2024 11
Q1 2025 38
Q2 2025 34
Q3 2025 47
Q4 2025 53
Q1 2026 89
Q2 2026 80

One methodology note, and then I'll shut up about it: these counts are earnings calls matching a full-text search. Keyword matching means the numbers are an upper bound — some matches are incidental co-occurrences, a passing analyst question or a throwaway aside. The shape of the curve survives that noise comfortably.

And the shape is unambiguous. 6 companies in Q1 2024. 38 a year later. 89 a year after that — a roughly 15x increase in eight quarters, against a steady backdrop of about 10,000 calls landing in the database every quarter. This is not a theme drifting through the language. It is a theme being adopted, quarter by quarter, by managements who have decided it belongs in their prepared remarks.

Look at the sub-structure too. Through 2024 the count idles in the low teens — a curiosity a handful of managements bring up when asked. Then Q1 2025 more than triples the prior quarter, from 11 to 38, and the count never comes back down. That kind of discontinuity in transcript data usually means something changed in the boardroom conversation between quarters: the theme stopped being a question analysts ask and became a line executives volunteer.

The interesting part is who adopted it.

The leaderboard reads like a bill of materials

Rank the companies by mentions since April 2026 and the top of the list is: Western Digital, MP Materials, Aptiv, Timken, Allegro MicroSystems.

That is storage, rare earth magnets, wiring and sensing, bearings, and motor-control chips. A hard-drive company tops a humanoid-robot search — exactly the kind of result no stock screener would ever hand you.

Not one pure-play humanoid maker in the top ranks. The reason is structural: the makers are mostly private. They don't hold earnings calls, so they leave no trail in a transcript database — and, more importantly for anyone allocating capital, they leave no ticker. The public market largely cannot buy the robot. It can buy the robot's magnets, its joints, its sensors, its wiring.

Which means the earnings call record is the supplier record. And when you actually read what these five say, you get something better than a mention count: you get three different time horizons inside the same theme.

Horizon one: the magnet ceiling

Start with the hardest constraint in the whole story. On MP Materials' August 6 call, management did the arithmetic out loud:

"If we were to be producing, say, 30 million or 40 million humanoid robots a year in the world, that alone would eat up 100% of the rare earth magnet production globally, including China."

Read that again. Not "demand is strong." Not "we see a large TAM." A physical ceiling: at 30–40 million units a year, humanoids alone consume the entire global supply of rare earth magnets — China included.

This is what I mean by suppliers telling the real story. A robot maker's pitch is about what the robot will do. A magnet maker's math is about what can physically be built. If the humanoid ramp is real, it collides with magnet capacity long before it collides with demand — and the company that mines and magnetizes the material said so on a recorded, transcribed earnings call.

That's the longest horizon: a capacity constraint that shapes the entire decade.

Horizon two: design wins are booked signals

Now the near end of the timeline. On Allegro MicroSystems' July 30 call, the language is not aspirational at all:

"This quarter, we secured current sensor wins with large Chinese humanoid robot OEMs."

And in the same breath:

"a large design win with a prominent North American humanoid robotics OEM that is using our inductive position sensors in robotics joints."

Design wins are the closest thing a semiconductor supplier has to booked future revenue — a customer has committed your part into their build. Allegro named two geographies and a specific component location: position sensors in the joints. On the call, robotics and automation were sized at about 3% of the revenue mix. Small — but it is present-tense, contracted, on-the-books small, not conference-keynote small.

The Timken Company's August 4 call frames the same near-term reality from the bearings side:

"whether you're talking about a robot, whether you're talking about humanoids, robot transfer units, medical robots, or just in general, factory automation, it's a pretty robust portfolio"

Followed by the line I'd put on the cover of this whole dataset:

"this demand is not new, but the implementation of it at scale is certainly ramping up."

A 100-year-old bearings maker is telling you the humanoid category is not a new invention to them — it is an existing motion-control business shifting into a higher gear. That is a much more credible bull case than any launch video.

Horizon three: the volume reality check

And then there is the corrective. On Aptiv's August 12 call — two days before I'm writing this — management was asked about humanoids and gave the most honest answer in the sample:

"we're doing work in our close to commercial agreements with a couple of the players in the humanoid space. We think that's an opportunity. We just think from a volume standpoint, that's further out."

"Further out." A wiring and sensing supplier that is already in near-commercial agreements with humanoid players is telling you volumes are not imminent.

Put the three horizons side by side and you get the actual state of the trade, straight from the people quoting parts:

No single call gives you that picture. The mention count alone doesn't either. You need the counts and the excerpts — which is the same lesson quants keep re-learning about transcript text; I wrote up the recurring patterns in five text signals quants pull from earnings calls.

GICS has no row for "humanoid exposure"

Here is the Q2 2026 sector split for companies matching the search: Information Technology 56, Industrials 24, Consumer Discretionary 23, Communication Services 6, Materials 5, Health Care 1.

IT plus Industrials dominate — chips, sensors, motion components, factory automation. That tracks perfectly with the supplier thesis. But notice what the split also tells you: you could not have built this list with a stock screener. "Humanoid exposure" is not a GICS category, not an industry code, not a tag in any classification system. It cuts across storage, mining, auto parts, bearings and semiconductors — five companies that share almost nothing except a customer.

The only place that exposure is declared, in the companies' own words, on the record, quarter after quarter, is the earnings call. Full-text search over transcripts is, as far as I can tell, the only systematic way to find it.

That single Health Care match is worth a second look, by the way — not because one company is a trend, but because it shows the search doing what a classification system can't: catching a theme the moment it crosses a sector boundary, one call at a time.

We've seen this movie before

If the pattern feels familiar, it should. When I traced "agentic AI" through 2,400 earnings calls, the earliest voice in the archive wasn't a hyperscaler — it was a mid-cap chip supplier, using the phrase about a year before NVIDIA said it on a call. The supply chain talked first, because the supply chain sees the purchase orders first.

Humanoids are running the same script, with one twist: this time the headline names may never show up in the transcript data, because they're private. The suppliers aren't just the early signal here. They are the only signal — and the only investable one.

Run it yourself

Everything above is one API query, re-run with different date windows. Here is the live version for calls since April 2026:

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

The q parameter supports exact phrases in quotes ("humanoid robot"), AND/OR operators, and -negation to strip terms you don't want. Full parameter reference is in the docs, and if you'd rather ask in plain English from Claude or another client, the same search is exposed through the MCP server.

The workflow I'd actually recommend: pin the query, re-run it in the week after each earnings season peak, and diff the company list against last quarter. New names entering the list are suppliers announcing exposure for the first time — that's the highest-signal event in the whole series. If you want the scaffolding for that, I walked through the full setup in build an earnings call theme tracker in Python; swap in the query above and you have a humanoid supply-chain tracker in an afternoon.

The robots get the keynotes. The suppliers get the purchase orders — and they talk about them every quarter, on the record, in a database you can query.

Search all 253,129 transcripts yourself — get an API key at earningscalls.dev.