Who Actually Buys From Industrial Software Startups?
Somebody finally counted the buy side. 2,223 verified customers of industrial software startups, and the five things the list turned out to say about who buys, what they buy, and why they buy it despite already owning a module that does the job.
AI Answer
ThreadMoat verified 2,967 purchase relationships between 2,223 organisations and 475 industrial software startups. Key findings: 86.3% of buyers hold exactly one startup vendor; 97.6% of SMBs bought exactly one, ever. Expansion is a cliff, not a gradient — 0.8% of companies under 50 staff buy across more than one software category versus 47.0% of companies above 50,000. Of the 614 purchases extending the Siemens ecosystem, 96.6% land in a category Siemens already ships a product for, so startups are not filling portfolio gaps. Buyer types diverge sharply: government and defence bodies hold 1.88 startup vendors each versus 1.02 for SMBs, and agentic-orchestration software is bought by startups (24.1%) and SMBs (18.4%) at well above the enterprise rate (12.3%).
Everyone in industrial software has an opinion about who buys from startups. Almost nobody has a list.
So we built one. Every customer we could verify for the startups we track — 2,223 organisations, 2,967 confirmed purchase relationships, 475 named startups. Not logos on a website. Not press releases. Confirmed relationships, each carrying the buyer's industry, country, employee band, revenue band, and the category of what they bought.
Here is what the list turned out to say.
The honest limit, stated first
439 of the 935 startups we track can name a customer at all — 47%. Counting vendors named by buyers but not yet matched to the tracked universe takes the ceiling to 50.8%.
Just under half the ecosystem has no findable buyer. Everything below describes the half that can be seen. "Verified" means the relationship was confirmed, not that contract value or deployment scope is known, and every count means "startups tracked by ThreadMoat that name this customer" — large buyers purchase from more vendors than we track.
1. There is no such thing as an industrial software buyer
Split 2,223 organisations by what kind of organisation they are and the market stops behaving like one market.
| Buyer type | Startup vendors held, average | Bought exactly one |
|---|---|---|
| Government & defence | 1.88 | 83.7% |
| Enterprise | 1.43 | 83.0% |
| Startup | 1.17 | 87.5% |
| Research institute | 1.09 | 93.2% |
| SMB | 1.02 | 97.6% |
Public bodies buy deepest. That number moves depending on where you draw the border, and it is worth being precise: 43 organisations are typed as government or defence and average 1.88 vendors. Widen it to all 92 bodies working in the public sector and it falls to 1.47 — still above enterprise, no longer dramatically. Both numbers are defensible; the honest claim is the range.
The mechanism is structural rather than cultural. The US Department of Defense alone appears as eleven separate purchasing entities holding 38 relationships between them — not one adventurous buyer, but eleven buyers each acting once or twice.
The SMB figure needs no such caveat. 97.6% of the 418 small and mid-sized businesses in the dataset bought exactly one startup, ever. Ten of them hold more than one. If a go-to-market plan assumes an SMB lands, expands and becomes an account, this data says that essentially never happens.
They also want different things. SMBs are two and a half times more likely than enterprises to buy progress monitoring (12.9% against 5.3%) — the small construction and field-services end of the market. Enterprises are nearly twice as likely to buy operations intelligence (21.6% against 11.7%).
2. Every industry buys a different thing, and none of them buys the average
Cross-tabulate what each customer bought against the industry it operates in, and the verticals separate hard.
| Industry → category | Times more likely than average | Absolute share |
|---|---|---|
| Logistics → Supply chain intelligence | 3.1× | 20.8% |
| Aerospace & defence → Digital thread | 3.0× | 49.7% |
| Construction → Progress monitoring | 2.8× | 17.8% |
| Automotive → Analysis & simulation | 2.6× | 32.9% |
| Food & beverage → Operations intelligence | 2.5× | 44.4% |
| Construction → CAD acceleration | 2.4× | 60.1% |
| Energy & utilities → Operations intelligence | 2.0× | 36.7% |
Read the first row as: a logistics company is 3.1 times more likely to have bought supply chain software than the average company in the dataset.
None of these verticals is buying "AI for manufacturing." Each is buying a fix for the specific thing that costs it money. Aerospace has a certification problem, so it buys traceability. Automotive has a prototype cost problem, so it buys simulation. Food and beverage runs continuous lines where an hour of drift is an hour of scrap. Construction has a drawing problem and always has.
Which makes the horizontal pitch — "we work across discrete manufacturing" — a pitch to the average column. The average column is the mean of twelve industries that each want something different, and no real company sits at it.
3. They already own a module that does this
The assumed answer to "why would a Teamcenter customer buy from a startup" is that the incumbent has a gap in its portfolio.
We tested it. Of the purchases in our data, 1,097 of 2,779 — 39% — involve a startup that declares compatibility with an incumbent platform rather than replacing one. So mostly these customers are not leaving. They are bolting something on.
Then we took the 614 purchases that specifically extend the Siemens ecosystem and asked a simple question: does this startup work in a category Siemens already ships a product for?
| Category | Purchases | Siemens product |
|---|---|---|
| Process optimization | 161 | Tecnomatix / Opcenter |
| Operations intelligence | 130 | Opcenter |
| CAD acceleration | 107 | NX |
| Digital thread | 99 | Teamcenter |
| Analysis & simulation | 96 | Simcenter |
| Progress monitoring | 18 | none |
| Supply chain intelligence | 3 | none |
593 of 614. 96.6% land on a category Siemens already sells. Twenty-one purchases sit in genuine white space.
These customers are not filling gaps in the portfolio. They already own something that nominally does the job, and they bought a startup anyway. Siemens does it to itself, too — it is the single largest buyer of industrial software startups in the dataset, and 25 of the 30 purchases we can classify are in categories Siemens itself sells.
What this cannot tell you: nothing in the data records why anyone bought anything. There is no reason field, no displacement flag, no record of what else was evaluated. It shows where these purchases land against the portfolio, not the argument that happened in the room.
What the vendors are doing is suggestive, though. Physics-informed surrogate models. Generative topology optimisation. 3D geometric part search. Code-driven CAD. Autonomous CAM. These are not cheaper versions of the module — they are capabilities that postdate the architecture the module was built on. Owning the category on the price list is not the same as owning the capability.
4. Nobody expands until ten thousand people
86.3% of these buyers hold exactly one startup vendor. One vendor, one category, then nothing.
Banding every organisation by employee count shows when the second purchase actually happens.
| Employee count | Buys across >1 category | Startup vendors per company |
|---|---|---|
| Under 50 | 0.8% | 1.01 |
| 50 – 250 | 3.6% | 1.05 |
| 250 – 1,000 | 4.0% | 1.07 |
| 1,000 – 10,000 | 8.6% | 1.14 |
| 10,000 – 50,000 | 19.7% | 1.38 |
| Over 50,000 | 47.0% | 2.97 |
That is not a gradient. It is a flat floor and then a cliff, and the cliff starts around ten thousand people. A company with 50,000 employees is roughly sixty times more likely to buy a second category than one with fifty.
The reason is boring and structural. A 500-person manufacturer has one person who evaluates software, and they do it between other jobs. A 50,000-person manufacturer has an innovation function, a digital group, and four plants running pilots that don't know about each other. Expansion at that scale is an org chart, not a product achievement.
Every organisation also carries an independently-estimated revenue band, and it finds the same cliff in the same place: 1.6% of sub-$10M companies buy across more than one category, against 36.3% of companies above $10B.
5. The leading edge isn't the enterprise
Score every product for how AI-intensive it actually is, then split by who bought it. Startups and university labs buy software materially more AI-intensive than enterprises do, from vendors half a generation younger — median founding year 2020 against 2018.
The sharper cut is agentic orchestration, the newest category in the market.
| Buyer type | Share of purchases with an agentic thesis |
|---|---|
| Startups | 24.1% |
| SMBs | 18.4% |
| Enterprises | 12.3% |
| Research institutes | 9.4% |
| Government | 6.1% |
That ordering is not what the discourse assumes. Agents are supposed to be an enterprise story with a long pilot phase in front of them. Here the small companies are ahead of the large ones, and the public sector is a full generation behind both.
Our read is that the mechanism is risk rather than appetite. A large manufacturer's procurement function is an insurance mechanism whose job is to not be wrong, and a 2023-founded vendor fails that underwriting however good the product is. A forty-person shop has no procurement function to fail — it has an owner who can decide on a Tuesday.
What's in the Q3 report
Five findings is what fits in a week. The dataset supports considerably more, and the following are finished, verified and deliberately held back for the Q3 report, where they get the full graphics and analysis treatment:
- The full industry × category matrix — all twelve industries against all seven software categories, with sample sizes.
- Portfolio overlap for every major platform — the 96.6% test run against Dassault Systèmes, PTC, Autodesk, SAP, AVEVA, Ansys and Aras. The spread between them is the story.
- Which industries actually spend — startup vendors held per customer by vertical, normalised for company size. One industry runs at nearly double the market rate, and the most distinctive buyer in the dataset turns out to be among the shallowest.
- Funding stage of what gets bought — seed through Series D+, and the conversion curve showing where a startup's odds of naming a customer peak, and where they start falling again.
- Geography, home bias and the ITAR fingerprint — which markets buy locally once you normalise for what is available, and the six-fold gap between US and non-US aerospace on cross-border purchasing.
- Vendor specialisation — how narrowly the sell side has specialised by buyer type, including one count suggesting an entire market segment has no native vendors at all.
Source: ThreadMoat customer census, export of 24 August 2026. 2,223 organisations, 2,967 verified relationships, 475 named startups, all rows verified. Counts mean "startups tracked by ThreadMoat that name this customer."