What the U.S. Data Actually Show
The subject here is the relationship between advanced-technology adoption, venture-capital finance, and the performance of U.S. firms. The immediate task is narrower than explaining why startups succeed: it is to determine what the Census evidence can say about the four combinations of technology and finance, and how much of their apparent advantage remains after observable differences are taken into account. Reading the data correctly gives readers a durable judgment rule—separate the population being measured, the comparison being made, and the causal claim being implied.
The distinction matters because “technology firm” and “VC-backed firm” are not natural species. They are categories constructed from surveys, finance records, and linked business histories. A category can be useful without being complete, and a large difference between categories can be real without revealing its origin.
Measurement before interpretation
The Annual Business Survey, or ABS, is the foundation for the technology measure. It asks nonfarm employer businesses whether they use selected technologies in producing goods or services. The selection includes advanced technologies such as artificial intelligence, robotics, RFID, additive manufacturing, and distributed ledgers, alongside related tools measured in particular survey years. The Census paper uses the word “tech” for this broader group. It is not a direct count of AI startups, and it is not a measure of whether a firm has adopted the latest generative model.
The survey instrument also changes over time. In one vintage, a firm might be counted as using a technology when it reports use in at least a small share of production; in another, the response options distinguish testing from actual use. The 2023 ABS asks about AI and robotics and records first adoption in five-year windows, such as 2016–2020 or 2021–present. That timing information is valuable, but it is coarse. It can establish an interval in which adoption occurred, not the exact month or year, nor the precise organizational decision that caused it.
The scope of the sample is equally important. The ABS covers employer businesses, not every person with an idea or every very early venture before it hires anyone. The 2018 and 2023 surveys each sampled about 850,000 businesses, while the annual surveys in between sampled roughly 300,000. Those are unusually broad sources for studying business technology, but broad coverage does not remove the boundary around what is observed. A conclusion about firms represented in these data should not be silently expanded into a conclusion about all entrepreneurial activity.
VC status is assembled through a different route. The paper primarily uses PitchBook records of venture financing, supplements them with a question about initial capital in the 2018 ABS, and adds information on initial public offerings. These records are matched to the Census Business Register using company names and addresses, then connected to the ABS and other Census data through firm identifiers. The result is a measure of whether a firm has been identified as receiving VC, not a complete account of every dollar, investor conversation, or informal source of entrepreneurial finance.
Employment, revenue, age, industry, and location come from the Longitudinal Business Database, which covers nearly all U.S. nonfarm employer businesses over a long historical span. This linkage lets researchers compare firms' outcomes with their technology and finance status and, in the timing analysis, follow size before and after an event. It also makes the comparison more meaningful than a one-time survey of founders. But the measures remain different kinds of evidence: technology use is largely self-reported, VC status is constructed from finance records and survey responses, and employment and revenue are longitudinal business outcomes.
The construction is a strength because it joins large-scale business information that is rarely available in one file. It is also a reason to keep the unit of analysis in view. The main descriptive results concern firms observed in ABS 2018–2022 and linked to LBD outcomes from the relevant years. The adoption-timing analysis uses the 2023 ABS and links it to a longer LBD history. A statistic from one design should not be treated as though it came from the other.
The weighting scheme matters as well. The reported firm, employment, and revenue shares are intended to represent the U.S. employer-business population, so they answer an aggregate question: how much of the economy is associated with firms in a category? The regression coefficients answer a different question: after accounting for specified industry, place, age, and later business characteristics, how different are the outcomes of firms in one category from those in the reference category? Neither calculation answers the counterfactual question of what the same firm would have achieved if it had chosen another technology or financing source. That question requires a comparison over time or across firms that is credible on the relevant pre-treatment conditions.
The linked records also make it possible to keep firm age and survival in the discussion. A young company has had less time to accumulate employees and revenue than an adult company, while a firm that remains observable long enough to appear in later outcomes is different from an idea that never becomes an employer. Controlling for age helps with one source of comparison error; it does not turn the observed population into a random draw of all possible startups. The strength of the data is therefore their breadth and longitudinal connection, not a guarantee that every selection process has been measured.
The four cells reveal a pattern, not a verdict
The first descriptive fact is a mismatch between the share of firms and the share of economic activity. In the weighted ABS 2018–2022 sample, advanced-technology firms represent 11.63% of firms, but account for 38.76% of employment and 44.11% of revenue. VC-backed firms represent only 0.64% of firms, yet account for 13.26% of employment and 14.98% of revenue. These numbers do not say that technology or VC created the firms' size. They do say that a small, selected population can matter greatly for aggregate employment and revenue.
This is the base-rate problem behind many startup narratives. If one looks only at successful VC-backed technology companies, it is easy to mistake a concentrated tail for a typical outcome. The data require two questions at once. How much larger is a firm in a category than a comparable firm? And how many firms belong to that category? A group can have a very large effect per firm but a limited reach, while a broader group can affect the economy through many smaller firms.
The Census analysis makes the comparison explicit by separating firms into four mutually exclusive groups. The first uses advanced technology and has VC backing. The second has VC backing but is not classified as an advanced-technology user. The third uses advanced technology without VC backing. The fourth has neither. This arrangement does not prove that either input is beneficial. It prevents a common analytical mistake: attributing every difference between the first and fourth groups to the two labels while ignoring what the middle groups reveal.
In the baseline regressions, the VC-and-technology group has the highest employment and revenue relative to the group with neither characteristic. The estimated coefficients for log employment and log revenue are 1.660 and 1.479. For VC-backed firms without advanced technology, the corresponding coefficients are 1.051 and 1.048. For non-VC technology users, they are 0.483 and 0.626. The ordering is therefore not merely “both versus neither.” VC is associated with a larger outcome gap among technology users, and technology is associated with a larger outcome gap among VC-backed firms than among the omitted group.
A log outcome requires care in translation. The paper converts the first pair of coefficients into approximate employment and revenue gaps of 425% and 338% relative to the omitted category. Those are model-based comparisons after the specified fixed effects, not percentage-point increases in employment or revenue caused by receiving a particular investment. They describe how the conditional outcome levels differ across categories. They should not be read as an expected return for an individual founder or investor.
The fixed effects account for year-by-four-digit-industry interactions, state, and firm age. This matters because industries differ in their production processes and revenue levels, regions differ in business ecosystems, and older firms have had more time to grow or fail. Holding those dimensions constant makes the four-way pattern more informative. It does not make the groups identical. The decision to adopt technology or pursue VC can still reflect motivation, technical capability, private information, customer demand, or an idea's potential.
The paper tests that concern with a richer set of observable controls from the ABS. These include the owner's advanced degree, prior business ownership, age, lifestyle motivation, process and product innovation, patents, the importance of intellectual property, and a growth strategy. Once those characteristics are added, the coefficient linking VC status to technology adoption falls from 0.106 to 0.061, but remains statistically significant. The gaps in employment and revenue also attenuate as more controls enter, yet remain meaningful in the fully controlled specifications.
That pattern tells us something specific about selection. Observable founder and business characteristics explain part of why VC-backed firms are more likely to use advanced technology and why the four groups differ in size. They do not explain all of the measured association in these specifications. But the remainder cannot automatically be called a financing effect. Unobserved ability and suitability are still plausible. A founder may know that a production method fits the idea before the data can measure it. An investor may recognize a demand signal that is absent from the control variables. A firm that is already on a steep growth path may adopt technology and raise capital in response to the same opportunity.
There are further reasons to resist a universal reading. The broad technology indicator combines technologies with different costs, uses, and maturity. The VC category does not separately identify angels, retained earnings, grants, corporate investors, or every bank arrangement. The employer-business frame does not describe the full population of non-employing or failed ventures. And the descriptive sample's employment and revenue shares are weighted aggregates, not experiments. These are not defects that invalidate the findings; they define the population and question to which the findings apply.
A useful way to summarize the evidence is to keep three statements separate. First, advanced-technology firms and VC-backed firms are disproportionately large in the observed U.S. business population. Second, the joint category is especially large even after adjustment for several observable characteristics. Third, the data do not yet tell us whether the joint advantage arose because VC transformed a technology-using firm, because high-potential firms selected into both choices, or because both forces operated together.
The third statement is the hinge. A regression can show that the gap persists after measured adjustment, but it cannot reveal what would have happened to the same firm under a different financing or technology choice unless the design supplies a credible counterfactual. That is why the timing of adoption and funding matters. If firms were already expanding before the event, a post-event difference may mostly reflect selection. If comparable firms track one another before the event and diverge afterward, a causal interpretation becomes more plausible—though still dependent on assumptions about anticipation, parallel trends, and unobserved shocks.
The descriptive evidence therefore earns a strong but bounded conclusion. It establishes a nontrivial relationship between advanced technology, VC, and firm outcomes in a broad U.S. employer-business dataset. It makes simple “VC alone” or “technology alone” stories inadequate. It also makes selection a central problem rather than a footnote. The next question is not whether the categories differ; they clearly do. It is when the differences emerge, and whether the changes around adoption and funding support a stronger claim than association alone.