Why big companies usually start with an engineer, and how to build one in 2026

Look at the ten most valuable companies in the world and the pattern is hard to miss. Nvidia was started by three chip engineers. Google by two computer science doctoral students at Stanford. Microsoft by two programmers. Broadcom by an electrical engineering professor at UCLA and the doctoral student he supervised. TSMC by a semiconductor engineer holding a mechanical engineering doctorate from MIT.

So the question is fair, and it splits in two. Is the pattern causal, or is it what a technology boom looks like from the outside? And does it still hold in 2026, now that a language model will write a working first version for anyone who can describe one clearly?

Stanford has the best dataset on the first half. Ilya Strebulaev and the Venture Capital Initiative identified and confirmed 4,975 people behind 1,110 American venture-backed companies that crossed a billion dollars between 1997 and 2021, then analysed 2,791 of them in detail (Stanford Graduate School of Business Venture Capital Initiative, published on Crunchbase News, 13 March 2025). The second half is a market question, and the 2026 funding numbers answer it in a way that is less flattering to the technical side than it first looks.

What follows is the evidence, the mechanism underneath it, the three places the evidence gets misread, and the order I would build in if I were starting a company this year.

Nine of the top ten were built by someone who could build them

Take the current ranking by market capitalisation and set the founding group next to it.

# Ten most valuable companies, and whether anyone
# in the founding group could build the product.
# Market caps: CompaniesMarketCap, retrieved
# 1 Sept 2026. Backgrounds are public record.
#
#   1  Nvidia         $5.28T  three chip engineers
#   2  Apple          $4.60T  one of two, Wozniak
#   3  Alphabet       $4.09T  two CS doctoral students
#   4  Microsoft      $3.79T  two programmers
#   5  Amazon         $2.79T  EE and CS degrees
#   6  TSMC           $2.15T  semiconductor engineer
#   7  SpaceX         $1.87T  physics, self-taught coder
#   8  Broadcom       $1.76T  two electrical engineers
#   9  Saudi Aramco   $1.68T  state oil company, 1933
#  10  Meta           $1.46T  one programmer

One entry does not belong. Saudi Aramco began in 1933 as a concession granted to an American oil company and was nationalised over decades, so it never had a founding group in the sense the others do. Apple needs a footnote too. Steve Jobs did not design the Apple I or write its code. Steve Wozniak did that alone, and Jobs sold it.

Set both aside and the count is this: in every one of the nine that began as a startup, at least one person in the room could produce the product without hiring anybody. Nine data points make a story rather than a finding, so the finding has to come from a bigger sample.

Stanford counted 2,791 of them

The Venture Capital Initiative built its list from Crunchbase, TechCrunch, PitchBook and VentureSource, kept only American venture-backed companies with a confirmed billion-dollar post-money valuation in a primary private round or a billion-dollar exit between 1997 and 2021, and verified each one by hand. That produced 1,110 companies and 4,975 confirmed people. Here is what they were doing before they started.

# Roles held before starting the company.
# Sample: 2,791 people behind 1,110 US
# venture-backed unicorns, 1997 to 2021.
# Source: Stanford GSB Venture Capital
# Initiative, Crunchbase News, 13 Mar 2025.
#
#   research or technology development   25%
#   had already run a company            22%
#   engineer                             21%
#   software engineer                    17%
#   product manager                      14%
#   had run engineering elsewhere         9%
#
# Rows overlap. One person can sit in more
# than one of them.

Degrees point the same way. Computer science is the most common undergraduate major in the sample at more than 500 people, engineering second, economics third. Those three account for 53 percent of the list, and at least 47 different majors appear on it, theology and anthropology included. The group is also heavily credentialed: set against the average American over 25, they are six times more likely to hold a doctorate, three times more likely to hold a master's and twice as likely to have finished an undergraduate degree (Stanford GSB Venture Capital Initiative, Crunchbase News, 24 February 2025).

The detail that makes this more than a gallery of winners is the control. Each of the 1,110 companies was matched against another American venture-backed company that raised its first round in the same year. Without that step, a table of unicorn biographies tells you what the survivors looked like and nothing about the odds. With it, you are comparing companies that started from the same funding position and then diverged.

A quarter came out of research or technology development, 21 percent had worked as engineers, and computer science is the single most common degree in the sample.

The mechanism is the length of the loop

None of that explains why. The usual explanation is talent, and I think it is simpler than talent. It is the number of steps between noticing something is wrong and it being fixed.

If you can build, the loop is: see it, change it, watch what happens. If you cannot, the loop is: see it, describe it to someone, wait, evaluate work you are not qualified to evaluate, describe it again. Ilya Levtov, who runs a supply chain software company called Craft, gave the most honest account of that cost I have read. He has raised $42 million, serves 35 federal agencies and reports double-digit millions in annual recurring revenue, and he says he has never written a line of code. His first developer was a $20 an hour hire on Upwork (Crunchbase News, 27 August 2026).

His own list of what it cost him: a slower start, because someone technical can prototype on nights and weekends while he had to recruit, explain, assess and raise money to pay for it. He hired the wrong engineers at points and lacked the expertise to catch it quickly, which he thinks held the company back.

They've got a direct line between the business concept and the code in which it's executed.

That is his description of the other side's advantage, and it is the mechanism in one sentence. The claim is not that engineers understand markets better. It is that during the first two years, when the only thing that matters is how many times you can be wrong and recover, the technical person can be wrong on a Tuesday and right by Thursday without booking a meeting.

I watch the small version of this every week in the studio I run. The projects that move fastest are the ones where the person deciding what to change is the person typing. As soon as those are two people, a day of work becomes three, and two of those days go on transmitting intent.

Three ways this gets misread

First, as an argument for starting young. It is not. Working from American census records covering firms, workers and owners, Pierre Azoulay, Benjamin Jones, J. Daniel Kim and Javier Miranda found that the mean age at founding for the 1 in 1,000 fastest growing new ventures is 45.0, and that the result holds inside high technology sectors and startup hubs rather than being an artefact of counting dry cleaners (Age and High-Growth Entrepreneurship, American Economic Review: Insights, vol. 2 no. 1, 2020, pages 65 to 82). Their finding, in their words, "strongly reject[s] common hypotheses that emphasize youth as a key trait of successful entrepreneurs".

Second, as an argument that the degree is the asset. The same paper reports that industry-specific prior experience "predicts much greater rates of entrepreneurial success", which is a different claim from knowing how to program. Twenty-two percent of the Stanford sample had already run a company and 25 percent came out of research or technology development. That describes people who spent a decade inside a problem before they sold anything against it.

Third, as an argument that someone who cannot code should not start. Levtov's answer to that is worth more than mine: "It really just takes both. It takes both sides." His read is that the technical side wins the early stage and the commercial side wins the scaling, and the Stanford numbers back the second half more than people expect. Forty percent of the sample had started a company before, and 60 percent of the 1,110 companies had at least one person on the team who had done it before.

One figure cuts against the myth in the other direction. Sixty percent of those companies were their team's first attempt. A second run at it is common, not a prerequisite.

The 2026 money is not the business

Now the part that changes the answer for anyone starting this year. The first half of 2026 took a record $510 billion of venture funding, more than the $440 billion invested across the whole of 2025 and above the previous half-year high of $375 billion set in the second half of 2021 (Crunchbase News, global funding report, 2 July 2026). Read where it went before you read the total.

# Where the record went.
# Global venture funding, H1 and Q2 2026.
# Source: Crunchbase News, 2 Jul 2026.
#
#   H1 2026, all stages          $510B
#   OpenAI and Anthropic         $217B    43% of H1
#   16 rounds above $1B        $108.6B    53% of Q2
#   AI share of Q2 capital        >70%    was <50%
#                                         a year before
#   Q2 seed, worldwide            $12B
#   Q2 seed at $10M and under      $5B

Two companies took 43 percent of everything. All of seed funding, worldwide, for three months, was $12 billion, and the slice arriving in rounds a normal company can raise was $5 billion. A record year and a hostile year for the median new company turn out to be the same year.

Richard de Silva, who runs Lateral Investment Management, draws the conclusion that follows from those numbers. Because code generation and design tools have cut what it costs to build, capital is no longer the limiting factor for most software companies, so most of them should need less risk capital than the previous cohort rather than more (Crunchbase News, 25 August 2026). Atlassian and Basecamp are his examples of the alternative.

Global seed funding for all of Q2 2026 was $12 billion. Two AI labs took $217 billion in six months. Both facts describe the same market.

This is also where the technical advantage compounds. If capital is not the constraint, the constraint becomes how fast you can produce something worth charging for, and that is the axis the person who can build is faster on.

The order I would build in this year

Seven things, in the order the evidence supports rather than the order that feels exciting.

One. Pick a domain you have already been inside for years, meaning a process you have watched go wrong from a desk rather than a market you read about. Prior experience in the same narrow industry is the strongest predictor in the census data, and it is the one part of the profile nobody acquires in a weekend.

Two. Be the person who can ship version one, or sit next to that person from the first week. A contractor, an agency or an advisor holding half a percent all sit outside the loop by definition. If you cannot build it and cannot recruit someone who will build it beside you, that loop becomes your permanent operating speed.

Three. Charge before you raise. A customer paying $200 a month tells you something no seed round can, and in a market where two labs absorbed 43 percent of global funding, revenue is the only capital most companies are going to see.

Four. Work somewhere that produces them, if that option is open to you. The Stanford sample had worked at 6,109 different organisations, and only 33 of those produced 15 or more people who went on to build a billion-dollar company. Google produced 96, Microsoft 64, IBM 42. Stanford, MIT and Harvard produced 43, 40 and 33 as employers rather than as schools. Israel Defense Forces alumni came out 3.1 times more likely than average. Two years inside one of those beats an accelerator application.

Five. Pick a buyer with a budget line. Craft did not become a serious company on the strength of its data. It became one when the United States Air Force needed 300,000 companies in the defence industrial base tracked and signed a five-year, $6.5 million contract 94 days after the first conversation. One person with a budget and a deadline beats a large addressable market on a slide.

Six. Plan for the possibility that this is attempt one of two. Forty percent of the Stanford sample had started something before. Structure the first attempt so that failing does not cost you the ability to try again, which mostly means not borrowing money you cannot repay and not signing anything that outlives the company.

Seven. Put it in front of one person every week. This is the only item here I have tested myself rather than read, and it is the one that decides whether the other six matter.

Six and a half years and about five rounds

The last thing worth calibrating is the clock. Across more than 1,500 American companies that crossed a billion dollars between 1997 and 2024, the average time from founding to that valuation was 6.6 years (Stanford GSB Venture Capital Initiative, Crunchbase News, 11 February 2025). A company started this week gets there, on average, in early 2033.

# Time and rounds to a billion-dollar valuation.
# Sample: 1,516 US companies, 1997 to 2024.
# Source: Stanford GSB Venture Capital
# Initiative, Crunchbase News, 11 Feb 2025.
#
#   average years from founding      6.6
#   typical number of rounds          ~5
#   companies needing under 2        125
#   companies needing 8 or more      331
#   most rounds any one company       18
#
#   fastest   Anthropic, xAI    about 1 year
#   slowest   Keyfactor         22 years
#             Kiteworks         25 years

The densest window is years four and five after founding, with more than 100 companies crossing in each of the peak years. Of the nearly 1,000 that stayed private the whole way, 453 raised their first venture round within months of starting and 281 more by the end of year two, so the early money mostly went to companies that were already moving.

The rate is rising. 250 companies joined the billion-dollar list in 2026 through 15 August, against 193 across the whole of 2025, with 56 percent headquartered in the United States and 19 percent in China (Crunchbase News, 19 August 2026). Y Combinator was the third most active investor behind Sequoia and Khosla, and the only accelerator in the top ten.

So the honest answer to the first question is that a technical background is not a credential, it is a shorter feedback loop, and for the first two years the loop is the whole company. The honest answer to the second is that 2026 hands that loop to more people than any year before it and hands the money to almost nobody. Pick a problem you have watched go wrong, build the first version yourself, charge somebody for it, and give it six years.