AI Made Seed Startups Cheaper to Build and Harder to Fund

The typical U.S. seed startup that closed a round in the first half of 2025 had 6.2 employees, down from 11.1 in 2021, Carta found. AI tools let a founder ship a working product in a weekend. That speed has flooded investors with finished-looking companies and fewer ways to tell which ones are real.

Aaron Tainter, director of accelerator programs at Innovation Works in Pittsburgh, argues that AI-native fluency is now table stakes and that founder-market fit is the scarce resource. The market around that claim is harsher than a cheap build suggests. North American venture totals hit records in 2026. Ordinary seed checks did not.

44% Fewer People on the Typical Seed Team

Carta’s compensation researchers, Kevin Dowd and Peter Walker, published the headcount collapse in their H1 2025 report. The typical seed team has 6.2 employees, 44% smaller than the 11.1-person average in 2021. Later stages shrank too, just less: Series A teams were 37% smaller than in 2021, Series B teams 19% smaller. Series C headcount was 2% larger.

SEED HEADCOUNT ON CARTA

Carta ties the early-stage shrink to AI. Founders are running leaner on purpose, using copilots and APIs to do work that used to need a bench of engineers. Hiring across the Carta universe keeps slowing. Startups added 28,299 people in January 2025, 17% below January 2024 and 62% below the 73,761 hires logged in January 2022.

The mix of those hires has shifted toward people who ship and people who sell. Engineering was 29.7% of new hires in the first half of 2025. Sales was 16.6%. Operations fell to 5.2%, from 12.4% in 2019. Product pay caught engineering: both functions averaged $189,000 for new hires as of June 2025.

Record Venture Dollars, Fewer Ordinary Seed Checks

Headline venture numbers in 2026 look nothing like a drought. Crunchbase data put North American seed-through-growth investment at $392 billion in the first half of the year. The second quarter alone was $137.2 billion, the second-largest quarter on that books, trailing only the first quarter’s blowout.

Almost none of that is a broad seed boom. About $4.9 billion went to North American seed and angel rounds in the second quarter, down 15% from the prior quarter and down 27% from a year earlier. Round counts dropped too. Five or more companies still closed seed or angel rounds of $100 million or more, including a $200 million round for Mirendil, a foundational AI research shop. Strip those outliers and the traditional seed market is thinner than the headline.

THE Q2 2026 SEED SQUEEZE

• North American total: $137.2 billion in the second quarter, inside a $392 billion first half.
• Seed and angel: about $4.9 billion, down 15% from the prior quarter and 27% from a year earlier.
• Early stage: just over $31 billion, nearly double a year earlier, with deal count at a five-quarter low.
• AI share: about 80% of investment across stages, with a $12 billion Prometheus round more than 40% of the early-stage total.

Anthropic raised $65 billion at a $965 billion post-money valuation. Anduril Industries raised $5 billion. Those are not seed outcomes, and they set the mood for everyone else. The remaining early checks keep clustering around companies that already look like category winners in AI, not around every team that can stand up a site in an afternoon.

Getting from seed to Series A got harder on a longer clock. Incisive Ventures, working from Carta cap tables in June 2025, found that 15.4% of companies that raised seed in the first quarter of 2022 reached a Series A within two years, compared with 30.6% of the first-quarter 2018 cohort. Tainter’s version is the same squeeze in plainer language: fewer seed-funded startups are graduating, and investors are concentrating capital in fewer bets.

Companies that raised before ChatGPT already felt a version of this split, a funding drought for pre-ChatGPT startups while newer AI names absorbed the oxygen. The 2026 wrinkle is that even the new names can look finished on day one.

A Weekend Demo Is Now Easy to Fake

Tainter calls the junk “startup slop,” the entrepreneurial cousin of empty AI copy. Software founders can fabricate a site, a deck, and a story of traction in an afternoon. Dealflow volume, he wrote, has become a vanity metric. The product is no longer the moat because everyone can build.

Founders who haven’t embraced these tools in their daily operations aren’t even at the table. They’re new-aged dinosaurs.

Aaron Tainter, director of accelerator programs, Innovation Works

That is the trap, not the pitch. If a demo takes a weekend, a demo stops being evidence. Y Combinator cofounder Paul Graham argued in August 2025 that vibe-coded apps are making money, after talking with an infrastructure founder who could see those apps in the wild. Revenue can be real. It still does not tell an investor why this team, in this market, will still be standing when the next weekend builder shows up with the same stack.

A clone of Instagram is the cartoon version of the same gap. Months of “anyone can ship” talk have not produced a vibe-coded photo network with a real audience, because Instagram was never a weekend of interface work. It was distribution, habits, and time. Software that looks done on Friday morning is now the easy part. Owning a market is not.

Some of the fake signal is financial, not visual. Scott Stevenson, cofounder and chief executive of the legal AI firm Spellbook, has called out AI startups for dressing contracted or unimplemented revenue as ARR. Several founders and finance operators told him the habit is common, and that some funds play along because a rocket-ship number helps the story. When the spreadsheet can be massaged as easily as the landing page, investors have to ask questions the deck cannot answer.

Why Hardware Still Survives a Diligence Call

Deep tech is drawing capital because a therapeutics pipeline, a plant, or a robot cannot be vibe-coded by lunch, which is a diligence fact as much as a science one. Walker’s Carta note on 2025 fundraising, built from more than 900 U.S. rounds in energy, hardware, semiconductors, and biotech, put deep tech’s 36% share of funding to Carta companies, up from 17% a decade earlier. He wrote that 2025 was a year of startups building in the physical world, and that 2026 would supercharge it.

Maybe some of that is the idea that moats in software are declining in the age of AI.

Peter Walker, head of Insights, Carta

Hardware is the exception on Carta’s headcount charts too. The average hardware startup that raised Series A in the first half of 2025 had 29.8 employees, 116% more than five years earlier. In that same half, hardware firms hired 1.3 people for every departure, the highest ratio of any major sector Carta tracked, ahead of medical devices at 1.2. Education ran the other way, at 0.6 hires per departure. Consumer, pharma/biotech, and gaming also lost net headcount.

Tainter’s sector map matches the data. A therapeutics company still needs real science, real key opinion leaders, and real partnerships. Hardware and advanced manufacturing have the same kind of friction. That friction is the filter. It is also why feature moats that AI made cheap look thinner next to a lab notebook, a bill of materials, or a hospital contract.

Not every physical category is winning. Walker flagged medical devices as still struggling in 2025, with IoT flat. The bet is not “atoms beat bits.” The bet is that whatever cannot be faked in an afternoon still has a way through a partner meeting.

AlphaLab Still Puts Founders in a Room

Innovation Works has been running AlphaLab in Pittsburgh since 2008, among the first ten U.S. accelerators. The program still writes up to $100,000 into each company, still runs six months, and still insists on an in-person cohort. In August 2026 AlphaLab said alumni had produced more than $1.4 billion in follow-on funding from more than 300 companies, with over 20 exits and two unicorns. The current class is 21 startups, its largest.

Tainter’s screening trick is local on purpose. AlphaLab asks why Pittsburgh is the right place to grow. It asks how customer discovery actually happened. It asks why the company exists. The point is not the city’s brand. The point is whether the answer sounds lived-in. AI can write a Pittsburgh paragraph. It is worse at surviving a follow-up from someone who has sat in that room for years.

HOW ALPHALAB TIGHTENED THE FILTER

1. 2008: Innovation Works launches AlphaLab as its first accelerator, later a founding charter member of the Global Accelerator Network.
2. January 28, 2025: Software, hardware, robotics, and life-science tracks unify under one curriculum built on customers, capital, and community.
3. February 10, 2026: AlphaLab names its 2026 class, later counted at 21 companies, and Tainter describes founders embedding intelligence into health care, energy, robotics, and industrial operations.
4. April 20, 2026: Applications open for the 2027 cohort, with a sharper push for AI-native founders and a new neuroscience subtrack with Allegheny Health Network’s Neuroscience Institute.
5. July 17, 2026: The 2027 application window closes.
6. August 18, 2026: AlphaLab says a national ranking of U.S. incubators and accelerators included it, the only Pennsylvania program in that top group as Innovation Works described it.

On the April 20, 2026 call for 2027 applications, Tainter said early-stage founders now face a higher bar than in previous years, in how they build and how they execute, and that selection has gotten more competitive as volume and quality both rise. Innovation Works, whose seed fund dates to 1999, says it has backed over 800 companies that went on to raise $3.74 billion and to generate or keep more than 20,000 Pennsylvania jobs.

The in-person rule is the product now. Plenty of accelerators moved remote and shipped recorded content. AlphaLab kept the room because the room is where slop breaks down. A founder who cannot explain why this city, this customer, and this problem belong together is not going to be saved by a prettier prototype.

What a Lean Seed Team Looks Like Now

With a 6.2-person seed company, every hire has to cover ground that used to belong to a department. Tainter’s list of early additions is short, and a bench of engineers no longer sits at the top of it. The product-minded builder still matters, because someone has to ship fast with AI tools. The other two seats are commercial: a person who owns the customer relationship and early revenue, and a person who can position the product and create demand.

THE THREE SEATS THAT STILL GET FUNDED

• The builder: A product-minded hire who can ship with copilots, APIs, and low-code tools at a pace that used to need a full engineering team.
• The customer owner: Someone with domain relationships who can run discovery and turn it into early revenue, not a waitlist screenshot.
• The demand lead: A person who can place the product in a market and generate pull, because a weekend website is not distribution.

That roster is a consequence of the 44% headcount cut. When the average seed company shed engineers, it did not shed the need for someone who has already lived the buyer’s problem. Tainter’s line is that AI can help a founder build anything, and that customers are what tell them what is worth building. Judgment, industry knowledge, and a clear view of what people will pay for are the scarce inputs. Code is not.

Coachability, hustle, and conviction are the soft labels he still uses. The hard version is documentary. Applications that show the founder has actually lived the problem survive. Applications that read like a market memo written last night do not. “Investors don’t want to write a check to someone who has vibe-coded a company they aren’t passionate about, and the tells are easier to spot than founders think,” he wrote.

The Inbox Test Investors Started Using

Because the technical chores got cheaper, Tainter wants founders to spend the leftover hours on judgment, creativity, storytelling, and relationships. The tell he watches is speed. There is no longer any excuse, he wrote, for taking four days to answer an email, skipping a weekly investor update, or failing to follow up after a meeting. AI drafted the update. A slow reply is now a statement about how the company will be run.

That is a brutal standard, and it is also a rational one in a market where the deck, the demo, and even the customer quotes can be generated. Soft interactions are what remain when the old artifacts got cheap. Investors are weighting those interactions more, not because manners suddenly became a thesis, but because manners are harder to batch-produce than a landing page.

The cost of building fell. The cost of being believed went up. AlphaLab’s 2027 cohort, already selected on paper after the July 17, 2026 deadline, will walk into that trade with a neuroscience track tied to a hospital system. Labs, customers, and a room in Pittsburgh are the parts a chatbot still cannot finish overnight.

Disclaimer: This article is news reporting and analysis of venture-market data, headcount figures, and accelerator practices. It is informational only and is not investment advice, a solicitation to buy or sell any security, or a recommendation to fund, join, or start any company. Readers who are considering startup investing, a fundraising process, or employment at an early-stage firm should consult a qualified investment adviser, securities lawyer, or career counselor before acting. Headcount, funding, valuation, and program figures reflect Carta, Crunchbase, Incisive Ventures, and Innovation Works materials as of the dates on those publications and can change as later deals are logged.