AI Boom Roadkill and Seed VC Math | David Frankel on 20VC

On 20VC with Harry Stebbings, seed investor David Frankel of Founder Collective frames the current AI surge as the biggest tech wave of his 18‑year career, but also one that will leave a trail of failures. He tells Stebbings that if he lines up the internet, SaaS, mobile, and AI, “nothing looks the same” next to AI in terms of scale and hype.

According to Frankel (and dated to the show’s August 2026 release), the numbers from the last 25 years already show how narrow true mega‑success is: among the 500 top company creations he studied, fewer than 100 ended up sustainably above a $10 billion valuation. He expects the AI cycle to be similar or more extreme, with a tiny number of platform‑level winners and a vast majority of disappointed investors.

Frankel repeatedly uses the phrase “Hollywood” economics to describe this era: a few superstar outcomes and a long tail of roadkill. Seed, in his view, stays interesting because you don’t need to hit the absolute #1 AI winner; owning a slice of the broader set of multi‑billion‑dollar outcomes can still return a fund.

Frankel’s central thesis on 20VC is that seed investing remains economically compelling in the AI era, but it has become crowded, commoditized, and structurally harder. He tells Stebbings that for the top 500 companies formed over 25 years, the median outcome was around $2.6 billion. In his math, if a seed fund owns roughly 5% of such a company, that single position can return an entire fund.

At the same time, Frankel argues the venture “pyramid” has intensified at the top. LPs and fund‑of‑funds increasingly judge managers on whether they were in a handful of multi‑trillion‑dollar companies. He says that missing those outliers makes asset‑gathering much harder, pushing larger firms to chase access and write “call‑option” checks across many hot AI stories.

Against that backdrop, Frankel positions Founder Collective as a small, return‑focused seed fund that refuses to scale into multi‑billion‑dollar vehicles. He says the firm’s GPs are the largest LPs in their own funds and are “greedy for returns, not management fees,” which he claims keeps them disciplined on fund size and follow‑on strategy even as AI valuations spike.

Frankel walks Stebbings through the specific data and deal structures he thinks characterize the current environment. He cites internal work looking at the top 500 company creations over 25 years, finding:

• Fewer than 100 ended up sustainably above a $10 billion valuation.
• The median outcome across those 500 was about $2.6 billion.

He builds on this to argue that a seed investor owning around 5% of one such median winner can fully return a fund, and even a 5% stake in a $500 million exit is “incredible” for a small, disciplined vehicle.

On the ground, Frankel says he sees:

• Seed rounds of $3–4 million in total capital still available, even amid headline $8–10 million “large seeds.”
• Caps around $20 million for some AI‑inflected vertical SaaS founders he meets (for example, people coming out of long SAP consulting careers and hacking with Claude‑based code).
• Uncapped notes at the seed stage, which he bluntly says “suck” for early investors and often back into very high implied entry prices.

He also stresses that many of the “hot hot AI” companies raising very large early rounds show little evidence of capital efficiency yet, and he believes “the jury’s out” on whether those capital‑intensive strategies will work.

Frankel repeatedly warns that AI euphoria masks significant risks for both investors and founders. He agrees with Stebbings that another crash is “definitely” coming, though he stresses that timing is unknowable.

He highlights several structural dangers:

• Mega‑fund call options: Frankel says many large platforms treat seed as cheap options. A junior principal might lead a high‑priced round, then leave the firm, at which point the partnership shifts focus to “real winners” and quietly stops supporting mid‑tier outcomes worth a few billion. In his view, this strands founders whose growth is slower or more linear than AI hype implied.
• Pro‑rata rights: Frankel calls broad pro‑rata a “call option against” founders. He’s seeing new rounds where only the lead or largest shareholder has pro‑rata, not all major investors. He believes that, while his fund wants the right to add slightly more, universal pro‑rata can be unhealthy in aggregate.
• Valuation shorthand: He admits that rigid valuation “frameworks” have at times saved his partnership, but also led them to miss large successes. He cites turning down Klaviyo because a pre‑set pricing framework made it too easy to say no, which he now describes as a “terrible mistake.”

Overall, he expects most AI startups to become “roadkill” despite this being, in his words, “the wave of our lives.”

Looking forward, Frankel tells 20VC listeners to watch several signals rather than any single stock or ticker. In consumer AI, he says he is surprised how little daily behavior has changed so far beyond him “typing much less” and speaking more into devices. Aside from examples like Suno, he believes many consumer categories remain “not yet played out,” and he views the shift from AI creation tools to true consumption platforms as a key milestone to track.

On capital markets, he urges investors to monitor:

• How often seed‑stage “orphan” companies emerge when big funds move on and label rounds as “seed plus” or “seed extensions.” He thinks those abandoned but still‑growing companies can represent attractive entry points for certain strategies.
• Whether uncapped or very high‑cap AI notes continue, or whether LP pressure and performance data force terms back toward more traditional pricing.

In the real‑economy impact of AI, Frankel says he is most excited about what compute‑heavy approaches could mean for chronic disease and cancer treatment. He calls current chemo “prehistoric” and predicts that AI‑driven discovery and simulation, combined with wider compute deployment in healthcare, could change millions of lives. He links this to broader automation trends, speculating that in 5–10 years his and Stebbings’ children may not need to learn to drive because autonomous systems will be mainstream.