Jun 2, 2023 · 24m · 20vc
20VC: Who Wins the AI Race; Startups or Incumbents & Does Having Proprietary Data Really Matter For Startups Today?
compilation · excluded from per-person scoring
Harry Stebbings · 8m spoken Emad Mostaque · 5m spoken Yann LeCun · 4m spoken Vince Hankes · 1m spoken Tomasz Tunguz · 1m spoken Clément Delangue · 46s spoken Sarah Guo · 39s spoken
0:00 / 0:00
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
This 20VC compilation episode brings together top venture capitalists and AI founders to examine whether agile startups or capital-rich incumbents will ultimately control the artificial intelligence market. The speakers analyze key dynamics including foundation model consolidation, startup defensibility, data moats, and execution speed across the evolving AI landscape.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Harry holds 35.4% of the talking time here. How this is scored →
Harry as informed peer 2.5 Guest teaching 5.2 Guest disagreement 2.2 Harry pushing back 1.8
speaking balance: gold is Harry, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 18:15 AI Expert rejects Geoff Hinton's AI dominance premise
The guest forcefully rejects mainstream AI doomer claims made by figures like Geoff Hinton, calling the idea that intelligence naturally seeks dominance a 'gigantic fallacy' and 'nonsense'.
Hardest push from Harry ▶ 3:36 Harry pushes back on application layer value using Tom Tungus dataHarry directly counters Imad's claim about application layer moats by bringing up Tom Tungus's quantitative breakdown showing average EV was far lower due to high company density.
Biggest teaching moment ▶ 16:00 AI Expert explains incumbent risk asymmetry with Galactica and BardThe expert educates the audience on why big tech failed to release ChatGPT first, citing how public backlash destroyed Meta's Galactica and a single error cost Google 8% in market cap.
Harry holds his own ▶ 3:36 Harry cites quantitative market EV metricsHarry demonstrates deep market knowledge by contrasting 3 infrastructure giants against 50 application layer companies to dispute valuation assumptions.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Harry as informed peer | Guest teaching | Guest disagreement | Harry pushing back | Why |
|---|---|---|---|---|---|---|
| Imad Mostak on Foundation Models and Enterprise AI Opportunities | 6 | 6 | 2 | 5 | Harry actively challenges Imad's framing on value accretion by quoting Tom Tungus's analysis regarding average enterprise value across application vs infrastructure layers. Imad agrees with the data point and expands with high-level details on foundation model costs and corporate AGI goals. | |
| Vince Hanks on Incumbent Agility and AI Talent Concentration | 2 | 4 | 2 | 1 | Harry acts primarily as a curator introducing clips from Vince Hanks and Clem Delang. Vince explains talent concentration in incumbents and the threat of fast product shipping against startups. | |
| Tom Tungus on Execution as the Ultimate AI Moat | 3 | 4 | 1 | 2 | Harry sets up the core debate between startup agility and incumbent distribution. Tom Tungus educates the audience on why execution rather than proprietary data serves as the true moat. | |
| Imad Mostak on Data Quality, Open Models, and Vertical Specialization | 2 | 6 | 3 | 2 | Imad details parameter shrinking, model evolution, and the need for clean national data sets. Harry interjects brief clarifying questions while Imad outlines his open-model vertical strategy. | |
| AI Research Expert on Open Ecosystems and Public Risk Aversion | 0 | 7 | 4 | 0 | In an extended monologue from the guest, the AI expert dismantles public risk aversion that crippled Meta's Galactica and strongly rejects the premise that machine intelligence inherently leads to a desire for dominance. Host-side scores are zero due to zero host involvement during the guest speech. | |
| Sarah Guo on Startup Speed and Data Moat Realities | 2 | 4 | 1 | 1 | Harry introduces Sarah Guo, who offers a concise perspective on startup speed vs incumbent data moats before Harry transitions into sponsor reads. |
Statements from this episode (14)
Mostak: Only 5 or 6 Foundation Model Companies Will Exist in 3-5 Years
“I think that there's only going to be five or six foundation model companies in the world in three years, five years.”
Mostak: DeepMind's annual salary budget is $1.2 billion
“DeepMind's salary budget is 1.2 billion a year.”
Mostak: Generative AI will have a bigger economic impact than COVID
“I think this will be a bigger economic impact than COVID.”
Hanks: Top AI talent remains heavily concentrated in big tech
“If you think about where the best talent in AI is right now, I think it's open AI. And then I think everyone would tell you it is Google and Facebook and Microsoft and Amazon. And maybe there are folks that have kind of dripped their way into the startup ecosy…”
Hanks: Betting on startups directly competing with fast-moving incumbents is risky
“If you're going toe-to-toe right now with an incumbent on their home turf and their shipping, I think it's a hard bet to take the opposite side of, you know, in our business.”
Hanks: Startups creating entirely new user experiences will beat incumbents
“But if you're creating something that's totally a different user experience that no incumbent has today, I'd bet on a startup 10 times out of 10 in that case.”
Delangue: Startups training their own models build better products than wrappers
“If you look at models closer to Runway ML or Stability AI or like PhotoRoom in Paris, you see these AI native startups that are actually building, training their own models. And how they can, in my opinion, build much better things than the ones that just use …”
Tunguz: Execution, not data moats, determines startup success in AI
“I think the answer is the one that it's always been, which is better execution is the moat.”
Tunguz: Only one or two startups can compete in foundation models
“I think at the foundational model layer, that's a big boys game or a big girls game just because of the capital intensity required, both for training and the GPU access and all those kinds of things. So maybe there's a startup or two. That's able to raise a co…”
Mostak: No AI models available today will be used in a year
“The reality is no models that are out today will be used in a year.”
Expert: Proprietary LLMs will fall behind open-source models
“The scenario I think will happen, and I'm certainly rooting for, is the scenario I described earlier, where you have some sort of open platform for base LLMs. So base LLMs basically would be seen as a basic infrastructure. TCP, IP, Linux, Apache, completely op…”
Expert: The ecosystem around open LLMs will create net jobs
“There'll be like a whole economy around this, which will create jobs by the way, not make them disappear.”
Expert: Superintelligence does not imply a desire to dominate humans
“There is this idea, somehow, the desire to, and the ability to dominate, is linked with intelligence. A statement that a lot of people are making, including Jeff Hinton recently, That somehow, as soon as a machine becomes intelligent, it becomes uncontrollable…”
Guo: Incumbent data moats are overblown in the AI race
“Much ado has been made about this idea of a data moat, but honestly, there's a lot of data out there, and entrepreneurs are incredibly creative about collecting it, and increasingly about generating it, and I don't think it's, ah, the incumbents are gonna win …”