What is the primary source of durable value in a foundation model company: the model, the data, or the talent?

Cristóbal Valenzuela · Aug 28, 2023

Models are not a moat and eventually don't matter; what matters is the people building them and how fast you can learn from and change them

State of the art in language and images has been completely overturned every few years, so the durable advantage is learning speed — getting models into people's hands, understanding the value users get, and pushing training and research boundaries

17:31 20VC: Why AI Models are not a Moat, Where Does the Value in AI Accrue; Startups or Incumbents, What the World Has Got Wrong About AI, Why AI Needs a New Story and Who is the Right People to Tell it with Cris Valenzuela, Co-Founder & CEO @ Runway

Aravind Srinivas · Jun 5, 2024

The value of OpenAI and Anthropic lies not in the models they have but in the machine that builds the machine — the specific group of people with the tacit knowledge to train frontier models plus their accumulated compute

Valuation should be judged by how hard it is to reassemble the whole thing, not just the output; the same people will produce the next model

25:22 20VC: Perplexity's Aravind Srinivas on Will Foundation Models Commoditise, Diminishing Returns in Model Performance, OpenAI vs Anthropic: Who Wins & Why the Next Breakthrough in Model Performance will be in Reasoning

Reid Hoffman · Jun 10, 2024

Model quality really does matter, and only a small number of teams currently know how to train models well

Training is not just data plus compute plus a button — there is art and science in the sequencing of training regimes, much of which is not written down in published papers

Scope: the set of capable teams will broaden over time

18:57 20VC: Reid Hoffman on Foundation Models: Who Wins & How Do Incumbents Respond | The Inflection AI Deal: How it Went Down | Why Trump is a Threat to Democracy | The Future of TikTok | Lessons from Sam Altman, Brian Chesky and the OpenAI Board

Harry Stebbings · Jun 19, 2024

The value in foundation model companies like OpenAI is not the model but the specific collection of people with the skills to build it, which is what justifies their valuations

Models depreciate, but the assembled teams are the scarce asset

7:18 20VC: Foundation Models are the Fastest Depreciating Asset in History, Lina Kahn is a Threat to American Capitalism, PE is Not Coming to Save the M&A Market & How China Could Overtake the US in the AI Race with Michael Eisenberg

Christian Kleinerman · Sep 22, 2023

Over 90% of the value in AI sits in the data rather than the model

Significant IP went into building models but that is becoming less differentiating while data grows in importance — people are even doing the math on running out of public data to improve models

16:50 20VC: Are Foundation Models Becoming Commoditised? Do OpenAI and Anthropic of the World Have a Sustaining Moat? Why Smaller Models May Work Better? Why Incumbents with Data Power Win the AI War with Christian Kleinerman, SVP Product @ Snowflake

Alexandr Wang · Jun 12, 2024

Of the three pillars of frontier model competition — algorithms, compute and data — data is the only one where a durable, sustainable moat can be built

Algorithmic IP eventually diffuses to the rest of the industry and compute can simply be bought with money, whereas differentiated data access can persist

19:24 20VC: Scale's Alex Wang on Why Data Not Compute is the Bottleneck to Foundation Model Performance, Why AI is the Greatest Military Asset Ever, Is China Really Two Years Behind the US in AI and Why the CCPs Industrial Approach is Better than Anyone Else's

David Cahn · Aug 5, 2024 · hedged

The data centre is the most important asset in AI, and increasingly a model is just a data centre.

As models get bigger, scaling laws dominate and researchers move freely between labs, so there is little remaining differentiation between the models companies launch — leaving the physical build as the differentiator.

16:46 The data center is the real source of value not the model data or talent

Michael Eisenberg · Jun 19, 2024

A company whose value rests on its talent rather than a lasting asset is a bad investment, and neither the data nor the models of foundation model companies constitute a lasting asset

Talent can walk out the door at night and is not owned — as Bill Gates' father asked, why invest in a business where the assets walk out at night; Microsoft's software was a durable asset whereas data and models are not

7:46 Neither data nor models are durable assets making talent dependent ai labs bad investments

Michael Eisenberg · Jun 19, 2024 · hedged

It is not clear how AI model companies build a moat; data is the most plausible candidate, but the models are visibly indistinguishable to consumers today

Users cannot visibly tell the difference between OpenAI, Anthropic and other models, unlike Google which built a moat by being far fastest plus a reinforcing ad business

39:26 Moat unclear data most plausible models indistinguishable to consumers today

Christian Kleinerman · Sep 22, 2023

Foundation models are becoming commoditized, and value will only shift back to the model layer when the next big innovation arrives

Teams of seven people have built models comparable to OpenAI's or Anthropic's for some use cases, so at the end of the day it's a data problem rather than a model problem

0:00 Value cycles between model and data layers with each innovation wave

Aravind Srinivas · Jun 5, 2024

Microsoft's cash flow advantage cannot buy it a frontier lab, because poaching individual scientists doesn't work — you need the whole team

The best people want to work with other best people, so hiring ten OpenAI researchers and handing them GPUs directly isn't happening; cash flow doesn't remove the dependence on the intact team

28:36 Team cohesion not individually poachable scientists is the asset so cash rich rivals cannot buy their way in