Are claims about the massive capital cost required to compete in AI a genuine barrier to entry, or a self-serving incumbent narrative?

Sarah Tavel · May 6, 2024

Training frontier models will keep getting more expensive even as underlying chip costs and research efficiency improve

Demand for compute grows faster than efficiency gains — like adding highway lanes and inducing more cars — and the race requires ever more specialized chips and power investment, with power becoming a real constraint for training and inference

33:19 20VC: Benchmark's Sarah Tavel on Are Foundation Models Commoditising | Why Frontier Models Will Be Closed Source | Why the Value is in the Application Layer | The Future of AI is "Selling the Work" Not the Tools

Eiso Kant · Oct 7, 2024

The $600M Poolside has raised is enough to be a credible entrant in the frontier race today but will not be enough over time

10,000 GPUs brought online this summer let them use reinforcement learning from code execution feedback to generate very large amounts of data and train very large models, but compute needs grow

Scope: enough for this moment in time; not enough over the longer term

25:13 20VC: Raising $500M To Compete in the Race for AGI | Will Scaling Laws Continue: Is Access to Compute Everything | Will Nvidia Continue To Dominate | The Biggest Bottlenecks in the Race for AGI with Eiso Kant, CTO @ Poolside

Arthur Mensch · Apr 29, 2024

Barriers to building a relevant foundation model company are not falling: you need enough capital for compute, scarce model-training talent, and a brand

Being relevant requires dominating the cost-efficiency/performance frontier; if your model is strictly dominated by another you have a problem, and only a few companies are well positioned

21:56 Capital talent and brand barriers to competitive foundation model building remain high and are not falling

Harry Stebbings · Aug 19, 2024 · hedged

The cost of competing in frontier AI is exorbitant and almost unlike anything seen before in technology.

OpenAI is reportedly spending on the order of $3B a year, which seems unprecedented for a technology cycle.

10:09 Current frontier spending scale is genuinely unprecedented in technology history

Cem Sertoglu · Nov 20, 2024

Larry Ellison's claim that it costs $100BN to enter the AI race is self-serving, because that price level would leave only about four players in the game

An incumbent benefits from a narrative that makes the entry cost prohibitive for everyone else

56:03 Hundred billion dollar entry cost claims are self serving incumbent narratives

Sarah Tavel · May 6, 2024 · hedged

The only defensible rationale for very large early rounds like Cognition's is that giving a talented team enough capital to buy GPUs and train their own model is a self-fulfilling prophecy that creates a capital-scale moat; absent that, it is just VC FOMO capital deployment, which does not lead to good outcomes

The capital required to even be on the field is itself a rare moat if the money is actually invested that way; otherwise VCs are repeating a familiar FOMO pattern

25:51 Self fulfilling capital scale moat justifies massive early rounds if actually invested in training

Eiso Kant · Oct 7, 2024 · hedged

$100B is roughly the entry price only for becoming an infrastructure hyperscaler, not for competing in the AI capabilities race

Cloud companies' CapEx over a few years already runs far above $100B if you want global data centers serving models to everyone

28:07 100b entry cost applies to becoming hyperscaler infrastructure not ai capabilities race