Capital Moats in Frontier AI

Capital is acting like product in frontier AI, because the labs with the biggest balance sheets can buy more GPUs, run larger training jobs, absorb inference losses, and sign the cloud partnerships that lock in future supply. That makes it harder for new labs to catch up on model quality, price, and distribution all at once, even as application companies increasingly route across multiple models instead of betting on one provider.

• OpenAI, Anthropic, and Cohere had already attracted about $14.5B by October 2023, and each used a different version of that capital moat. OpenAI paired model R&D with ChatGPT and Microsoft distribution, Anthropic aligned with Amazon and Google, and Cohere leaned into enterprise deployments across public cloud, private cloud, and on prem.
  1 sacra 2 sacra 3 sacra 4 sacra
• The moat is not just training spend. It is also the ability to subsidize serving costs while customers ramp, reserve scarce compute ahead of rivals, and fund the humans, tooling, and evaluations needed to improve model behavior. That is why model labs sit upstream of a growing middleware layer that makes switching easier for developers, but does not remove the need for frontier scale.
  1 sacra 5 sacra 6 sacra
• This starts to look like multi cloud. Customers such as Ramp, DuckDuckGo, Scale, and Quora already mix providers, which creates room for several large labs to coexist. But the winners are still likely to be the few labs with enough capital to stay on the quality frontier while also meeting enterprise demands around uptime, deployment flexibility, and pricing.
  1 sacra 2 sacra 3 sacra 4 sacra

Going forward, the strongest labs will keep turning financing into durable infrastructure advantages, then into distribution and product advantages on top. That points to a market with a small number of heavily capitalized foundation model providers underneath, and a much larger layer of apps and tooling companies building on top of them.