Who is Chamath Palihapitiya?
He's is a Sri Lankan-born venture capitalist who founded Social Capital and co-hosts the All-In podcast alongside Jason Calacanis, David Sacks and David Friedberg.
His views carry weight in AI circles because he sits at the intersection of early-stage investing and public markets commentary, giving him visibility into both start-up strategy and how institutional investors are pricing the sector.
From brains to bodies
Palihapitiya argues that AI development has moved through two distinct phases, and is now entering a third.
The first phase, he says, was about building the "brain": large language models capable of answering questions but little else.
The second phase added "harnesses" and agents, giving that brain eyes, hands, a notebook for memory and a keyboard, effectively turning it into something that could act autonomously.
Why phase two falls short
Palihapitiya says autonomous agents, as they currently exist, are not yet good enough for enterprise use, which is why he believes a third phase is now beginning.
This depends on making agents more informed, giving them the specific context needed to do a job properly rather than answer generically.
He argues this means training AI for defined roles, such as lawyers, customer service staff or sales agents, all of which require access to large volumes of contextual data.
Systems of record advantage
Palihapitiya's central claim is that companies holding "large systems of record", the databases that already contain a business's core operational data, are unusually well placed in this next phase, provided they execute correctly.
He points to Salesforce, the customer relationship management software company, as an example of a business navigating that shift successfully, through integrating with existing AI models, supporting agent harnesses and building its own proprietary models.
Palihapitiya credits that approach for Salesforce's strong net dollar retention, revenue growth and upgraded guidance, alongside a 43% rise in its share price since he called the stock's bottom.
Data, not model spect
The argument positions incumbents with deep operational data as potential winners over AI labs that lack direct access to enterprise systems, a view that runs counter to the assumption that model quality alone will determine who captures value from AI.
