Open-source gains force Anthropic to face IPO valuation cuts

Open-source artificial intelligence models are devouring proprietary market share, threatening to derail the multi-billion-dollar IPO plans of frontier labs like Anthropic.

The shift began with sharp defensive moves from closed providers. On This Week in Startups on September 23, 2026, venture capitalist Jeff Clavier described recent 50 percent token price cuts by OpenAI and Anthropic as a venture-subsidized race to the bottom. Clavier noted that frontier labs spend billions training massive base models only to sell raw compute at razor-thin margins, while open-source alternatives rapidly erase the performance gap.

Everywhere VC managing partner Jenny Fielding observed that early-stage founders care primarily about unit economics and independence, leading portfolio founders to systematically abandon closed APIs. Relying on proprietary models exposes startups to sudden platform changes or direct competition from the labs themselves. As hardware costs rise, value is fleeing the middle layer and shifting to open infrastructure and chipmakers.

Three days later, on September 26, 2026, the All-In podcast detailed how dramatic this migration has become. David Friedberg cited data from Vercel showing open-source models surged from 20 percent to 80 percent of total token volume in just twelve weeks. Releases from open-weight developers like DeepSeek and Alibaba now match last year’s top proprietary systems, running locally on Mac Studios and consumer hardware for a fraction of hosted API costs.

Chamath Palihapitiya argued on All-In that because base model tokens have become interchangeable commodities, closed labs will be forced to build vertically into domain software or face severe valuation write-downs. Palihapitiya predicted that Anthropic will have to accept a lower valuation during its upcoming IPO as institutional buyers demand a wider safety margin.

"Basic intelligence is now free software."

- Chamath Palihapitiya, All-In

David Sacks added on All-In that Anthropic's messaging actively undermines its public listing. Chief executive Dario Amodei publicly warns of catastrophic biological threats while simultaneously building wet labs and raising venture capital. Sacks noted that public market investors hesitate when founders demand super-voting control alongside existential risk disclosures, especially as federal authorities reject requests for regulatory liability shields.

By September 28, 2026, the enterprise fallout reached a tipping point on TFTC. Nous Research strategist Tommy Eastman reported that corporate legal teams and financial institutions are panicking over data sovereignty. Sending proprietary intellectual property into external black-box models operated by OpenAI or Anthropic risks breaching fiduciary duties, driving enterprises toward self-hosted open architectures.

Eastman explained on TFTC that open-source models paired with execution harnesses like Hermes Agent now handle 90 percent of routine corporate tasks at a fraction of API costs. Show host Marty Bent noted that local agentic harnesses maintain self-referential memory and strict data isolation, turning raw models into predictable automated labor without leaking operational secrets to closed providers.

"Model-agnostic harnesses convert raw AI outputs into persistent operational labor."

- Tommy Eastman, TFTC

This enterprise migration to open weights has inverted hardware dynamics. Eastman pointed out that sales reps at compute providers are turning away corporate buyers as local inference demand consumes available GPU clusters. High debt loads, squeezed data center capacity, and collapsing token prices leave closed labs holding expensive, commoditized infrastructure while open-source software takes over the enterprise stack.