Anthropic has entered early-stage discussions with Samsung Foundry to explore the manufacturing of custom AI chips on a 2-nanometer process. While the company has declined to confirm the report, the move signals a calculated effort to decouple its compute infrastructure from its current reliance on Google’s Tensor Processing Units (TPUs).
The timing of these discussions is inseparable from Anthropic’s confidential filing for a U.S. IPO on June 1, 2026. For prospective investors, the ability to control the hardware stack is a primary metric for assessing long-term valuation and operational risk. By moving toward proprietary or semi-custom silicon, Anthropic is attempting to mitigate the supply chain vulnerabilities inherent in relying on a single cloud provider’s ecosystem.
In June 2026, the lab hired Clive Chan, formerly OpenAI’s second-ever hardware engineer on its custom chip program. Chan’s title — ‘perplexity per picojoule’ — explicitly frames the lab’s objective: optimizing inference efficiency at the physical level.
Anthropic’s interest in Samsung Foundry places it within a broader, high-stakes pipeline. Samsung’s 2nm process is currently being courted by Tesla, which holds a $16.5 billion contract for AI6 chips, as well as Meta for its third-generation MTIA and AMD for its SF2P process. The competition for foundry space is becoming a structural bottleneck.
The industry-wide pivot toward custom silicon is a direct response to Nvidia’s 85% market dominance. Every major AI lab is now engaged in a race to escape this dependency. OpenAI is collaborating with Broadcom for 2027, Meta is utilizing Broadcom for its MTIA, Google continues to iterate on its TPU program, and both Microsoft and Amazon are advancing their respective Maia and Trainium initiatives.
For Anthropic, the transition from Google’s TPU infrastructure to a proprietary model represents a multi-year capital commitment. By positioning itself within Samsung’s 2nm pipeline, Anthropic is signaling to the public markets that it intends to control its own compute destiny.