Anthropic is in talks with Samsung to develop a custom silicon chip for running its Claude AI models, according to people familiar with the matter. The discussions represent the AI safety startup's push to reduce dependence on Nvidia's dominance in GPU supply and lower inference costs for its large language models.
The timing reflects broader industry momentum. OpenAI unveiled its own custom chip through a Broadcom partnership just days earlier, signaling that leading AI labs now view proprietary silicon as essential infrastructure. Both moves target the same problem: Nvidia's H100 and newer chips command premium pricing, and supply constraints persist despite easing.
For Anthropic, a Samsung partnership makes strategic sense. Samsung manufactures advanced semiconductors and has existing relationships with major cloud providers that could distribute Claude services. Custom chips tailored to Claude's architecture could improve performance on inference workloads, where Anthropic runs billions of tokens daily for paying customers and enterprise clients.
The conversations remain early stage, meaning no deal is finalized. Custom chip development typically takes 18-24 months from design to production, so any Samsung silicon wouldn't arrive until 2026 at earliest. That timeline mirrors OpenAI's expected deployment window for its Broadcom-based chip.
Anthropic has raised $7 billion to date, including a $5 billion Google commitment. The company operates at scale but remains unprofitable as most frontier AI labs burn cash on compute costs. Custom silicon offers a path to better unit economics. Even modest efficiency gains compound across millions of API calls annually.
The broader narrative shows AI incumbents fragmenting Nvidia's grip on the market. Google already manufactures TPUs for internal workloads. Meta built custom accelerators. Now OpenAI and Anthropic are joining that club. None of these chips will replace Nvidia entirely, but collectively they reduce leverage over the AI supply chain.
Samsung's participation suggests