Anthropic is negotiating with Samsung Electronics to design and manufacture a custom AI chip, according to reports. The company has already hired chip engineers for the project, which remains in early stages. Anthropic emphasized that Nvidia remains a key part of its infrastructure strategy, signaling the custom chip effort complements rather than replaces current GPU partnerships.
The move positions Anthropic alongside other AI labs pursuing vertical integration. OpenAI recently unveiled "Jalapeño," its own chip initiative designed to reduce reliance on third-party accelerators and lower training costs. Meta has similarly developed custom silicon for its AI workloads. For companies burning through billions on compute, manufacturing proprietary chips offers control over specifications, supply chains, and long-term economics.
Samsung's involvement matters. The chipmaker has extensive manufacturing expertise through its foundry business, competing with TSMC on advanced process nodes. Partnering with Samsung rather than designing alone gives Anthropic access to proven production capabilities. Samsung has previously worked with other tech companies on custom silicon projects.
The timing reflects broader industry pressure. As AI models grow larger and more expensive to train, companies face hard choices: depend on Nvidia's H100 and H200 GPUs and accept supply constraints plus vendor lock-in, or invest heavily in custom silicon with uncertain timelines and higher upfront costs. Neither path is simple. Building competitive chips takes years and billions in R&D. Nvidia's dominance in AI compute persists because its software ecosystem, driver support, and performance-per-watt advantages remain difficult to match.
Anthropic's public insistence that Nvidia "still matters" suggests it is not betting entirely on Samsung. This is pragmatic. Custom chips typically optimize for specific workloads, training, inference, or both. Heterogeneous infrastructure using both Nvidia GPUs and custom accelerators for different tasks may prove more efficient than either alone