Anthropic has built its reputation on measured, safety-first rhetoric—the AI company that would rather slow down than move fast and break things. Yet behind the careful public posture, the company is making a decidedly aggressive infrastructure play: it is in discussions with Samsung to develop a custom AI chip, according to people familiar with the talks.
The move marks a strategic inflection point not just for Anthropic but for the broader AI industry. Until now, the company has been content to rent compute from cloud providers and purchase Nvidia's H100s and B200s like everyone else. Custom silicon is a different game entirely—one that requires billions in upfront capital, years of development, and a bet that your architecture will still be relevant when the chips finally tape out.
Why Samsung, not TSMC?
The choice of Samsung as a potential partner is the most telling detail. TSMC dominates advanced chip manufacturing, producing silicon for Apple, Nvidia, AMD, and most of the industry's cutting-edge designs. But TSMC's fabs are booked solid, with capacity allocated years in advance to its most lucrative customers. Samsung's foundry business, by contrast, has struggled to win marquee clients and has excess capacity at advanced nodes.
For Anthropic, this creates an opening. Samsung can offer faster timelines, more flexible terms, and likely more favorable pricing than TSMC would to a first-time customer. The trade-off is yield risk—Samsung's 3nm and 4nm processes have historically lagged TSMC's in manufacturing consistency. But for a company racing to reduce its dependence on Nvidia, an imperfect chip that exists beats a perfect chip that doesn't.
The economics of AI independence
Anthropic's interest in custom silicon reflects a brutal arithmetic. Nvidia's data-center GPUs carry gross margins north of 70 percent, and the company has used its market dominance to extract premium pricing from every major AI lab. Google, Amazon, and Microsoft have all responded by developing proprietary chips—TPUs, Trainium, Maia—to reduce their exposure. Anthropic, despite raising over $7 billion in funding, remains a customer rather than a peer in the silicon hierarchy.
A custom chip would allow Anthropic to optimize specifically for its Claude architecture, potentially achieving better performance-per-watt than general-purpose GPUs. More importantly, it would give the company leverage in negotiations with cloud providers and reduce the existential risk of Nvidia prioritizing competitors during the next GPU shortage.
Our take
This is Anthropic acknowledging what every serious AI company eventually learns: in the long run, you cannot build a defensible business on someone else's silicon. The safety-focused lab that once seemed content to let others race ahead on infrastructure has concluded that compute independence is itself a safety issue—if you cannot train your models, you cannot make them safer. Samsung gets a prestige customer; Anthropic gets optionality. Whether either gets a chip that actually works remains the expensive question.
