Anthropic Opens Talks With Samsung to Build Its First Custom AI Chip

Anthropic is in early discussions with Samsung Electronics to manufacture a custom AI accelerator chip, according to reporting by The Information, first published July 2. The talks are preliminary enough that Anthropic hasn't yet decided what the chip should do, how powerful it needs to be, or how it would physically fit into a server rack. What the company apparently has decided: renting compute forever isn't a business model.

According to a report reviewed by TechCrunch, the Claude developer has relied entirely on chips rented from Amazon, Google, and Nvidia, and that dependence is colliding with the mounting cost of running its largest models. The company reportedly carries a compute bill running above $1.25 billion per month. Custom silicon won't fix training costs, but inference is a different story. As noted in reporting aggregated by Bloomberg, shaving even 10 to 30 percent off inference costs matters enormously at Anthropic's scale, where the margin between profitable and burning can be that thin.

The Samsung angle is worth examining on its own terms. According to UPI, Anthropic is considering Samsung's 2-nanometer manufacturing process and advanced chip-packaging facilities. That's a direct bet on Samsung's foundry arm at a moment when the division is fighting hard for flagship customers. As reported by The Information, Samsung's foundry business has been pushing aggressively for partnerships with high-profile AI customers, particularly as TSMC faces capacity constraints and longer lead times. A win here gives Samsung a marquee name to sit alongside its existing Nvidia partnership, through which Samsung already produces chips that Nvidia needs to train and run AI models.

The competitive pressure behind Anthropic's timing is hard to miss. As TechCrunch reported, Anthropic's move follows OpenAI's announcement of its own custom inference processor, called "Jalapeño," built with Broadcom, which OpenAI claims delivers better performance-per-watt than competing chips. OpenAI, Google, Amazon, and Microsoft all have custom silicon embedded in their data center stacks now. Anthropic, as UPI noted, is entering this race later than most.

The infrastructure math explains why every major lab is chasing the same outcome. According to data cited by UPI from research firm TrendForce, server shipments using custom application-specific integrated circuits are projected to grow 44.6 percent in 2026, compared to 16.1 percent growth for servers running general-purpose GPUs. When your inference workload is that large and growing that fast, paying GPU-rack rates for commodity silicon is a compounding disadvantage.

What Anthropic doesn't have yet is an actual chip. According to reporting by The Information, the company hasn't begun detailed chip design, testing, or manufacturing. It's hired silicon engineers and is still in the specification phase. In a statement reported by Technology.org, Anthropic said a diversified hardware stack including chips from Google, Amazon, and Nvidia would remain central to its compute strategy, and added nothing further on the Samsung talks. Samsung declined to comment when contacted by TechCrunch.

The gap between "talks" and "shipping inference silicon" is wide and lined with examples of custom chip programs that took longer and cost more than expected. The story here isn't that Anthropic has a chip. It's that the company has finally concluded it needs one, and that conclusion took until mid-2026 to arrive.

Sources cited:
- The Information (https://www.theinformation.com/)
- TechCrunch (https://techcrunch.com/2026/07/02/anthropic-is-discussing-a-new-custom-chip-with-samsung/)
- Bloomberg (https://www.bloomberg.com/news/articles/2026-07-02/anthropic-in-talks-with-samsung-for-custom-ai-chip-information-mr3l34t4)
- UPI (https://www.upi.com/Top_News/World-News/2026/07/03/Anthropic-Samsung-Electronics/7811783128641/)
- Technology.org (https://www.technology.org/2026/07/04/anthropic-samsung-custom-ai-chip/)