Anthropic reportedly explored acquiring AI chip startup MatX for approximately $7 billion to accelerate custom accelerator development.

The competitive race among leading artificial intelligence model developers is now extending into semiconductor design. Recently, OpenAI unveiled its first proprietary inference chip, codenamed “Chili Pepper,” claiming performance advantages over NVIDIA, while reports also surfaced regarding Anthropic’s internal chip initiative, dubbed “Ambition.”

According to a Reuters exclusive, Anthropic previously held discussions to acquire AI chip startup MatX for approximately $7 billion, aiming to accelerate its internal chip development. The announcement prompted industry observers to remark: “First there was OpenAI, then Anthropic—now large models are entering the chip industry.” Ultimately, however, the transaction did not proceed. Sources familiar with the matter indicate that talks between the two parties have since pivoted from an acquisition to explore potential collaboration.

While Reuters has not disclosed specific reasons for the termination of the acquisition talks, Anthropic’s commitment to the deal was evident. Seeking faster and more cost-effective compute power for its Claude model family, the company has begun moving downstream from the model layer all the way to the chip layer.

MatX, the target of Anthropic’s high-value interest, was founded in 2023 by co-founders Reiner Pope and Mike Gunter, both former Google employees. Pope previously worked on Google TPU software and large model infrastructure, while Gunter has extensive experience in TPU hardware design. Earlier this year, in February, MatX closed a $500 million Series B funding round backed by investors including Jane Street and Situational Awareness. The company specializes in designing chips for large language models, with a primary focus on the training phase.

This focus likely explains Anthropic’s engagement. Citing sources familiar with the negotiations, Reuters reported that Anthropic may intend to develop its own training chips, with plans to eventually launch inference-optimized processors as well. This approach contrasts with OpenAI’s recently publicized strategy: OpenAI’s first in-house chip, “Jalapeño,” currently emphasizes inference—optimizing efficient model deployment post-training through lower latency, higher throughput, and improved energy efficiency. Anthropic’s revealed interests, by contrast, extend deeply into the training side.

Anthropic’s pivot toward training chips aligns with broader industry trends. As model sizes continue to expand, training has evolved into a super-scale engineering challenge. From pre-training and post-training to reinforcement learning, each iteration requires mobilizing massive accelerator clusters. Even marginal gains in training efficiency yield substantial cost savings when scaled across tens or hundreds of thousands of chips. Co-designing the chip, model architecture, and training system from the outset can further amplify these advantages. This vertical integration explains why Google developed TPUs years ago, Amazon launched Trainium, and OpenAI is now pursuing Jalapeño—custom silicon is gradually becoming integral to a model’s capabilities.

MatX is not the only chip company Anthropic has engaged with. Beyond acquisition talks, Anthropic is actively investing in internal chip design capabilities. Over the past few weeks, the company has held meetings with multiple AI chip startups to help its engineers and leadership systematically evaluate existing market architectures. However, it has yet to decide on an acquisition target or finalize the technical direction for its in-house development.

Concurrently, Anthropic is aggressively recruiting semiconductor talent. Bloomberg reported earlier this week that Anthropic is forming an internal chip team and has hired Amir Salek, former head of Google’s TPU core, to join its computing division and advance its custom chip initiative. Salek brings deep industry experience: he joined Google in 2013, helped establish its custom chip business, and oversaw the TPU project until his departure in 2022, driving the development and delivery of Google’s first seven TPU generations. Prior to Google, he spent roughly eight years at NVIDIA as a Senior Engineering Director, where he founded and led the System-on-Chip (SoC) design team, gaining extensive experience with GPUs and mobile processors. After leaving Google in 2022, Salek transitioned into investment, joining Cerberus Capital Management as Senior Managing Director and becoming a partner at its deep tech platform, Tracker Ventures, focusing on semiconductors, AI, and edge computing. He now returns to hardware development, reporting to James Bradbury, Anthropic’s computing lead.

This hiring follows another strategic move in June, when Anthropic recruited Clive Chan, a former chip engineer at OpenAI who contributed to OpenAI’s in-house chip project. Viewed together, these actions clarify Anthropic’s hardware strategy: recruit specialized talent, build an internal team, research diverse architectures, engage with chip startups, and pursue multi-billion-dollar acquisitions.

Despite this push, Anthropic does not intend to fully transition to in-house silicon. According to Reuters, the company will maintain a multi-chip strategy and continue collaborating with established providers such as NVIDIA and Google. Chip design remains inherently capital-intensive and time-consuming; even with substantial funding, developing a functional advanced processor typically takes over a year from design to production, with single-generation design costs potentially reaching hundreds of millions of dollars. Acquiring an established startup like MatX offers a pragmatic alternative, rapidly injecting in-house design expertise while potentially lowering long-term costs.

Readers are invited to share their perspectives on Anthropic’s strategic shift into hardware.

https://www.reuters.com/business/finance/anthropic-planned-then-abandoned-7-billion-purchase-matx-sources-say-2026-08-27/

https://www.bloomberg.com/news/articles/2026-08-21/anthropic-taps-google-chip-veteran-as-part-of-push-into-hardware

This article is from the WeChat public account "Machine Heart" (ID: almosthuman2014), authored by someone interested in AI.