Anthropic previously discussed acquiring AI chip startup MatX for about $7 billion, but the deal did not move forward. Reuters, citing people familiar with the matter, reported that discussions between the two sides have now shifted from a takeover to a potential partnership.
The report adds to a growing pattern in the AI industry: competition between model companies is no longer limited to software and model performance. It is moving down the stack into chips. In Anthropic’s case, Reuters said the company had looked at MatX as a way to speed up its in-house chip effort and secure faster, cheaper computing power for Claude.
Reuters did not disclose why the acquisition talks ended. Even so, the fact that Anthropic explored a transaction of that size shows how seriously it is evaluating direct involvement in chip development rather than relying only on outside suppliers.
MatX is focused on training chips for large language models
MatX was founded in 2023. Its co-founders, Reiner Pope and Mike Gunter, both came from Google. Pope worked on Google TPU software and large-model infrastructure, while Gunter spent years on TPU hardware design.
In February, MatX raised $500 million in a Series B round. Investors included Jane Street and Situational Awareness. The company focuses on chips built for large language models, with its primary emphasis on training workloads.
Reuters, citing people familiar with the talks, said Anthropic’s engagement with MatX indicates the company may be considering its own training chip. It could also launch inference chips in the future.
That stands in contrast to the path recently made public by OpenAI. According to the input material, OpenAI’s first in-house chip, Jalapeño, is currently positioned around inference: running trained models with lower latency, higher throughput and better energy efficiency. Anthropic’s apparent interest reaches into the training side as well.
Training economics are pushing model companies deeper into hardware
For Anthropic, a move into training chips is not hard to understand from a cost and infrastructure standpoint. As model sizes grow, training has become a large-scale engineering task. Pre-training, post-training and reinforcement learning all require vast accelerator clusters. A small gain in training efficiency can translate into a meaningful cost difference when the system is running across tens of thousands, or even hundreds of thousands, of accelerators.
If chip design, model architecture and training systems are developed in tandem from the start, that advantage can become larger. That is part of the reason Google built TPUs years ago, Amazon has Trainium, and OpenAI has now introduced Jalapeño. Chips are becoming part of what defines model capability.
Anthropic has been talking to more than one chip startup
MatX is not the only chip company Anthropic has approached. Reuters reported that in recent weeks the company has met with several AI chip startups. It has not decided which company, if any, it would eventually acquire, and it has not fully settled on the technical path for its internal chip effort.
One purpose of those meetings, according to Reuters, is to give Anthropic’s engineers and executives a structured view of the AI chip architectures currently available in the market. In that sense, the MatX talks appear to be one part of a broader review rather than a one-off event.
Internal hiring is moving in parallel
Anthropic is also building out its own chip bench. Bloomberg reported days earlier that the company was assembling an internal chip team and had hired former Google TPU leader Amir Salek into its compute organization to help advance its in-house chip plans.
Salek joined Google in 2013 and helped create and lead its custom chip business. He oversaw TPU work for years until leaving in 2022. During that period, he drove the development and delivery of Google’s first seven generations of TPU products and helped build the company’s custom silicon capability for data centers.
Before Google, Salek spent about eight years at Nvidia as a senior engineering director. He founded and led Nvidia’s system-on-chip, or SoC, design group and built experience across GPUs and mobile processors.
After leaving Google in 2022, Salek moved into investing. He joined private equity firm Cerberus Capital Management as a senior managing director and also became a partner at its deep-tech investment platform, Tracker Ventures, focusing on semiconductors, AI and edge computing. He has now returned to front-line chip development at Anthropic, where he is expected to report to compute head James Bradbury.
Earlier this year, in June, Anthropic also hired former OpenAI chip engineer Clive Chan, who had worked on OpenAI’s internal chip project, according to the input material.
Anthropic is still keeping a multi-chip strategy
Taken together, Anthropic’s direction is coming into view: recruiting chip talent, building an internal team, studying different architectures, meeting with startups and even weighing multibillion-dollar acquisitions.
Reuters also said Anthropic does not plan to switch fully to in-house chips. The company still expects to keep a multi-chip approach and continue working with chip and cloud suppliers including Nvidia and Google.
That reflects the realities of chip development. The input material notes that an advanced chip can take a year or longer to go from design to deployment, and the design cost for a single generation can run into the hundreds of millions of dollars. Against that backdrop, a company like MatX carries obvious appeal for Anthropic: an acquisition could bring internal chip-design experience much faster and may lower costs over time.