Anthropic is eyeing "training chips"? It was revealed that it once planned to acquire AI chip company MatX for 7 billion US dollars.

The "battlefire" among large AI model companies is spreading all the way to the chip sector.

The other day, OpenAI just grandly announced that its first self-developed inference chip "Jalapeño" outperforms Nvidia, while on the other side, Anthropic's "ambition" for chips has also surfaced...

According to an exclusive Reuters report, Anthropic previously discussed acquiring AI chip startup MatX for approximately 7 billion US dollars, hoping to accelerate its in-house chip R&D through this move.

Seeing this news, netizens have expressed their views one after another: after OpenAI comes Anthropic, "now large model companies are all starting to step into the chip field."

However, it is regrettable that this transaction was not pushed forward in the end. According to people familiar with the matter, the current discussions between the two sides have shifted from acquisition to potential cooperation.

Reuters has not disclosed the specific reason for the termination of the acquisition negotiations, but it is worth noting the level of seriousness Anthropic has shown towards this deal: to find faster and cheaper computing power for Claude, this model company has begun to dig deep from the model layer all the way to the chip layer.

And what kind of outstanding enterprise is this chip company that can make Anthropic want to acquire it at a high price?

According to public information, MatX was founded in 2023, and its co-founders Reiner Pope and Mike Gunter are both from Google. Reiner Pope once participated in work related to Google TPU software and large model infrastructure, while Mike Gunter has long been engaged in TPU hardware design.

In February this year, MatX completed a 500 million US dollars Series B financing, with investors including Jane Street, Situational Awareness and others.

It is worth noting that MatX focuses on designing chips for large language models, mainly serving the "training" segment.

This may be an important reason why Anthropic chose to contact it. Citing people familiar with the matter, Reuters said that the negotiations with MatX indicate that Anthropic may intend to develop its own "training chips", and of course, it may also launch chips for inference in the future.

This point clearly forms a very interesting contrast with the path that OpenAI has recently made public: OpenAI's first self-developed chip "Jalapeño" currently emphasizes inference: after the model training is completed, how to run the model services with lower latency, higher throughput and better energy efficiency. The interest exposed by Anthropic goes deep into the training side at the same time.

In fact, it is not unexpected for Anthropic to start laying out the "training chip" strategy.

At present, as the scale of models is getting larger and larger, model training is increasingly approaching a super project. From pre-training, post-training to reinforcement learning, behind every model iteration, a huge chip cluster needs to be mobilized. Even if the training efficiency is increased by a few percentage points, at the scale of tens of thousands or even hundreds of thousands of accelerators, it may eventually correspond to a quite considerable cost gap.

If the chip, model architecture and training system can be co-designed from the very beginning, this advantage may continue to be amplified.

This is why Google started building TPU many years ago, Amazon has Trainium, and OpenAI now follows up to launch Jalapeño... Chips are gradually becoming part of model capabilities.

MatX is not the only chip company that Anthropic has contacted

In fact, Anthropic has not only contacted MatX, but is seriously laying out chip design.

According to Reuters, in recent weeks, Anthropic has met with many AI chip startups, but it has not yet decided which company to acquire in the end, nor has it fully determined which technical route its self-developed chips will adopt. One important purpose of these contacts is to let Anthropic's engineers and management systematically understand different AI chip architectures currently on the market.

At the same time, Anthropic is also frantically recruiting talents in the chip field.

Just a few days ago, Bloomberg reported that Anthropic is forming an in-house chip team, and has hired Amir Salek, the former core head of Google TPU, to join its computing department to promote the self-developed chip plan.

Amir Salek is a senior veteran in the chip industry. He joined Google in 2013, participated in the founding and led its custom chip business, and was in charge of the TPU project for a long time until he left in 2022. During this period, he promoted the development and delivery of the first seven generations of Google TPU products, and participated in building Google's custom chip capabilities for data centers.

Before joining Google, Salek worked at Nvidia for about eight years, serving as senior engineering director, and founded and led Nvidia's system-on-chip (SoC) design department, accumulating long-term experience in GPU, mobile processors and other chip fields.

After leaving Google in 2022, Salek moved to the investment field, joined private equity firm Cerberus Capital Management as a senior managing director, and at the same time served as a partner of Tracker Ventures, its deep tech investment platform, focusing on semiconductors, AI, edge computing and other fields.

Now he returns to the front line of chip R&D to join Anthropic's computing team. It is reported that he will report to James Bradbury, the company's head of computing.

Earlier this year, in June, Anthropic also hired Clive Chan, a former chip engineer at OpenAI, who once participated in OpenAI's self-developed chip project.

If you connect these operations and look at them comprehensively, Anthropic's chip roadmap will be very clear: poaching chip talents, building in-house teams, researching different architectures, contacting chip startups, and even directly considering acquisitions at the multi-billion-dollar level...

Of course, Anthropic does not intend to fully shift to self-developed chips. According to Reuters, it still plans to continue to adopt a multi-chip route and maintain cooperation with chip and cloud computing suppliers such as Nvidia and Google.

Because chip design itself is an expensive and time-consuming large project, even if you have enough funds, it is not easy. It should be noted that a truly usable advanced chip may take one year or even longer from design to mass production, and the design cost of a single generation of chips may reach hundreds of millions of US dollars.

That's why acquisition targets like MatX are very attractive to Anthropic. After all, directly acquiring a mature AI chip startup can quickly obtain in-house chip design experience, and may also reduce costs in the long run.

What about you? What do you think of this move by Anthropic?

Reference links:

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 Official Account Synced (ID: almosthuman2014), the author is a follower of AI, and 36Kr publishes it with authorization.