The "fire" among large AI model companies is now spreading to chips.
A few days ago, OpenAI proudly unveiled its first in-house inference chip, "Chili Pepper," outperforming NVIDIA, while Anthropic's chip, "Ambition," has also come to light...
According to a Reuters exclusive report, Anthropic previously discussed acquiring AI chip startup MatX for approximately $7 billion to accelerate its internal chip development.
Upon seeing this news, netizens also remarked: "First there was OpenAI, then Anthropic—now large models are entering the chip industry."
Unfortunately, the transaction ultimately did not move forward. According to sources familiar with the matter, discussions between the two parties have now shifted from an acquisition to potential collaboration.
Reuters has not disclosed the specific reasons for the termination of the acquisition talks, but it is noteworthy that Anthropic has demonstrated serious commitment to this deal: in pursuit of faster and more cost-effective computing power for Claude, the model company has begun moving downstream from the model layer all the way to the chip layer.
Just who is this chip company that Anthropic paid a high price to acquire?
According to available information, MatX was founded in 2023 by co-founders Reiner Pope and Mike Gunter, both formerly from Google. Reiner Pope worked on Google TPU software and large model infrastructure, while Mike Gunter has long been involved in TPU hardware design.
In February of this year, MatX completed a $500 million Series B funding round, with investors including Jane Street and Situational Awareness.
Notably, MatX focuses on designing chips for large language models, primarily serving the training phase.
This may be a key reason Anthropic chose to engage with MatX; according to Reuters, citing sources familiar with the matter, negotiations with MatX suggest Anthropic may intend to develop its own training chips, and it could also launch chips optimized for inference in the future.
This presents an interesting contrast with OpenAI’s recently publicized approach: OpenAI’s first in-house chip, “Jalapeño,” currently emphasizes inference—how to efficiently deploy models after training, with lower latency, higher throughput, and better energy efficiency. Anthropic’s revealed interests, by contrast, extend deeply into the training side as well.
In fact, it is not surprising that Anthropic has begun to pursue a "training chip" strategy.
Currently, as model sizes continue to grow, model training has increasingly become a super-scale engineering endeavor. From pre-training and post-training to reinforcement learning, each model iteration requires mobilizing massive clusters of chips. Even a few percentage points of improvement in training efficiency can translate into significant cost differences when scaled across tens of thousands or even hundreds of thousands of accelerators.
If the chip, model architecture, and training system are co-designed from the outset, this advantage could be further amplified.
This is why Google began developing TPUs years ago, Amazon has Trainium, and OpenAI is now following suit with Jalapeño... Chips are gradually becoming part of a model's capabilities.
MatX is not the only chip company Anthropic has engaged with.
In fact, Anthropic has not only engaged with MatX but is genuinely investing in chip design.
According to Reuters, over the past few weeks, Anthropic has held meetings with several AI chip startups, but it has not yet decided which company to acquire or fully determined the technical direction for its in-house chip development. A key objective of these discussions is to enable Anthropic’s engineers and leadership to systematically understand the various AI chip architectures currently available in the market.
Meanwhile, Anthropic is also aggressively recruiting talent in the chip industry.
Just a few days ago, Bloomberg reported 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.
Amir Salek is a seasoned veteran of the semiconductor industry. He joined Google in 2013, helping to establish and lead its custom chip business, and oversaw the TPU project for many years until his departure in 2022. During this time, he drove the development and delivery of Google’s first seven generations of TPUs and played a key role in building Google’s custom chip capabilities for data centers.
Before joining Google, Salek worked at Nvidia for approximately eight years as a Senior Engineering Director, where he founded and led Nvidia’s System-on-Chip (SoC) design team, gaining extensive experience in chips including GPUs and mobile processors.
After leaving Google in 2022, Salek transitioned into investment, joining private equity firm Cerberus Capital Management as Senior Managing Director and becoming a partner at its deep tech investment platform, Tracker Ventures, with a focus on semiconductors, AI, and edge computing.
He has now returned to the front lines of chip development, joining Anthropic’s computing team. He will report to the company’s computing lead, James Bradbury.
Earlier this year, in June, Anthropic hired Clive Chan, a former chip engineer at OpenAI, who had worked on OpenAI’s in-house chip project.
When these actions are viewed together, Anthropic’s chip strategy becomes very clear: recruit chip talent, build an in-house team, research different architectures, engage with chip startups, and even consider acquisitions worth billions of dollars...
Of course, Anthropic does not intend to fully shift to in-house chips; according to Reuters, it still plans to maintain a multi-chip strategy and continue collaborating with chip and cloud computing providers such as Nvidia and Google.
Because chip design is inherently a costly and time-consuming endeavor, even with ample funding, it is no easy task. A truly functional advanced chip can take a year or more to move from design to production, with the design cost for a single generation potentially reaching hundreds of millions of dollars.
Therefore, acquisition targets like MatX are particularly attractive to Anthropic, as directly acquiring an established AI chip startup can rapidly provide in-house chip design expertise and potentially reduce costs over the long term.
What about you? How do you view Anthropic’s move?
Reference link:
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.