Abstract:
According to people familiar with the matter, artificial intelligence company Anthropic has entered into negotiations to acquire AI chip start-up MatX, with a transaction valuation of approximately US$7 billion. Anthropic hopes to speed up the process of developing custom hardware for its rapidly expanding AI business.
Two of the people familiar with the matter said that the acquisition plan is aimed at helping Anthropic obtain the resources and talents needed to build its own chips. Discussions later turned to a partnership, a third person familiar with the matter said, but Reuters was not yet informed of why the original acquisition talks were no longer moving forward.
MatX was founded by engineers who had participated in the development of Google's tensor processor (TPU). Currently, the company is seeking a new round of financing with a valuation of approximately US$4 billion. The sources spoke on condition of anonymity because the discussions involved undisclosed matters. Anthropic declined to comment on deal negotiations with MatX. MatX did not respond to a request for comment.
As the Claude family of AI models continues to expand in scale, Anthropic is seeking to produce hardware that can meet its huge data processing needs and reduce its reliance on Nvidia chips. The chip design process can take years. Anthropic is expected to go public this year, and the company has recruited engineering and management talents to speed up chip design work.
Reuters reported earlier this month that Anthropic is expected to IPO months after SpaceX goes public at a valuation of $1 trillion and will seek a valuation of about $2 trillion; this valuation is expected to depend on its revenue target of up to $200 billion in 2028.
Anthropic said it is expanding its internal silicon chip team to design custom chips to make Claude models run faster and more efficiently. The company also plans to maintain a multi-chip supply strategy and continue to work with suppliers ranging from Nvidia to Google.
Designing chips is expensive and takes a long time. Development of a working piece of hardware typically takes a year or more, and the design cost of a single generation of chips can run into hundreds of millions of dollars. People familiar with the matter said that acquiring AI chip startups such as MatX would allow Anthropic to gain in-house chip design capabilities and potentially reduce costs in the long term. MatX is developing a chip that can be used to train large AI models.
Anthropic plans to invest tens of billions of dollars to rent computing power from cloud service providers and will also directly purchase chips. The company plans to purchase Google AI chips worth US$36 billion and sign an AI cloud computing lease agreement with Nscale worth US$45 billion; in addition, it also agreed to pay SpaceX US$1.25 billion per month until May 2029 to obtain the computing power of its data center cluster.
Contacts with MatX also indicate that Anthropic may be interested in developing training chips; in contrast, other chip startups and rival OpenAI are more focused on inference chips suitable for generating chatbot responses. Anthropic may also choose to develop inference chips, people familiar with the matter said.
Anthropic has met with several AI chip startups in recent weeks but has not yet decided to make an acquisition. The talks are intended to give the company's engineers and management a sense of what's currently happening on different chip design routes.
This week, Anthropic hired Google chip veteran Amir Salek to boost its hardware business. In June this year, the company also hired Clive Chan, a former OpenAI chip engineer who was involved in the development of the chip that OpenAI released earlier this year.
Leading AI laboratories such as Anthropic and OpenAI are increasingly paying attention to customized chips, hoping to achieve significant performance and cost advantages by optimizing their own models and workloads. At a conference this week, OpenAI said its first custom chip, "Jalapeno," outperforms a similar Nvidia processor and consumes less energy when performing inference calculations.
Customized chips could also help Anthropic hedge against tight supply of Nvidia processors. Nvidia said on a conference call on Wednesday that its chip supply will remain tight until 2027. Google has developed TPU, and Amazon has launched Trainium and Inferentia chips.
Anthropic is one of the first companies to run models on hardware from multiple vendors, including Nvidia, Google and Amazon. To meet the rapid growth in demand for its models, the company has been seeking more computing power, including a deal with SpaceX to use its Colossus 1 facility, which houses more than 220,000 Nvidia chips.
Comments