Anthropic abandoned its $7 billion MatX purchase, according to sources
Two?people who were briefed about the issue said that Anthropic had discussed purchasing artificial intelligence chip company MatX for around $7 billion. The goal was to speed up efforts to develop custom-made hardware for its rapidly growing AI business.
The merger talks that a third party said had evolved into a conversation about a partnership?underscores?the AI Lab's ambition to secure resources and talent required to build its chips. Anthropic wants to accelerate the development of its own chips.
Reporting on the discussions is the first time, but I could not find out why they are no longer in progress. MatX, founded in 2011 by ex-Google tensor processor unit (TPU), engineers, is seeking new capital to be valued at about $4 billion. One person, who requested anonymity for private discussions, said that MatX was now looking to raise new funds.
Anthropic refused to comment on the deal negotiations with MatX. MatX didn't respond to an inquiry for comment.
Anthropic is on the road to IPO with its chip plans. As it scales up its?Claude AI family, Anthropic wants to produce hardware capable of satisfying its voracious appetite for crunching data and reducing its dependence on Nvidia chips. Anthropic is expected to go public 'this year' and has recruited executive and engineering talent to speed up the chip design process. This could take many years. Anthropic, which is expected to list 'this year,' has hired executive and engineering talent to accelerate the chip design process, which could take years. Anthropic announced it was expanding its in-house silicon team, which will design custom chips to allow Claude models run faster and more effectively. It will also continue to use a multichip approach by working with suppliers from Nvidia and Google. The process of designing chips is time-consuming and expensive. Design costs for a single generation can be hundreds of millions of dollars and it may take an entire year to create a piece of viable hardware. Anthropic could lower its costs in the long run by acquiring an AI chip startup like MatX. MatX is working on a chip which would be helpful for training large AI models. Anthropic will spend tens and tens billions to rent computing power to cloud providers, as well as to purchase chips. It has signed a $45-billion deal with Nscale to rent AI cloud computing. It agreed to pay SpaceX a total of $1.25 billion a month for its computing power across all data center clusters through May 2029. DISCUSSIONS with other AI?CHIP STARTUPS
Anthropic's discussions with MatX suggest that it may also be interested in developing a training chip. Meanwhile, other chip startups, including rival OpenAI, are pursuing processors more suited for generating chatbot responses, or inference. Sources said that Anthropic may also choose to manufacture an inference chip.
In the last few weeks, Anthropic has met with a number of AI chip startups. The company has not made a decision to acquire. Anthropic engineers and executives are attending the meetings to better understand current chip design approaches.
Anthropic has hired Amir Salek, a Google chip veteran, as part of its push into hardware. The AI lab hired Clive Chan, a former OpenAI chip developer who worked on OpenAI's chip that was unveiled in June.
Anthropic, OpenAI and other leading AI labs are increasingly focused on custom-made chips that can be tailored for their models and workloads. They 'hope to achieve significant economic and performance advantages through this.
OpenAI announced at a recent conference that its custom chip, Jalapeno, outperformed an Nvidia processor of a similar design. OpenAI executives said that the chip performed inference calculations more efficiently. Anthropic may be able to hedge against the shortage of Nvidia processors. Nvidia announced in a Wednesday conference call that they would only have a limited supply until 2027. Google developed TPUs while Amazon built their Trainium and Inferentia processors. Anthropic was the first company to run their models on hardware made by several vendors, including Nvidia and Google. The company is seeking to increase its computing capacity in order to meet the demand for its models. It has signed a deal with SpaceX, to use their Colossus 1 facility which contains more than 220,000 Nvidia chip. Max A. Cherney in San Francisco and Krystal H. Hu, editor David Gregorio.
(source: Reuters)
