(Reuters) - Enfabrica, a California-based startup that aims to make AI chips work more efficiently together at scale, said on Tuesday it raised $115 million in funding and plans to release its newest chip early next year.

Founded by veterans from Broadcom and Alphabet, Enfabrica is tackling one of the biggest technical problems that has emerged in the AI field - how to tie tens of thousands or more chips together with a network.

If that network is too slow, expensive chips from firms such as, which is an investor in Enfabrica, end up sitting idle and waiting for data.

Enfabrica's chip aims to address the bottlenecks by letting AI computing chips talk to more parts of a network at once than current networking chips. Enfabrica co-founder and Chief Executive Rochan Sankar told Reuters that current technologies can string together about 100,000 AI computing chips before the network starts to bog down.

Sankar said Enfabrica's technology could boost that figure to about 500,000 chips and make it possible to train even larger AI models. This process often takes weeks or months, and millions of dollars can be wasted if the resulting AI model is not reliable or accurate.

"It's become apparent in the last six to nine months that the attributes of that network really drive the capability of that (computing power), whether it's bandwidth, resiliency or recovery from loss," Sankar said. "All these things matter when you start running at scale."

The funding round announced Tuesday was led by Spark Capital, joined by new investors Maverick Silicon and VentureTech Alliance. Also joining the funding round were existing investors that include Atreides Management, Alumni Ventures, IAG Capital, Liberty Global Ventures, Sutter Hill Ventures and Valor Equity Partners.

(Reporting by Stephen Nellis in San Francisco; Editing by Cynthia Osterman)