Energy-efficient inference chip developer FuriosaAI is raising $620 million to address data center power shortages

South Korean inference chip startup FuriosaAI Inc. announced today that it is ready to launch its challenge to Nvidia Corp. after raising a record $620 million in funding to scale production of its energy-efficient artificial intelligence processors.

The round brings FuriosaAI’s valuation to a whopping $2.9 billion, the startup said.

The investment was led by South Korea’s National Growth Fund, which brought in $280 million of the total, part of a government-backed plan to boost the country’s domestic semiconductor industry. Global investors such as Cisco Investments, Abu Dhabi-based NeoFort Holdings, QUDA Capital Partners, Jericho Capital and Worldline Capital contributed another $110 million, while existing backers such as Korea Development Bank, Kako Investment and NAVER contributed the rest.

The nine-year-old startup and its investors believe it has a decisive advantage over Nvidia. According to her, the bottleneck in AI data centers currently is not the availability of computing power, but rather the lack of enough energy to run all of these processors.

Nvidia’s power-hungry chips still dominate the industry, but many data centers are now struggling to scale because they can’t get enough power to power all those processors. Power grids and cooling systems are utilized to the maximum possible extent, which is why FuriosaAI’s inference chips, designed to maximize tokens per watt, are such a tempting alternative.

FuriosaAI introduced its energy-efficient Tensor contraction processor earlier this year. It is the second generation of inference chips based on the RNGD architecture. Unlike Nvidia’s graphics processors, which use a rigid two-dimensional matrix to process calculations, the TCP-based chips, including RNGD, use a higher-dimensional primitive, tensor contraction, as the basic unit of calculation

By treating tensor contractions as the primary computing unit and using deterministic execution on silicon, the TCP chips keep energy-consuming data movements to a minimum. This means they use less than half the power of Nvidia’s GPUs while offering compatibility with AI frameworks like PyTorch and vLLM.

The startup’s chips have already gained a large following in South Korea. For example, LG Group, the parent company of LG Electronics Co. Ltd., uses FuriosaAI chips to power its EXAONE models. Meanwhile, Korean AI startup Upstage is using the chips to power live search summaries for millions of users, and Samsung SDS is providing access to the company’s silicon through its commercial Neural Processing Units-as-a-Service offering, which launched this summer.

Funds from today’s round will help FuriosaAI increase production volumes of next-generation TCP chips, the startup said. It plans to produce 20,000 this year and double that number to 40,000 in 2027 before scaling up further in the coming years. At the same time, work is also underway on a third-generation processor, which is being developed together with Broadcom Inc.

Co-founder and chief executive June Paik (pictured) said the size of today’s round reflected investors’ strong confidence in the company’s technology. “We are entering a new phase in which we bring our technology to the global AI infrastructure market at scale,” he emphasized. “By accelerating the commercialization of RNGD and advancing the development of Gen 3, we want to help shape the next generation of AI infrastructure.”

Yong Jung Park, a partner at NAVER Ventures, which has participated in every funding round since FuriosaAI’s pre-seed in 2017, said the company has overcome the most difficult hurdle facing semiconductor companies – converting architecture into proven, commercial hardware. “We’re looking for the moment when tech companies move from testing their technology to testing their business model,” Park said. “FuriosaAI is now at this point, the chip is in mass production and deployed to customers.”

Image: FuriosaAI