FuriosaAI raises $620 million to tackle AI data center power limits

A massive capital injection

FuriosaAI recently announced a Series D funding round totaling $620 million. This investment brings the company's valuation to $2.9 billion, which is roughly four times its previous valuation of $735 million. The round highlights a significant shift in investor confidence toward energy-efficient hardware.

The funding was led by a mix of government and private interests. South Korea’s National Growth Fund contributed the largest portion at $280 million as part of a national push to develop a domestic chip industry. International investors added $110 million, with US-based Infinitum reportedly leading that group.

Other notable participants include Cisco Investments, QUDA Capital, and several Korean entities like NAVER and the Korea Development Bank. The Export-Import Bank of Korea also participated, marking its first-ever direct investment in a startup. This funding marks a record-breaking Series D for any semiconductor firm based in South Korea.

Solving the power bottleneck

The primary goal for FuriosaAI is to address the electricity crisis facing modern data centers. Currently, many facilities are running out of power faster than they are running out of physical chips. Because AI models require massive amounts of energy, the bottleneck for growth is shifting from hardware availability to grid capacity.

FuriosaAI focuses specifically on AI inference, which is the process of running a trained model to generate outputs for users. This is different from training, where models are first built. As companies move toward deploying AI for real-world use, the demand for inference-specific chips is growing rapidly.

The company’s flagship chip, the RNGD, uses a design called the Tensor Contraction Processor. This technology is designed to reduce the energy-heavy data shuffling that standard GPUs perform. By optimizing how data moves, the chip aims to deliver more computational power while consuming significantly less electricity.

Performance and production

The RNGD chip is designed to be highly efficient for large-scale operations. FuriosaAI reports that its 180-watt, air-cooled RNGD card can match the throughput of a GPU that draws twice as much power or more. In specific tests using LG’s EXAONE models, the chip achieved a 2.25x improvement in performance-per-watt compared to the GPU hardware it was measured against.

The company is moving beyond laboratory testing into full-scale commercial production. The RNGD chip entered mass production in January and is being manufactured by TSMC. This move signifies a transition from proving the technology works to proving that the business can scale to meet market demand.

Several major companies are already utilizing the hardware for their services. LG Group runs its EXAONE models on the RNGD chip, while Upstage uses it to power live search summaries for millions of people. Additionally, Samsung SDS debuted a cloud offering powered by the chip in July to serve its customers.

Scaling the roadmap

With the new capital, FuriosaAI has set ambitious manufacturing targets for the coming years. The company plans to produce 20,000 RNGD chips within this year. Looking further ahead, they aim to scale production to more than 40,000 chips by 2027.

The company is also looking toward future hardware iterations to maintain its competitive edge. A third-generation chip is currently in development in partnership with Broadcom. These efforts are intended to provide a long-term alternative to the current market dominance of NVIDIA.

While NVIDIA remains the leader in the AI chip space, the focus on power efficiency changes the math for data center operators. If a chip provides more tokens for the same amount of power, operators can serve more users without needing to build new power plants. FuriosaAI is reportedly eyeing a public listing as early as 2027 to further its growth.