South Korean startup FuriosaAI declined an offer to sell to Meta that was worth up to 800 million dollars, both supporting and advancing its own vision on the AI hardware market. Barely three months afterwards, the firm revealed a groundbreaking collaboration with LG AI Research to provide its customized RNGD AI chip, as one of the enterprises that can utilize the EXAONE 4.0 platform. This strategy is foundational to an increasing movement in chip startups that script sovereignty and niche innovation rather than absorption by tech conglomerates.
The partnership places RNGD accelerators produced by FuriosaAI in the center of South Korean AI sovereignty efforts. LG EXAONE, a larger language model accessible in various verticals, including biotechnology, finance, will use hardware provided by FuriosaAI, as they enable its operations to be more cost-efficient. Trial runs indicated that RNGD upgraded inference performance by 2.25 times that of conventional GPU systems, a decisive benefit given how data centers all over the world are struggling to keep up with the rising energy requirements of AI.
Why FuriosaAI Walked Away From Meta
CEO June Paik argues that the fall of negotiations on the Meta acquisition was caused by those incompatible visions. As Paik put it, “We are working to carry on our mission.” FuriosaAI is devoted to sustainable AI computing. It has been reported that the disagreements were over strategy post-acquisition and organizational independence and not over financial terms. The offer made by Meta also indicated its desperate need to diminish its reliance on the chips manufactured by Nvidia, a fact that has now been made more difficult with FuriosaAI going rogue.
In the case of the 15-employee startup, the LG agreement proves its inference-optimization emphasis, or the practice through which prepared AI locates responses. In contrast to the general-purpose GPUs, RNGD chips deploy a unique architecture Tensor Contraction Processor, designed to excel on AI tasks. This design reduces the data movement and cuts down on power requirements, as LG is able to install air-cooled 4U servers instead of liquid-cooled solutions mandated by the high-wattage GPUs.
Technical Breakthroughs Behind the LG Win
LG’s selection followed seven months of rigorous testing. FuriosaAI’s RNGD accelerators (manufactured using TSMC’s 5nm process) achieved 512 teraflops of FP8 performance at just 150-180 watts thermal design power. In real-world benchmarks using LG’s 32-billion-parameter EXAONE model, a four-card RNGD configuration processed prompts at 50-60 tokens per second with near-human response latency (0.3 seconds for initial tokens).
FuriosaAI is efficient because of its very high bandwidth memory (HBM3) and compiler optimizations. It recycles the memory across the transformer layers, eliminating data transfers that are energy-intensive. In the case of scaling to many chips, FuriosaAI designed PCIe 5.0 peer-to-peer interplays to prevent bottlenecks, a less expensive option to the proprietary NVLink of Nvidia.
Challenging Nvidia’s Dominance
The collaboration is also a significant move by a large enterprise that will support a non-NVIDIA AI technology. Even though Nvidia already enjoys a market share of more than 98 percent in data center GPUs, FuriosaAI goes after the inference direction as varied application use cases leave market gaps that can be filled by experts. The analysts reveal that the inference market might grow up to 255 billion by 2030 compared to the training segments.
“RNGD has an architecture lending itself to high-queries-per-second and lower-precision inference,” said Matt Kimball, Moor Insights & Strategy analyst. Applications like customer service bots or financial analysis (core to LG’s target industries) benefit from such tailored designs.
Global Ambitions
Though rooted in South Korea’s “sovereign AI” movement, FuriosaAI eyes worldwide expansion. Paik confirmed LG’s global clients will access RNGD-powered EXAONE services, with negotiations underway in the U.S., Middle East, and Southeast Asia. The startup plans new funding rounds ahead of a potential IPO, betting that efficiency-conscious enterprises will seek alternatives to GPU giants.
As data centers consume more energy than small countries, power-sipping chips developed by FuriosaAI are much more than competitively inferred AI; they point toward a long-term scenario of the AI revolution. Being amongst five firms of AI hardware companies that have successfully shipped units in value scaling levels, this 15-man company has shown that, at times, autonomy can be the stage to take the largest jumps.
