Read this if you run or plan inference workloads and your data center is limited by power rather than floor space. FuriosaAI targets that case with air-cooled servers drawing about 3kW, and Samsung SDS and LG AI Research are named customers. It sells to enterprises and cloud providers, not individuals, and every price goes through sales.
FuriosaAI is a Seoul-based AI chip company founded in 2017, led by co-founder and CEO June Paik. It designs the RNGD inference accelerator, a PCIe card built on TSMC's 5nm process with HBM3 memory, and sells it alone or inside the eight-card NXT RNGD Server. RNGD entered mass production in January 2026.
Founded: 2017 · HQ: Seoul, South Korea · CEO: June Paik · Funding: $246M total after a $125M Series C in July 2025; board approved up to KRW 686B of new preferred shares in July 2026
About FuriosaAI
FuriosaAI is a Seoul-based chip company that designs processors for AI inference, the stage where a trained model answers requests. Its about page dates the start of R&D in Seoul to 2017. June Paik (백준호) is co-founder and CEO, and press reports say he previously worked at Samsung Electronics and AMD. The company's head office is in Gangnam-gu, Seoul, with offices in Silicon Valley and Lisbon, and a Singapore subsidiary, FuriosaAI Singapore Pte. Ltd., announced on September 10, 2026 to cover Asia-Pacific sales under Addison Chi, its Vice President of APAC Sales.
The product history runs in generations. The company says its first chip, a vision NPU built on Samsung's 14nm process, launched in 2021, and that it was the first AI chip startup to outperform Nvidia on MLPerf Inference that year. In 2022 it reported commercial deployment with Kakao and ecosystem partnerships with Hugging Face, Samsung, SK Hynix, TSMC, ASUS and LG AI Research. The first generation reached volume production in 2023 with Samsung Foundry and ASUS. The second generation, RNGD (pronounced "renegade"), was unveiled in 2024 on TSMC's 5nm process, sampled with global technology companies in 2025 and, per the company's blog, entered mass production in January 2026.
RNGD is sold two ways. The PCIe card has 48 GB of HBM3 memory, 1.5 TB/s of memory bandwidth, 256 MB of on-chip SRAM, 512 TFLOPS of peak FP8 compute and a 180W power rating. The NXT RNGD Server, launched in September 2025, packs up to eight cards into one chassis that the company says runs air-cooled in ordinary data centers. Software comes as the Furiosa SDK and furiosa-llm, a runtime that serves models through an OpenAI-compatible API and installs with pip. In its own test on Qwen3-32B at FP8 with 1K input and 1K output tokens, Furiosa reports up to 2× the tokens per second per kilowatt of an NVIDIA RTX PRO 6000. That is a vendor measurement against one GPU, not an independent benchmark.
Customers and partners are named in the company's 2026 announcements. LG AI Research adopted RNGD for its EXAONE models in July 2025 and reported 2.25 times the performance per watt of legacy GPUs, according to Verdict's coverage; HokAI lists the family under K-EXAONE 2.0. On July 7, 2026 Furiosa began installing RNGD servers at Equinix's Lisbon LS2 data center as a European evaluation site. On July 20, 2026 Samsung SDS launched what the two companies call Korea's first domestic NPU-as-a-Service on RNGD, hosted in Samsung SDS's Dongtan data center. On May 27, 2026 the company announced a partnership with Broadcom to build a third-generation accelerator for very large frontier models.
Funding comes from Korean state banks and strategic investors. A $125M Series C in July 2025 took total funding to $246M, with two Korean state-backed banks, Kakao's investment arm and two venture firms taking part. Korean press reported in March 2025 that Furiosa declined an $800M acquisition offer from Meta. In January 2026 Bloomberg reported a Series D target of $300M to $500M, advised by Morgan Stanley and Mirae Asset Securities, with a listing considered as early as 2027. A notice on the company's site says its board resolved on July 1, 2026 to issue up to 403,656 redeemable convertible preferred shares at KRW 1,699,449 each, up to KRW 685,992,785,544 in total, with payments scheduled between July 27 and October 28, 2026. No valuation is published, so this page does not state one.
Competition splits into two groups. Among inference-chip startups, Rebellions is the closest Korean rival, and Groq and SambaNova sell inference systems abroad. Against Nvidia GPUs, Furiosa competes on power per token rather than peak speed: a 180W card is sized for dense, air-cooled racks, and the company argues that is what limited data centers need. The weakness is software reach, since CUDA remains the default target for most model code, and the customer references published so far lean on Korean groups.
There is no public price list. Buyers request a quote or a trial from the company's contact and request-access pages, and the evaluation program offers bare-metal access to a dedicated server or a hosted endpoint that follows the OpenAI API format. None of the pages HokAI read lists a security certification such as SOC 2 or ISO 27001, so enterprise buyers should ask for audit reports directly.
Mission
Build AI chips capable of running the world's most advanced models efficiently, making AI computing more sustainable.
Products
- Furiosa RNGD PCIe Card (AI inference accelerator card): https://furiosa.ai/rngd
- Furiosa NXT RNGD Server (Eight-card data center inference server): https://furiosa.ai/nxt-rngd-server
- Furiosa SDK and furiosa-llm (Compiler and OpenAI-compatible serving runtime): https://developer.furiosa.ai/latest/en/
Links
Frequently Asked Questions
How much funding has FuriosaAI raised?
FuriosaAI raised a $125M Series C in July 2025, which brought its total funding to $246M at that point. Investors in that round included Korea Development Bank, Industrial Bank of Korea, Keistone Partners, PI Partners and Kakao Investment. Bloomberg reported in January 2026 that it was seeking $300M to $500M in a Series D, and a July 2026 notice on the company's site approves up to KRW 685,992,785,544 of new preferred shares. No valuation has been published.
What products does FuriosaAI make in 2026?
It makes the RNGD PCIe card, a 180W inference accelerator with 48 GB of HBM3, and the NXT RNGD Server, an eight-card appliance rated at about 3kW. The software side is the Furiosa SDK and the furiosa-llm runtime, which exposes an OpenAI-compatible API. None of these has a public price: buyers request a quote from the company.
How does FuriosaAI handle data privacy and compliance?
FuriosaAI sells hardware and an SDK, so data handling depends on where RNGD runs: a customer's own data center, an Equinix site or a partner cloud such as Samsung Cloud Platform. The pages HokAI read list no security certification, and no trust center was found. A privacy policy is published at furiosa.ai/privacy-policy, and enterprise buyers should request audit reports directly.
Who are FuriosaAI's main competitors?
Nvidia is the main rival: its GPUs are faster per card and have the CUDA software base, while Furiosa argues RNGD wins on tokens per kilowatt in power-limited racks. Rebellions is the closest Korean competitor in inference chips. Groq and SambaNova also sell inference systems, so buyers should test their own models on each.
How do you start using FuriosaAI's products?
Start with the request-access form on furiosa.ai, which offers bare-metal time on an NXT RNGD Server or an OpenAI-compatible API endpoint. Developers can also install the runtime with pip install furiosa-llm and launch a model with furiosa-llm serve. Samsung SDS offers RNGD instances as a cloud service in Korea for teams that do not want to buy hardware.