FuriosaAI
Last reviewed
Jun 7, 2026
Sources
20 citations
Review status
Source-backed
Revision
v1 · 2,401 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 7, 2026
Sources
20 citations
Review status
Source-backed
Revision
v1 · 2,401 words
Add missing citations, update stale details, or suggest a clearer explanation.
FuriosaAI is a South Korean fabless semiconductor startup that designs data-center AI chip accelerators for deep learning inference. It was founded in 2017 by June Paik, a former Samsung Electronics and AMD engineer, and is headquartered in Seoul with an additional office in Santa Clara, California. The company has shipped two generations of silicon: Warboy, a first-generation vision-oriented neural processing unit (NPU), and RNGD (pronounced "Renegade"), a second-generation inference accelerator built on a custom "Tensor Contraction Processor" (TCP) architecture and aimed at running large language model workloads efficiently. FuriosaAI drew unusually wide attention in 2025 when it reportedly rejected an roughly $800 million acquisition offer from Meta and instead signed LG AI Research as an anchor customer for RNGD. It positions itself as one of a small group of startups attempting to challenge Nvidia in the AI inference market.
FuriosaAI was established in 2017. Its founder and chief executive, June Paik (Korean name Paik June-ho), had previously worked as an engineer at Samsung Electronics and at AMD, where he saw demand building for specialized processors to run deep learning workloads. He started the company to build energy-efficient AI accelerators for data centers, betting that a purpose-built chip could beat general-purpose graphics processors on the narrow but growing task of model inference.
The company is fabless: it develops the chip architecture and software but outsources fabrication to external foundries. Early backers included the South Korean internet company Naver and the venture firm DSC Investment. According to a February 2025 report in The Korea Herald, Paik held about 18.4 percent of the company at that time. By 2025 FuriosaAI employed on the order of 150 people across its Seoul and Santa Clara offices, a small team relative to the established chipmakers it competes against.
FuriosaAI's first-generation product was Warboy, an NPU introduced in 2021. It was fabricated by Samsung's foundry on a 14-nanometer process, with the Korean SoC design platform company SEMIFIVE assisting in bringing it to production. Warboy was oriented toward computer-vision inference such as image classification and object detection rather than large language models.
| Warboy specification | Value |
|---|---|
| Vendor / generation | FuriosaAI, first generation |
| Foundry / process | Samsung, 14 nm |
| Introduced | 2021 |
| Peak performance | 64 TOPS (INT8) |
| Architecture | Two processing elements, 32 TOPS each, independently deployable |
| On-chip SRAM | 32 MB |
| Memory | 16 GB LPDDR4X, about 66 GB/s |
| Host interface | PCIe Gen4 x8 |
| TDP | 40 to 60 W (configurable) |
| Form factor | Single-slot, half-height half-length PCIe card |
RNGD, FuriosaAI's second-generation accelerator, was first detailed publicly at the Hot Chips conference in August 2024 and is the product on which the company has staked its future. It is built specifically for AI inference, including the reasoning and language-model workloads that dominate modern data-center demand. RNGD is manufactured by TSMC on a 5-nanometer process, carries roughly 40 billion transistors, and uses two stacks of HBM3 memory supplied by SK hynix.
The defining feature of RNGD is its underlying architecture, which FuriosaAI calls the Tensor Contraction Processor. A tensor contraction is a generalization of matrix multiplication that operates directly on multidimensional arrays. FuriosaAI's argument is that the building block of conventional GPUs and many AI accelerators is the two-dimensional matrix multiply, so those designs must first flatten the multidimensional tensors that AI models actually use into 2D matrices before computing on them. The company says that flattening step destroys parallelism present in the original data and wastes hardware. TCP instead treats the tensor contraction itself as the native hardware primitive, which FuriosaAI says lets the chip keep more data on-chip and cut the costly movement of data between the chip and external memory; it notes that moving data to and from HBM3 can cost far more energy than the arithmetic itself. FuriosaAI presented a peer-reviewed paper on the architecture, "FuriosaAI RNGD: A Tensor Contraction Processor," at the MICRO 2025 conference.
RNGD's headline pitch is performance per watt. The chip is rated at about 512 INT8 TOPS (and a comparable FP8 figure), with roughly 1.5 TB/s of memory bandwidth across its HBM3 stacks, and it draws about 150 to 180 watts depending on form factor. That power envelope is a fraction of the 1,000-plus watts drawn by the latest flagship data-center GPUs, which is the basis for FuriosaAI's claim that RNGD can deliver competitive inference throughput at far lower power. A single RNGD card can also be partitioned into 2, 4, or 8 isolated virtual NPUs for multi-tenant environments such as Kubernetes.
| RNGD specification | Value |
|---|---|
| Generation | FuriosaAI, second generation |
| Architecture | Tensor Contraction Processor (TCP) |
| Foundry / process | TSMC, 5 nm |
| Transistors | About 40 billion |
| Memory | 48 GB HBM3 (two stacks), about 1.5 TB/s, supplied by SK hynix |
| Compute | About 512 TOPS (INT8); up to 1024 TOPS (INT4) |
| Host interface | PCIe Gen5 x16 |
| TDP | About 150 to 180 W |
| Virtualization | Partitionable into 2, 4, or 8 isolated NPUs |
| Public debut | Hot Chips, August 2024 |
FuriosaAI also offers the NXT RNGD Server, a 4U air-cooled chassis holding eight RNGD cards. The company says five such servers fit in a standard air-cooled rack, delivering around 20 petaFLOPS (INT8) per rack, and that a rack of RNGD provides greater compute density than comparable H100-based systems while fitting within ordinary air-cooled power and thermal limits.
In early 2025 FuriosaAI became the subject of reported acquisition interest from Meta, which has been building its own custom AI silicon to reduce its dependence on Nvidia. On February 12, 2025, Forbes reported, as relayed by The Korea Herald and others, that Meta was in talks to acquire the Seoul-based startup, with a deal potentially finalized within that month.
On March 24, 2025, Korean outlet Maeil Business Newspaper reported, and Bloomberg and TechCrunch subsequently covered, that FuriosaAI had turned down the offer, which was widely reported at roughly $800 million. According to those reports, the talks collapsed over disagreements about post-acquisition business strategy and organizational structure rather than over price. The concern attributed to FuriosaAI was that, inside Meta, its engineers would have been redirected to build custom chips solely for Meta's own services, rather than continuing to develop a broadly available accelerator to compete with Nvidia and AMD. Both the rejection figure and the reasons trace back to Korean media reporting and should be read as reported rather than company-confirmed; FuriosaAI did not publicly disclose the offer terms. Speaking afterward to the press, Paik framed the decision around independence, saying the company wanted to "continue our mission."
FuriosaAI's most significant publicly announced customer is LG AI Research, the AI lab of South Korea's LG group, which adopted RNGD to serve its EXAONE family of large language models. The two organizations announced the partnership in July 2025, after what LG described as roughly two years of evaluation. FuriosaAI reported that in LG's testing RNGD delivered up to 2.25 times better LLM inference performance per watt than GPU-based systems. The Register, citing FuriosaAI's data, noted that the comparison was against Nvidia A100 GPUs and that LG ran the EXAONE 32B model across four RNGD cards in a tensor-parallel configuration, reaching roughly 50 to 60 tokens per second at batch size one. Kijeong Jeon, a product unit leader at LG AI Research, described RNGD as offering strong real-world performance, lower total cost of ownership, and straightforward integration. The companies said they would jointly offer the RNGD Server to enterprises deploying EXAONE across sectors such as electronics, finance, telecommunications, and biotechnology.
On September 11, 2025, at the opening of OpenAI's first office in Seoul, FuriosaAI and OpenAI staged a joint demonstration in which OpenAI's open-weight gpt-oss-120b model ran in real time on two RNGD cards using MXFP4 precision. FuriosaAI's chief technology officer, Hanjoon Kim, took part in the demonstration, which the companies presented as evidence that a large open model could run within the power budget of an ordinary enterprise data center rather than requiring high-power GPU clusters. The event was a public technical showcase rather than an announced commercial supply agreement.
Saudi Aramco was reported to have been among the parties testing RNGD. Press accounts and investor materials have also referenced interest from other large technology buyers, but beyond the LG AI Research supply deal and the OpenAI demonstration these remain reported or exploratory rather than confirmed commercial deployments.
FuriosaAI raised an early base of capital from Korean investors including Naver and DSC Investment, reaching a cumulative total of around $115 million by early 2025. The pivotal round came later that year. On July 30, 2025, the company announced a Series C bridge round of about $125 million (roughly 170 billion won), drawn from approximately 40 institutional investors. The round was based on a pre-money valuation of about 830 billion won (roughly $630 million) and pushed the company past a $1 trillion won (roughly $769 million) post-money valuation, making FuriosaAI one of South Korea's newest AI unicorns. The company said the new round brought its cumulative funding to roughly $246 million since its 2017 founding.
| Funding event | Date | Amount | Notes |
|---|---|---|---|
| Early rounds (cumulative) | Through early 2025 | About $115 million total | Investors included Naver and DSC Investment |
| Series C bridge | July 30, 2025 | About $125 million (about 170 billion won) | About 40 institutional investors; unicorn valuation; cumulative total about $246 million |
Investors named in the Series C included the Korea Development Bank, the Industrial Bank of Korea and its affiliates IBK Securities and IBK Venture Investment, Keistone Partners, and Kakao Investment. FuriosaAI said the proceeds would fund the mass production and go-to-market of RNGD and the development of a next-generation chip aimed at agentic and reasoning workloads. Reporting in 2026 indicated the company was seeking substantial additional capital to fund further expansion.
On January 27, 2026, FuriosaAI announced that RNGD had entered mass production, stating that it had received an initial 4,000 units from its manufacturing partners TSMC and ASUS. The company indicated it was scaling toward production on the order of tens of thousands of units per year. RNGD's commercial rollout was framed as a test of whether a Korean inference chip could win volume orders against entrenched GPU suppliers.
At its Renegade 2026 Summit in Seoul on April 2, 2026, FuriosaAI laid out its inference-focused strategy and named additional Korean partners alongside LG AI Research, including LG Uplus, Samsung SDS, and MegazoneCloud. Samsung SDS was reported to be preparing a subscription-based NPU service using RNGD. The company continued to argue that inference, rather than training, will account for the majority of new global AI data-center capacity over the rest of the decade, and that efficiency at the rack and total-cost-of-ownership level, rather than peak per-chip performance, will decide which accelerators data-center operators choose.
FuriosaAI is notable on several fronts. It is one of the most prominent attempts to build a credible alternative to Nvidia for AI inference from outside the United States and China, and a flagship of South Korea's ambition to develop domestic AI semiconductor capability beyond its existing strength in memory chips. Its TCP architecture is a genuine departure from the matrix-multiply-centric designs that dominate the field, and the company has backed the claim with peer-reviewed work and third-party validation from LG AI Research rather than relying on internal benchmarks alone. The reported rejection of Meta's roughly $800 million offer turned the company into a symbol of a Korean startup choosing independence over an early exit, a decision whose wisdom hinges on whether RNGD can convert design wins and demonstrations into sustained, high-volume revenue. As of mid-2026 the company had its anchor customer in LG, a high-profile demonstration with OpenAI, and chips in mass production, but had not yet disclosed the kind of blockbuster order that would confirm RNGD as a durable competitor to GPU-based inference at scale.