South Korea's FuriosaAI raises $620M to take on Nvidia in AI chips

South Korean AI chip startup FuriosaAI has raised US$620 million in its latest funding round, quadrupling its valuation from the previous round to US$2.9 billion, DIGITIMES reports. In local currency, the company announced on October 29 that it had completed an 820 billion won Series D round at a valuation of approximately 3.8 trillion won, according to a TechFlow report citing Yonhap News Agency.

The round is pitched as fuel for a challenge to Nvidia, and the money is earmarked for the global commercialization of FuriosaAI’s second-generation AI chip, RNGD, and a third-generation AI accelerator co-developed with Broadcom. For developers who mostly meet AI hardware as a line item on an inference bill, it is another sign of capital flowing toward alternatives in the chips that run models, not just the ones that train them.

What FuriosaAI raised and from whom

FuriosaAI launched the financing in January this year with an original target of 450 to 750 billion won; the final 820 billion won exceeded that range. The investor list is heavy on South Korean state and institutional money:

• National Growth Fund: 370 billion won, the largest investor in the round.
• Existing shareholders: 170 billion won combined, including 48 billion won from Nextrans and 30 billion won from the Industrial Bank of Korea.
• Follow-on investors: NAVER, DSC Investment and KDB Capital.
• New domestic investors: UAMCO invested 40 billion won, the largest amount among new domestic backers, and the Korea Export-Import Bank made its first direct investment in a startup.
• Overseas investors: US-based Infinitum made the largest overseas investment; Singapore’s QUDA Capital Partners and Cisco Investments also participated.

Item	Series D (2026)
Amount	820 billion won; US$620 million
Valuation	About 3.8 trillion won; US$2.9 billion
Original target	450 to 750 billion won
Round opened	January 2026
Announced complete	October 29, 2026
Largest investor	National Growth Fund, 370 billion won
Use of funds	RNGD commercialization; third-gen accelerator with Broadcom

How the valuation jumped fourfold

DIGITIMES characterizes the new valuation as four times the previous round’s. The last detailed public picture of FuriosaAI’s fundraising comes from April 2025, when ChosunBiz reported on a Series C bridge round. At that time, the company had originally planned to raise 70 billion won at a pre-money valuation of 830 billion won, then lifted its target to 90 billion won because of investor demand. ChosunBiz reported that, if that 90 billion won round closed, the post-investment valuation would reach 920 billion won.

Against that 920 billion won projection, a valuation of about 3.8 trillion won is roughly four times higher, consistent with the DIGITIMES description. The sources do not confirm the final closing figures of the 2025 bridge round, so the comparison rests on the projected number ChosunBiz reported at the time.

The jump also clears a milestone the market was watching. In 2025, ChosunBiz noted growing expectations that FuriosaAI would reach “unicorn” status, defined in that report as a valuation above 1 trillion won, and quoted a person in the venture capital industry predicting the next round would recognize a valuation of 1 trillion won. The Series D valuation lands well above that mark.

What FuriosaAI builds: inference chips

FuriosaAI is a fabless startup, meaning it designs chips and relies on others to manufacture them. It was founded in 2017 and focuses on AI chips specialized for inference, the stage where a trained model such as a large language model (LLM) answers requests. ChosunBiz described it as a company that designs AI semiconductor chips “like NVIDIA”.

The company’s product line, as the sources describe it, spans three generations:

1. First generation: the Warboy chip, shown in a server image in the 2025 ChosunBiz report.
2. Second generation: RNGD, the chip the new funds will commercialize globally. ChosunBiz’s translated 2025 report referred to the company’s second-generation chip as “Renegade2” and said it is equipped with HBM3, a type of high-bandwidth memory used to feed data to the processor fast enough to run LLMs.
3. Third generation: an AI accelerator co-developed with Broadcom, which the new round will also fund.

The sources do not give performance figures, power draw, memory capacity or pricing for any of these chips, and they do not include benchmark comparisons against Nvidia hardware. Any claim about how RNGD stacks up against competing accelerators would need sources beyond these reports.

From 2025 skepticism to a state-backed round

The road to this round was not smooth. ChosunBiz reported in 2025 that the second-generation chip had at one point faced investor skepticism over performance concerns. The same report said a rumor that Meta had begun negotiations to acquire FuriosaAI became a turning point for investor sentiment. That acquisition talk was reported as a circulating rumor, and none of the sources confirm any deal; the latest round shows FuriosaAI raising as an independent company.

Political support was already part of the story in 2025. ChosunBiz reported that investors were optimistic about policy benefits for AI semiconductors as a national strategic industry, and that Lee Jae-myung, then declaring his presidential candidacy, chose FuriosaAI for his first public schedule and met CEO Baek Joon-ho to discuss research and development investment. In the Series D, the single largest check came from the National Growth Fund, and the Korea Export-Import Bank chose FuriosaAI for its first direct startup investment.

Commercial traction in 2025 was early. ChosunBiz reported that FuriosaAI had begun technology verification, or proof-of-concept testing, with customers including LG AI Research and Saudi Arabia’s Aramco, and that revenue stood at 3 billion won at the time. The sources do not give updated revenue or customer figures for 2026.

What it means for builders

Most developers reading this never buy an accelerator directly; they pay per token or per request to an API provider, which in turn pays for the hardware underneath. Inference chips matter to that bill because they determine how much it costs a provider to serve each response. A well-funded company focused specifically on inference hardware adds one more potential supplier in a market where DIGITIMES frames the competition squarely against Nvidia.

Some practical points from what the sources do and do not say:

• No developer access details yet. The sources do not say whether RNGD is available through any cloud provider, hosted API or developer program. The funding’s stated purpose is global commercialization of RNGD, which implies wider availability is the goal, but no timeline or channel is given.
• Software support is unknown. Moving a model from one chip family to another usually depends on compiler and framework support. None of the sources address FuriosaAI’s software stack, so builders cannot yet judge how easily their existing models would run on it.
• The Broadcom accelerator is a future product. The third-generation chip is described as co-developed with Broadcom and funded by this round, but the sources give no specifications or launch date.
• Customer pilots are the signal to watch. In 2025 the company had proof-of-concept work with LG AI Research and Aramco. Production deployments by model providers would matter more to API users than funding totals.

Open questions and limits of the reporting

This article rests on news coverage rather than a primary company document. The detailed investor breakdown comes from TechFlow, which attributes it to Yonhap News Agency reporting on the company’s October 29 announcement. The dollar figures and the “quadruples” framing come from DIGITIMES, whose full article is behind a paywall; only its opening summary was available. The sources report the amount in won and in dollars separately and do not state the exchange rate used.

Several things remain unaddressed in the sources:

• How RNGD performs against Nvidia or other inference accelerators on any public or independent benchmark.
• Current revenue, shipment volumes or named production customers in 2026.
• Where RNGD is manufactured and whether supply is secured for the global rollout.
• The timeline for the Broadcom co-developed third-generation accelerator.
• The size of the stake held by state-linked investors after the round, and what, if any, conditions came with that money.

What to watch next

The next markers are commercial, not financial. FuriosaAI now has to turn the round into RNGD shipments outside South Korea, since global commercialization is the stated use of the money. Watch for named customers moving from pilots to production, any independent benchmark results for RNGD, and the first technical details of the Broadcom accelerator. For API builders, the most relevant signal will be whether any hosting provider starts serving models on FuriosaAI hardware, and at what price relative to Nvidia-based capacity. Until those details appear, the funding shows strong investor and state backing for a Nvidia alternative in inference, but not yet how well that alternative performs.