FuriosaAI raises $620 million Series D at a $2.9 billion valuation to scale RNGD chips

What FuriosaAI announced

FuriosaAI said on 29 October 2026 that it had closed a $620 million Series D to scale production of its inference accelerator. Coverage from Pulse 2.0, Dealroom and Converge Digest puts the post-money valuation at about $2.9 billion, roughly four times the $735 million figure from the company's July 2025 round. Korean outlet The Elec reported the valuation as about 3.8 trillion won.

South Korea's National Growth Fund, a state-backed vehicle created to finance strategic industries, was the largest investor with $280 million. Reports break down the rest as about $110 million from international investors and $125 million from existing shareholders, a group that includes NAVER, Korea Development Bank, DSC Investment and Kakao Investment. Some coverage dated the announcement 28 October Korean time; the company's post is the reference. One outlet described it as the largest Series D ever raised by a South Korean semiconductor company, a claim we could not check independently.

The investor mix is notable. The National Growth Fund's 80 million is about 45 percent of the round, which makes the Korean state the anchor of a private chip company at a stage where startups usually rely on venture and strategic investors. Strategic backers such as NAVER and Kakao, which run large Korean AI services, are also potential customers. FuriosaAI has not said whether the government money comes with conditions on where it manufactures or whom it sells to.

RNGD and the production plan

FuriosaAI's current product is RNGD, pronounced renegade, its second-generation chip. The company's product page rates it at 512 TFLOPS in FP8, with 48 GB of HBM3 memory, 256 MB of on-chip SRAM and a thermal design power of 180 W, and sells it as a passively cooled PCIe card and in an eight-card server that FuriosaAI says delivers 4 PFLOPS in about 3 kW. The pitch is power efficiency: the chip is designed to run large language models and other inference work in existing air-cooled racks, where many data centers cannot add the power and liquid cooling that the largest GPU systems now require.

RNGD entered mass production in January 2026. With the new money, FuriosaAI says it aims to ship 20,000 accelerators this year and more than 40,000 in 2027. FuriosaAI lists Samsung SDS, LG Group and Upstage among RNGD customers. The company has not published revenue, and those unit targets are its own plans rather than booked orders.

A third chip with Broadcom

The round also funds FuriosaAI's third-generation accelerator, which the company is developing with Broadcom and other technology partners for large-scale AI infrastructure and agentic AI workloads. Reports describe a 2 nm, multi-die design with HBM4 or HBM4E memory, and one report says samples are not expected until 2028. Broadcom already designs custom accelerators for several hyperscalers, so the partnership gives FuriosaAI access to packaging and interconnect technology that a startup would struggle to build alone.

For Broadcom, the deal adds another name to a custom accelerator business that already drives much of its growth; the company reported 6.7 billion in AI revenue for its third quarter. For FuriosaAI, working with a large partner reduces technical risk but also means sharing margins and depending on Broadcom's capacity and priorities. Neither company has disclosed the commercial terms or whether any hyperscaler has committed to the chip.

That timeline matters because the market is moving fast. Nvidia ships new rack-scale systems each year, and the large cloud providers increasingly use their own inference chips. A third-generation Furiosa part sampling in 2028 would compete with whatever those vendors offer then, not with today's products.

Why it matters

Inference, the work of running models after they are trained, is where AI spending is growing fastest as agents and long-context applications multiply token volumes. That has opened room for specialist chip companies such as Cerebras, Groq and FuriosaAI, which argue that GPUs built for training are not the cheapest way to serve models. The state's role in this round also fits South Korea's wider industrial policy, which aims to turn the country's memory strength, led by SK hynix and Samsung, into a broader position in AI hardware. Our coverage of SK hynix's record HBM revenue shows how much of the current boom already flows through Korean suppliers.

Skeptics point out that funding does not guarantee data center scale. Selling tens of thousands of cards means winning software support from model developers, passing qualification at large operators and competing on price with vendors that have far larger volumes. FuriosaAI's numbers for 2026 and 2027 will show whether RNGD can move beyond early adopters. The round is listed in our AI funding tracker, and more deals sit in the AI business section.

Questions people ask

How much did FuriosaAI raise?

$620 million in a Series D at a valuation of about $2.9 billion, announced on 29 October 2026. South Korea's National Growth Fund was the largest investor with $280 million.

What is FuriosaAI's RNGD chip?

An AI inference accelerator that FuriosaAI rates at 512 TFLOPS in FP8 with 48 GB of HBM3 memory and a 180 W thermal design power, sold as a PCIe card and in eight-card servers, per its product page.

What will FuriosaAI do with the money?

Scale RNGD production to 20,000 units in 2026 and more than 40,000 in 2027, and develop a third-generation accelerator with Broadcom for hyperscale data centers.

Sources

1. FuriosaAI closes $620 million Series D to scale efficient, production-ready AI inference (FuriosaAI, 29 Oct 2026)
2. Furiosa RNGD PCIe Card (FuriosaAI, 29 Oct 2026)
3. FuriosaAI Raises $620 Million Series D At About $2.9 Billion Valuation (Pulse 2.0, 29 Oct 2026)
4. FuriosaAI Raises $620M at $2.9B Valuation for AI Silicon (Converge Digest, 29 Oct 2026)
5. FuriosaAI Raises 820 Billion Won in Series D Funding, Reaches 3.8 Trillion Won Valuation (The Elec, 29 Oct 2026)