> ## Content Index > Fetch the complete content index at: https://nexi.fund/llms.txt > Use this file to discover other available public pages before exploring further. # FuriosaAI's $620M Bet: Can a 180-Watt Korean Chip Take On Nvidia? - URL: https://nexi.fund/furiosaai-620m-series-d-2026/ - Published: 2026-10-31T10:30:08.000Z - Updated: 2026-10-31T10:30:08.000Z - Description: Nine years after its seed round, FuriosaAI closed a $620M Series D at a $2.9B valuation. RNGD is shipping to Samsung SDS, LG Group and Upstage. The next test is 40,000 cards in 2027, and a 2nm Gen 3 with Broadcom. - Author: Nexi.fund Labs - Tags: AI & Infrastructure, #mode-5, #hook-character, #track-A, #layout-asymmetric-long At 7:57 p.m. Eastern on 28 October, a press release from Seoul put a number on nine years of work: $620 million. FuriosaAI closed a Series D at a $2.9 billion post-money valuation. Its second-generation inference chip, RNGD, has been in mass production since January. The company plans to ship 20,000 cards this year and more than 40,000 in 2027. Most AI chip startups die between a working demo and a purchase order. FuriosaAI is past that gap. The round does not settle the harder question. Can a 180-watt Korean accelerator win repeat volume against Nvidia's software moat, at a valuation that already assumes it will? ## Nine years from a seed cheque to a state-backed round DSC Investment backed FuriosaAI at seed in 2017 and has joined every round since. NAVER came in through D2SF, its early-stage arm, and returned in this round through NAVER Ventures in Silicon Valley. KDB Capital, another seed investor, participated again. RNGD reached mass production in January 2026, nine months before the round closed. The company says it has secured memory and fab capacity to more than double output next year. That supply commitment matters more than the headline cheque for anyone planning a procurement cycle around the chip. The card is a 180-watt, air-cooled PCIe device. Eight of them make an NXT RNGD Server: 4U, 3 kilowatts, 4 petaFLOPS at FP8, 384 GB of HBM3\. FuriosaAI's pitch is that this fits racks and power budgets enterprises already have. ``` CHRONICLE: FuriosaAI ──────────────────────────── 2017 · 🌱 Seed from DSC, NAVER D2SF, KDB Capital 2025 · 🧠 RNGD taped out; $735M valuation in July 2026 JAN · 🏭 RNGD enters mass production 2026 MAY · 🤝 Gen 3 partnership with Broadcom announced 2026 JUL · 🌍 Equinix Lisbon deployment; Samsung SDS NPUaaS ◉ NOW · 💰 $620M Series D at $2.9B 🔥 NEXT · 📦 40,000+ cards targeted for 2027 ``` Company announcements and Jon Peddie Research, 2026. ## How the $620 million splits South Korea's National Growth Fund committed $280 million, about 45% of the round and the single largest cheque. It carries a public mandate to expand national AI computing infrastructure and the domestic semiconductor ecosystem. International investors added $110 million. Infinitum, a US firm with a long record in AI semiconductors, led that tranche. Cisco Investments, Singapore's QUDA Capital, Abu Dhabi's NeoFort Holding, Jericho Capital and Worldline Capital joined. Existing shareholders put in $125 million of follow-on capital. Korea Development Bank added $22 million, UAMCO $30 million, and the Export-Import Bank of Korea made its first direct venture investment. Hanwha Asset Management joined as a new investor. The round started in January with a target of 450 billion to 750 billion won, according to DigitalToday. It closed at 820 billion won. Demand, not plan, set the size. $620M Series D, Oct 2026 #### FuriosaAI Series D At a $2.9 billion post-money valuation, roughly four times the July 2025 mark. The National Growth Fund put in $280 million. *BusinessWire, 2026* ## Why inference is the wedge Training builds the model. Inference serves it, every time a user types a prompt. That second job runs continuously, so power and memory bandwidth dominate the bill. RNGD is built for that workload. It uses TSMC's 5nm process at 1.0 GHz and reaches 256 TFLOPS in BF16, 512 TFLOPS in FP8 and 512 TOPS in INT8\. Two HBM3 modules supply 1.5 TB/s of bandwidth and 48 GB of memory. A single card can be partitioned into as many as eight independent instances through SR-IOV. The architecture is the differentiator FuriosaAI sells. GPUs and most accelerators break model operations into fixed-size, two-dimensional matrix multiplications. TCP treats tensor contraction, a higher-dimensional operation, as the basic unit of work across hardware and compiler. Fewer data moves, more compute per watt. ## The engineering bet: fewer watts per token The company claims a concrete result. RNGD matches RTX PRO 6000 throughput at the same card count and serves twice as many users per kilowatt. In one measured configuration it delivered 78 tokens per second per user against 85 on the Nvidia card. That is about 8% slower. Both claims come from FuriosaAI's own sweep, SDK 2026.4 against vLLM 0.28.0\. No independent benchmark is published yet. Treat the parity claim as a vendor number until someone outside the company reproduces it. | Metric | RNGD | RTX PRO 6000 | | ----------------------------- | ------- | ------------ | | **Card power** | 180 W | 2× or more | | **Throughput per card** | Matches | Baseline | | **Users per kW** | 2× | Baseline | | **Tokens/s/user, one config** | 78 | 85 | FuriosaAI company benchmark, SDK 2026.4 vs vLLM 0.28.0, four-card instances scaled to eight-card servers. Not independently verified. Software is moving the number fast. SDK 2026.4, released on 16 October, raised token generation per watt on 18 models. On gpt-oss-120B the gain was 138% over the previous release; on gpt-oss-20B it was 57%. > We are entering a new phase of bringing our technology to the global AI infrastructure market at scale. By accelerating the commercialization of RNGD and advancing the development of Gen 3, we aim to help shape the next generation of AI infrastructure.— June Paik, co-founder and CEO, FuriosaAI ## Named customers, mostly in Korea Samsung SDS launched a commercial NPU-as-a-Service on RNGD in July. LG Group runs EXAONE deployments on the chip. Upstage uses RNGD for live search summarisation, and Daum is also named as a deployment. Those are production workloads, not pilots. They also cluster in one domestic market. FuriosaAI opened a Singapore subsidiary in September for Southeast Asia and deployed RNGD at Equinix's Lisbon data centre in July. The Series D is meant to turn those footholds into repeat orders. 🎯 $620 million buys volume, not a verdict. RNGD is in production and in named accounts. The valuation quadrupled in fifteen months. The 2027 target doubles shipments, not market share. Gen 3 with Broadcom is the hyperscale test. It has no public tape-out date. ## Gen 3 is the hyperscale test Gen 3 is a 2nm, multi-die design with HBM4 and HBM4E memory, developed with Broadcom and other partners. FuriosaAI announced the partnership in May. The target is frontier agentic models in large clusters. That is a different sale. Enterprise buyers purchase a card that fits an existing rack. Hyperscalers buy integrated systems, software and a multi-year roadmap. Nvidia sells all three today. No public tape-out date exists for Gen 3\. The release names no process risk, no customer and no volume target. That gap is where the valuation is most exposed. > Taking an advanced AI chip from design to volume production is an exceptionally difficult undertaking, and few companies have demonstrated the ability to do so successfully. FuriosaAI has gone further, securing production deployments with leading enterprises such as Samsung SDS and LG Group.— John Yetimoglu, chief investment officer, Infinitum ## The delivery test starts now Forty thousand cards in 2027 is roughly 5,000 eight-card servers, before any shift toward denser configurations. Getting there requires repeat orders from customers who have already qualified the chip. That is a supply-chain problem first and a sales problem second. As we wrote in October, [Nvidia funded d-Matrix, a startup built to compete with it](https://nexi.fund/d-matrix-nvidia-nvlink-inference-2026/). The pattern repeats here at a larger scale: the incumbent hedges, the challenger gets capital, and shipments decide who was right. The $2.9 billion valuation is a bet on 2027 output. Watch the card count. Everything else is commentary. [ FuriosaAI secures $620 million Series D The primary release with the round size, investor list, production targets and customer deployments. BusinessWire ](https://www.businesswire.com/news/home/20261028655033/en/FuriosaAI-Secures-%24620-Million-Series-D-to-Meet-Surging-Enterprise-AI-Inference-Demand-and-Accelerate-Next-Gen-Silicon?ref=nexi.fund) The source for the funding terms and shipment guidance used above. [ FuriosaAI raises $620M to attack the data centre power bottleneck Independent context on the TCP architecture and the efficiency claim behind RNGD. SiliconANGLE ](https://siliconangle.com/2026/10/28/energy-efficient-inference-chip-developer-furiosaai-raises-620m-to-solve-the-data-center-power-bottleneck?ref=nexi.fund) Useful for the architecture explanation and the Nvidia comparison. [ FuriosaAI raises $620 million for RNGD A detailed breakdown of the round's commitments and the open questions around Gen 3. Jon Peddie Research ](https://www.jonpeddie.com/news/furiosaai-raises-620-million-for-rngd/?ref=nexi.fund) The clearest accounting of who put in what, and what the release does not say.