---
title: "FuriosaAI Raises $620M at $2.9B to Scale Inference Chips"
description: "FuriosaAI raised a $620M Series D at a $2.9B post-money valuation. We examine RNGD's customer workflow, production claims, investor thesis and closest inference-chip rivals."
date: 2026-10-29T17:00:00.000Z
source: https://venturecapitaltracker.com/furiosaai-620m-series-d-2-9b-valuation
---

# FuriosaAI Raises $620M at $2.9B to Scale Inference Chips

> FuriosaAI raised a $620M Series D at a $2.9B post-money valuation. We examine RNGD's customer workflow, production claims, investor thesis and closest inference-chip rivals.

FuriosaAI has closed a **$620 million Series D at a $2.9 billion post-money valuation** to scale its RNGD inference accelerator. The Seoul company says South Korea’s National Growth Fund supplied KRW 370 billion, about $280 million, while international and existing investors filled out the round. No formal lead investor was named. The financing matters because Furiosa has crossed the line that stops many chip startups: RNGD entered mass production in January and is now serving production workloads, rather than remaining a design or benchmark project.

Korean reporting describes the round as KRW 820 billion. Using the European Central Bank’s October 29 reference rates, that is approximately $613 million; the company’s announcement rounds the transaction to $620 million. The same sources describe a KRW 3.8 trillion post-money valuation, approximately $2.84 billion at that rate, consistent with the company’s $2.9 billion headline.

## What customers buy from FuriosaAI

Furiosa designs accelerators for inference: the repeated work of turning a trained AI model into answers for users. Customers can buy RNGD PCIe cards and eight-card NXT servers for their own data centers, or rent RNGD capacity through Samsung SDS’s cloud service. Public card prices, revenue, backlog and gross margin are undisclosed.

Each RNGD card uses 180 watts and contains 48GB of HBM3 memory. An eight-card server is a 4U, air-cooled system drawing about 3kW. Furiosa’s commercial argument is therefore less about winning every raw-performance benchmark than fitting useful model throughput into existing racks without liquid-cooling retrofits.

The before-and-after workflow is clearest in Samsung SDS’s NPU-as-a-Service. A customer that would otherwise procure accelerator servers, integrate drivers and operate the hardware can select an instance with one, two, four or eight RNGD cards through Samsung Cloud Platform. Samsung handles the physical infrastructure and metering; the customer deploys its inference workload through familiar cloud operations.

Furiosa’s Upstage case study provides a more concrete production example. Upstage moved a 100-billion-parameter model used for Daum search summaries onto three NXT servers. Furiosa says the application integrated through an OpenAI-compatible endpoint, requiring mainly an endpoint change. It reports 1.3 times the performance per watt of an H200 server and says Upstage subsequently requested a tenfold capacity expansion. Those are company-produced case-study figures, not independently audited economics.

## The Series D and why investors moved now

The company named Cisco Investments, Infinitum, Jericho Capital, NeoFort Holding, QUDA Capital and Worldline Capital among the international participants. New Korean investors included the Export-Import Bank of Korea, UAMCO and Hanwha Asset Management. NAVER Ventures, DSC Investment, KDB Capital, KDB, Nextrans and Kakao Investment returned. Individual allocations beyond the National Growth Fund commitment were mostly undisclosed.

Furiosa previously raised a $125 million Series C in July 2025, taking disclosed cumulative funding to $246 million at that point. Bloomberg reports that the Series D quadrupled Furiosa’s valuation from the prior round’s $735 million.

The investor case is execution rather than architecture alone. Infinitum’s investment statement highlights the difficulty of moving an advanced chip into volume production and securing enterprise deployments. NAVER Ventures says Furiosa is moving from proving its technology to proving its business model. Furiosa plans to produce 20,000 RNGD cards in 2026 and more than double output in 2027 after securing memory and fabrication capacity. It is also developing a 2nm, multi-die third-generation accelerator with Broadcom and other partners.

## The closest inference-chip financings

[Positron AI raised $875 million at a $5 billion valuation](/2026-positron-ai-875m-series-c-5b-nea) in September 2026. It is the closest direct comparison because it also sells purpose-built inference silicon and argues that memory architecture and power efficiency can reduce serving costs.

[Etched raised $700 million at a $21 billion valuation](/2026-etched-700m-21b-jane-street) in August. Etched is more specialized around transformer inference, whereas Furiosa emphasizes programmability across changing open models. [Tenstorrent’s $1.5 billion Series E](/tenstorrent-1-5b-series-e-open-ai-compute) is broader, spanning AI processors, chiplets and intellectual-property licensing rather than inference deployments alone.

The financing buys Furiosa time and supply, but it does not prove repeatable unit economics. Nvidia’s software ecosystem, rapid GPU product cycles and customer reluctance to adopt another accelerator remain powerful barriers. Furiosa also faces semiconductor yield, HBM availability, customer concentration and long enterprise qualification cycles. The next decisive figures are paid card shipments, revenue, contracted backlog, average selling price, gross margin, production yield, uptime and independently reproduced tokens-per-watt results. A tenfold capacity request is promising; repeat orders and profitable deployments would show whether RNGD can become more than a sovereign alternative.

**By:** [Venture Capital Tracker](https://venturecapitaltracker.com/editorial-policy)

**Editorial note:** AI tools assisted with research, structure, or drafting. Venture Capital Tracker retains human editorial responsibility for factual accuracy, relevance, and source quality before publication.