FuriosaAI closes $620M Series D at $2.9B valuation; RNGD inference cards in mass production (20k cards in 2026)

Why it matters

【Key points】

FuriosaAI (Seoul) on 2026-10-28 closed a $620M Series D at a $2.9B post-money valuation. International LPs: Cisco Investments, Infinitum, Jericho Capital, NeoFort Holding, QUDA Capital, Worldline Capital. Korea’s National Growth Fund ~$280M (KRW 370B); new Korean LPs: Export-Import Bank of Korea, UAMCO, Hanwha AM; seed backers NAVER Ventures, DSC, KDB Capital re-upped.

1. Production: RNGD inference cards in mass production — 20k cards in 2026; memory/fab locked to more than double in 2027. ~180W air-cooled PCIe; 8 cards → NXT RNGD Server (4U / 3 kW / 4 PFLOPS FP8 / 384 GB HBM3) for existing racks, no liquid cooling. Claims RTX PRO 6000-class throughput at same card count and ~2× users/kW.

2. Positioning: inference-only; ~100 tok/s/user; cites Gemma 4 and Qwen 3.8-27B. Architecture: Tensor Contraction Processor (TCP) + compiler. Mass production since 2026-01; customers include Samsung SDS (NPU-as-a-Service) and LG Group on-prem hybrid cloud.

3. Secondary: Bloomberg reports the step-up from ~$735M valuation in 2025.

【Verifiable】

Primary: https://furiosa.ai/blog/furiosaai-620-million-series-d

Newsroom: https://furiosa.ai/newsroom

Bloomberg via Techmeme

【Our read】

1. Background: inference hardware shifting from demos to deployability in existing power/cooling. Nvidia still leads training/high-end inference; enterprises and sovereign AI want air-cooled rack-friendly alternatives. A stated 20k-card 2026 run is Furiosa’s lab→supply-chain pivot; National Growth Fund fits Korea’s domestic AI-compute push.

2. Peer compare: vs Groq, Cerebras, Tenstorrent, Intel D-series — Furiosa sells drop-in 4U air-cooled tokens/kW vs RTX PRO 6000, not B200 clusters. Software maturity and independent benches remain the failure mode that sank prior non-Nvidia GPUs.

3. Impact: real volume puts Furiosa on the second-tier inference list (most relevant APAC/Korea on-prem). $2.9B still needs verified perf/TCO and 2027 capacity. Open: benches, ASP, contracts, open-model support.