Private companySeries D

FuriosaAI

silicon · KR

furiosa.ai

Research revised Nov 20, 2026

Offering and investment case

FuriosaAI supplies RNGD inference hardware and software; the question is whether supported models meet production quality and latency requirements at a competitive complete-system cost.

The company distinguishes an RNGD PCIe card, NXT RNGD server and furiosa-llm serving stack. An OpenAI-compatible API can simplify application connection; it does not establish identical model support, operators, output behavior or migration effort for every GPU workload.

Research mapping

Physical AI / AI Factory & Cloud Training Infrastructure (L1) / silicon

Evidence boundary

Public-source analysis is included below. Issuer descriptions identify the offering; they do not independently establish commercial scale or durable technical advantage.

A capable compiler/runtime and supported models can make alternative silicon usable. Defensibility depends on sustained application performance and support, rather than peak arithmetic throughput or a nominal card-power rating alone.

The Furiosa site offers hardware or endpoint evaluation and describes an API-compatible serving interface. Check the intended model/version, precision, context, batching, decoding features and software dependencies. An application can send the same request shape while receiving different quality, throughput or tail latency. A new SDK release is not evidence that all previous unsupported models now run.

Compare accepted tokens or completed tasks at the same quality and latency target, including host CPUs, memory, networking, cooling and utilization. Card watts and wall power are different denominators. GPU inference, another ASIC and optimized existing hardware are relevant alternatives. Include porting, software maintenance, fleet operations and hardware supply/service. Repeatable customer workloads and repeat orders strengthen the case; a narrow benchmark, missing kernels or low utilization can weaken it. Public marketing and funding announcements do not establish independent performance leadership, supplier profitability or customer-specific savings.

Substantive research revision: 2026-11-20. AI-assisted public-source research; no human approval attested.

Data Center Operators

Samsung SDS, LG AI Research, Upstage, Megazone Cloud, Equinix, Saudi Aramco

June Paik — Co-founder & CEO — Former Samsung and AMD engineer who founded FuriosaAI in Seoul in 2017 to build specialised inference silicon.

Hanjoon Kim — Co-founder — Worked with Paik at Samsung before leaving to co-found FuriosaAI.

Source
Reported valuation
US$2,900,000,000 (Post-money) As of 2026-10-28

Issuer explicitly states post-money.

Source 1
Disclosed capital
US$246,000,000 As of 2025-07-30

Historical issuer total at Series C; excludes the later Series D.

Source 2
Financing announcement
US$620M Series D As of 2026-10-28

Issuer announcement.

Source 1

The current US$620M Series D and US$2.9B post-money valuation are issuer statements. The separately dated US$246M historical total predates Series D. KRW370B participation described within Series D is not added again on top of the full round.

  1. FuriosaAI Series D announcement — Company announcement; published 2026-10-28
  2. FuriosaAI Series C funding announcement — Company announcement; published 2025-07-30

Financing sources checked 2026-11-20. Figures describe the cited reports and dates; they do not establish current market value, available cash or investment returns.

Mapped funds and investors

Direct relationships supported by published evidence.

Reported holdings as-of

No sourced relationships mapped yet.

Disclosed investment history

No sourced relationships mapped yet.

Showing up to four relationships in each group. ETF weights describe the fund’s portfolio, not ownership of this company. Private investment history does not establish current ownership.

The weekly brief

Get a weekly digest of what’s changing.

Research updates, thesis changes and company developments. One email a week.