FuriosaAI announced a $620 million Series D on 28 October at a $2.9 billion post-money valuation, backing an expansion of its RNGD inference accelerators. The company says it expects to produce 20,000 cards in 2026 and has secured memory and fabrication capacity to more than double that volume in 2027.[1]
The funding is intended for wider RNGD deployments and development of a third-generation chip with Broadcom and other partners. The annual card figure is a production target; the company’s capacity statement for next year is not a report of customer deliveries.[1]
There is a concrete manufacturing baseline. In February, LG CNS reported that Furiosa had received 4,000 RNGD cards from TSMC in January. LG CNS also described technical validation of the chips for its AgenticWorks enterprise platform as the first step in a broader cooperation agreement.[3]
Customers can rent the hardware
Samsung SDS provides a commercial route already in operation. Its 20 July launch put RNGD on Samsung Cloud Platform as a subscription service, with customers selecting one, two, four or eight accelerators. The service connects with the platform’s storage, computing and networking products.[2]
That arrangement changes the buying decision. A customer can start with allocated accelerator capacity without purchasing and installing an NPU server. The cloud provider handles the infrastructure investment, while the user buys access. This is useful distribution for a chip vendor seeking customers beyond companies prepared to operate its hardware themselves.[2]
Power figures need a system boundary
Furiosa’s NXT RNGD appliance is a 4U, air-cooled server with eight accelerators. Its specification lists 3,000 watts of system power, alongside two AMD CPUs and one terabyte of DDR5 system memory. That full-server figure is the relevant starting point for a facility’s equipment power budget.[4]
Furiosa advertises up to twice the tokens per kilowatt in its Qwen3-32B FP8 comparison. Those tests use four-card instances, then scale results to eight-card server power envelopes: 3.0 kilowatts for NXT RNGD and 6.5 kilowatts for the RTX PRO 6000 comparison. The claimed advantage therefore applies to the specified model and serving conditions.[4]
For the expansion, manufacturing and distribution therefore have different roles. More cards create supply; a subscription channel lets customers take smaller allocations and build demand around running applications. Samsung’s existing service gives Furiosa a way to reach that second group as production grows.[2][1]
Sources & further reading
- FuriosaAI closes $620 million Series D to scale efficient, production-ready AI inference FuriosaAI · 28 Oct 2026 · Accessed 29 Oct 2026
- Samsung SDS Officially Launches Korean NPU-Based NPUaaS Samsung SDS · 20 Jul 2026 · Accessed 29 Oct 2026
- LG CNS and FuriosaAI introduce NPUs for agentic AI, targeting public-sector AI transformation (English translation) LG CNS, via LG newsroom · 5 Feb 2026 · Accessed 29 Oct 2026
- Furiosa NXT RNGD Server FuriosaAI · Accessed 29 Oct 2026