FuriosaAI Raises $620 Million: RNGD Production and the Gen3 Roadmap

FuriosaAI announced a $620 million Series D on October 28 to expand its AI inference business and develop its next chip. The company says it will produce 20,000 RNGD cards in 2026 and has secured memory and fabrication capacity for more than twice that number in 2027.

This is not a new RNGD launch. It is a manufacturing and product-roadmap update for an existing alternative to general-purpose GPU infrastructure.

What the funding announcement changes

The company's announcement values FuriosaAI at $2.9 billion after the round. It says the proceeds support production, deployments and its Gen3 accelerator, which is being developed with Broadcom using a 2nm process, a multi-die design and HBM4/4E memory.

Gen3 remains a development programme. The announcement does not provide a customer availability date or public pricing, so its proposed architecture should not be treated as hardware buyers can deploy today.

The figures also need careful reading. Planned production and secured capacity are not the same as cards shipped, installed or generating revenue. The announcement describes intended supply, not a verified count of completed customer deployments.

What RNGD offers today

FuriosaAI's NXT RNGD Server specifications describe an air-cooled, 4U inference system containing eight RNGD accelerators. The listed configuration provides 384 GB of HBM3 accelerator memory, 12 TB/s of aggregate memory bandwidth and approximately 4 PFLOPS of peak FP8 compute at 3 kW system power.

Those are product specifications, not proof that every model runs faster or more cheaply than on a competing server. A useful comparison needs the same model, acceptable output quality, latency target and serving workload.

The physical configuration is nevertheless relevant. A buyer evaluating an air-cooled system can assess it against an existing rack's power and cooling limits. That is a different purchasing question from selecting the highest peak performance accelerator, particularly where a facility upgrade would delay deployment.

Software support determines how much hardware is useful

FuriosaAI's recent SDK 2026.4 release provides a useful view of the work behind the hardware. It includes model-support improvements and experimental hierarchical KV caching, which is disabled by default.

A KV cache stores intermediate attention data so a language model does not need to recompute all of it during generation. Managing that data across memory tiers can affect the cost and responsiveness of long-running inference. The experimental status matters: a feature described in release notes should not automatically be assumed production-ready for every workload.

For an operator, the questions are specific. Does the intended model compile and run correctly? Which quantisation formats are supported? What happens under concurrent requests, long prompts and strict latency limits? How difficult is it to update the model or integrate the server with existing monitoring?

A hardware alternative becomes easier to adopt when those answers are predictable. Otherwise, a lower power figure can be offset by engineering work that the initial specification sheet does not show.

What to watch after the round

The funding gives FuriosaAI more room to turn available manufacturing capacity into repeatable deployments while developing another architecture. Those are related but separate execution problems: supplying more of the current product and bringing a new one to market.

The most useful follow-up evidence would be customer workloads with disclosed serving conditions, sustained utilisation and an understandable migration path. That would show where the existing RNGD system is a practical choice, rather than asking buyers to make a decision based on the promise of Gen3.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical teams turn hardware, infrastructure and AI product updates into clear stories that developers and founders can understand and share.

Book a 30-minute call.