Today we’re announcing Furiosa’s $620 million Series D at a post-money valuation of $2.9 billion, with participation from international investors Infinitum, Jericho Capital, QUDA Capital, Cisco Investments, NeoFort Holding, and Worldline Capital. Our RNGD inference accelerator is in mass production: we’ll produce 20,000 cards this year, and we’ve secured the memory and fab capacity to more than double that in 2027. This capital lets us bring RNGD to many more enterprises worldwide and build what comes next.
We founded FuriosaAI in 2017, a year after AlphaGo defeated Go champion Lee Sedol in Seoul. It was clear to us then that artificial intelligence would be a transformative technology and that serving AI to billions of people would require a new kind of chip. To be economically and environmentally sustainable, AI infrastructure would need to meet three crucial needs:
- Performance: the most tokens from every chip and every rack.
- Efficiency: power is expensive and scarce, and a chip that draws too much raises total cost of ownership and limits where it can be deployed.
- Programmability: high throughput on new models and techniques, without months of hand-tuning kernels.
Building AI inference accelerators that deliver all three of these requirements has been Furiosa’s priority ever since.
We are grateful to all our investors – new and returning, Korea-based and global – for their confidence in our team. In addition to the international investors noted above, Korea’s National Growth Fund committed KRW 370 billion (about $280 million); the Export-Import Bank of Korea, UAMCO, and Hanwha Asset Management joined as new investors; and all three of our original seed investors, NAVER (through NAVER Ventures), DSC Investment, and KDB Capital contributed again, along with Korea Development Bank, Nextrans, Kakao Investment, and many of our early investors.
Performance, efficiency, programmability
Formula One fixes the engine's displacement and asks engineers for the most performance from a set amount of fuel. That is how we think about our RNGD chip, which was built to deliver the most tokens possible from less than 200 watts. RNGD comes as a 180-watt, air-cooled PCIe card. Eight of them make the NXT RNGD Server: 4U, 3 kW, 4 petaFLOPS (FP8), 384 GB of HBM3, designed to go into the racks and power budgets enterprises already have, with no cooling retrofit. RNGD matches RTX PRO 6000 throughput at the same card count and serves twice as many users per kW of power.
We were equally deliberate about what RNGD is for. It’s built specifically for inference because that is where efficiency is non-negotiable. You train a model once, but run it many billions of times in production.
RNGD targets models running at up to about 100 tokens per second per user, because that’s what matters most for enterprise workloads. It excels at running the latest state-of-the-art open models like Gemma 4 and Qwen 3.8-27B. And, as described below, RNGD is highly programmable.
RNGD deployed in production
The most important test for any AI accelerator is meeting customers’ needs in large-scale deployments. This is only possible when your silicon moves past testing and into volume production – a significant logistical and business challenge with semiconductor supply tight and demand soaring.
RNGD reached this mass production milestone in January 2026 and we’re now deploying RNGD with major customers across AI cloud providers, sovereign AI initiatives, AI-native companies, and enterprises building on-premises and hybrid cloud infrastructure. RNGD is running some of the most capable open-weight AI models for a wide range of our customers’ AI applications.
For example, Samsung SDS has launched its NPU-as-a-Service offering powered by RNGD, while LG Group companies are deploying RNGD in their on-premises hybrid cloud environments.
Thank you to these customers. They deployed real inference workloads on a second-generation chip from a company most of the industry was still learning about, and their feedback has helped us tremendously.
The architecture behind RNGD and our next generation
Everything we build rests on one architectural innovation, the Tensor Contraction Processor (TCP), which we built alongside our compiler and software stack. This is what enables performance to coexist with efficiency and programmability. GPUs and most other AI accelerators break model operations into fixed-size, two-dimensional matrix multiplications. TCP makes tensor contraction, a higher-dimensional operation, the basic unit of work for both the hardware and the software.
Our compiler can plan across a whole model and, thanks to deterministic execution on the silicon, minimize energy-wasting data movement. Reconfigurable compute slices keep the chip flexible enough to run new models and new optimization techniques, to excel at both prefill and decode, and to keep utilization high no matter how big or small the batch size.
That is what TCP delivers, today and for the models to come: performance, efficiency, and programmability.
Scaling production, software development, and Gen 3
This Series D round is focused on two priorities: scaling RNGD production and deployments globally, and building our third-generation chip to serve frontier agentic models in large clusters. With development underway now with Broadcom and other partners, Gen 3 will leverage our TCP architecture with a 2nm, multi-die design and HBM4/4E memory.
We will continue our aggressive cadence of software releases to make sure we’re serving our customers’ needs. We’ve shipped four major SDK releases this year, adding hybrid batching, which combines prefill and decode requests in one batch, prefix caching, and native Kubernetes support. SDK 2026.4, released earlier this month, raises token generation per watt across 18 models compared with 2026.3 – including by 138% on gpt-oss-120B. It also expands vision-language and embedding model support.
At the PyTorch Developer Conference earlier this month, we detailed how customers can deploy in PyTorch as they would on GPUs, and, when they need more granular control, easily write their own kernels in our Tensor Contraction Language and Virtual ISA. These software innovations let customers get more tokens from Furiosa’s hardware without slowing them down or changing how they deploy.
Building for the world’s inference needs
Even as the field of AI continues to advance, we are confident that the needs we are solving for will remain essential: Data centers will need to run the most capable AI models at massive scale. Frontier models will grow more general and more capable, while domain-specialized models will be refined through inference-heavy post-training. Physical AI systems will perceive the world and respond in real time.
All of this points in one direction: the world will need far more inference compute, and it will need to maximize the number of tokens produced per watt. Every facility – from hyperscalers with hundreds of racks to on-prem deployments serving a few dozen employees – will be constrained by the cost and limited availability of power.
We’re looking forward to sharing more customer wins, more hardware roadmap details, and more software advances soon.
If you want to see RNGD running your own inference workloads, please reach out here.
And if you want to build the future of AI inference with us, we are hiring in Seoul, Silicon Valley, Lisbon, and Singapore.
What our investors are saying
“Taking an advanced AI chip from design to volume production is an exceptionally difficult undertaking, and few companies have demonstrated the ability to do so successfully,” said John Yetimoglu, Chief Investment Officer with Infinitum. “FuriosaAI has gone further, securing production deployments with leading enterprises such as Samsung SDS and LG Group. This combination of technological innovation and commercial execution gives us strong conviction in FuriosaAI's ability to scale RNGD globally and deliver on its ambitious Gen 3 roadmap.”
“We look for the moment a technology company moves from proving its technology to proving its business. FuriosaAI is at that point now, with its chip in volume production and in deployment with customers," said Yong Jung Park, General Partner of NAVER Ventures. “As a Silicon Valley-based investor, helping the company grow in North America is where we can add the most value. We're proud to back June and the team as they take Furiosa to the global stage.”
Written by
The Furiosa Team
