FuriosaAI Raises ₩820 Billion… The AI Chip Battle Is Shifting to Customers’ Sites

Date
Empty
Classification
  1. Investment/Funding
작성자
에디터 L
FuriosaAI has completed its Series D funding round. The funds will be used to expand the supply of RNGD, its semiconductor for AI inference, strengthen its software, and develop next-generation chips. The next challenge is moving beyond building the technology and integrating it into real-world customer operations.
South Korean AI semiconductor company FuriosaAI has raised KRW 820 billion in Series D funding. The English announcement released by the company on October 28 put the investment at $620 million and the post-money valuation at $2.9 billion. In its domestic announcement on October 29, it gave the post-money valuation as approximately KRW 3.8 trillion.
This funding will help the company supply its second-generation chip RNGD, also known as Renegade, which is already in mass production, to more businesses. FuriosaAI is simultaneously pursuing growth in its overseas customer base, software development, and development of a third-generation AI accelerator. As notable as the size of the investment is the fact that these three efforts are linked as a single business challenge.

From making chips to supplying them reliably

The National Growth Fund invested KRW 370 billion in this round. International participants included Infinitum, Jericho Capital, QUDA Capital, and Cisco Investments. Early investors such as Naver, DSC Investment, and KDB Capital also took part in the follow-on investment.
RNGD entered mass production in January this year. In its official announcement, FuriosaAI cited Samsung SDS’s NPU service and LG Group’s private and hybrid cloud environments as real-world deployments. It also said it had secured the memory and production capacity needed to manufacture 20,000 cards this year and increase production next year to more than double this year’s planned output. These are production plans outlined by the company, not confirmation of year-end production results.
In the semiconductor business, a good design and reliable supply are two different challenges. After a customer decides to adopt a product, what matters is whether the equipment can be delivered when needed, integrated with existing systems, and supported when operational issues arise. Investment after mass production begins helps expand this execution capability.

The AI inference race extends to power and software

Inference is the process by which a trained AI model answers questions or processes data. As the number of service users grows, this process is repeated, making not only the performance of individual computations but also the ongoing cost of running the service important.
RNGD is an accelerator specialized for inference. The company identifies performance, power efficiency, and programming flexibility to support new models as its key design criteria. The card featured in the official announcement is an air-cooled 180W PCIe model. However, these specifications alone do not establish that it is superior to existing GPUs in every customer environment. The benefits of adoption depend on the models used, request volume, and response-time requirements.
Software matters for the same reason. Even if customers buy a new chip, the burden of switching increases if they cannot easily run the models they use. This is why FuriosaAI has emphasized improving its SDK alongside expanding production. Improving a chip’s performance and enabling customers to use that performance in practice are challenges that must be addressed together.

Expanding the current product while developing the next generation

With this funding, FuriosaAI will increase the global supply of RNGD while developing a third-generation chip with Broadcom and others. The company said the next-generation product will be designed to process agentic AI models in large-scale clusters. Since this product is still under development, it should be distinguished from the performance and supply record of products currently available.
This presents the challenge of keeping two timelines in sync. The product currently on sale must be supplied reliably to customers, while the next product must adapt to changing models and ways of using them. Progress on only one front is not enough. The results of this investment should likewise be measured not by the funding amount itself, but by actual growth in supply and sustained customer use.

Implications for the startup ecosystem

One lesson early-stage teams can take from this case is how to explain the value of their technology. Rather than focusing only on how novel or fast it is, they should also show what burdens it reduces for customers after adoption. For AI infrastructure, that explanation should include power and operating costs, software compatibility, and support systems.
The same standard applies to teams building AI services. When comparing new infrastructure, they should test it with the models and request conditions they actually use, rather than relying on peak performance figures in promotional materials. Shifting the selection criterion from “a more powerful chip” to “can we run our service reliably?” also makes for a more concrete way to assess investment announcements.
Sources
#FuriosaAI #AISemiconductors #AIInference #StartupInvestment #Infrastructure
[Advertisement]
Collaboration proposals / Business inquiries: hello@startlounge.kr
IR/Pitch deck production inquiries: hello@pitchmound.kr
Everything for startups. Start Lounge - Launchpad 「Lunch Pad」
ⓒ 2026. Start lounge. All rights reserved.