A server rack hums in a data hall outside Seoul, drawing less power than the racks beside it while pushing through the same flood of AI queries.
Inside sits a palm sized chip built to answer, not to learn.
Its maker just closed one of the largest funding rounds any chip challenger has landed this year.
The company’s worth has moved in a way few private firms see happen in twelve months.
So how did one accelerator pull in that kind of money?
What makes this chip different from the ones it is chasing
The startup is called FuriosaAI, and its product is an inference focused accelerator named RNGD, built to answer AI queries rather than train new models from scratch. That distinction matters because training chips and inference chips are increasingly built, priced and sold as separate categories of hardware.
Inference is the unglamorous half of AI computing, the part that runs every time a chatbot answers a question or a search box returns a summary. It happens constantly, at enormous scale, and it rewards power efficiency over raw peak speed.
That is the gap FuriosaAI says it designed RNGD to fill. According to the company’s own announcement, FuriosaAI, an AI semiconductor company building sustainable, high-performance inference compute for data centers, announced the completion of its $620 million (KRW 820 billion) Series D round, bringing its post-money valuation to approximately $2.9 billion.
The pitch to buyers is simple: similar throughput to a popular rival card, but running more simultaneous users on the same electricity bill.
Inside the round that reshaped the company’s price tag
The scale of the raise is what pulled outside attention toward a company most American readers have never heard of. FuriosaAI announced the completion of its $620 million Series D round, bringing its post-money valuation to approximately $2.9 billion , a jump that happened in about a year.
That new number is not a modest step up. The startup’s valuation rose to $2.9 billion, approximately four times its previous valuation of $735 million. Few private semiconductor companies anywhere have seen their price multiply that fast outside the handful of giants already locked into the AI boom.
South Korea’s state backing gave the round an unusual shape. The round was led by Infinitum Partners and the Korea National Growth Fund, which jointly invested $280 million, and this marks Korea Eximbank’s first direct investment in a startup since the bank’s founding.
Investors include U.S.-based Infinitum Partners, South Korea’s National Growth Fund, Jericho Capital, Worldline Capital, Singapore’s QUDA Capital, and Abu Dhabi’s NeoFort Holding , alongside earlier Korean backers staying in.
Where the evidence for that claim actually comes from
None of this would carry weight without real customers already running the chip, and FuriosaAI points to several. FuriosaAI has gone further, securing production deployments with leading enterprises such as Samsung SDS and LG Group , giving the efficiency pitch a commercial track record rather than just a lab benchmark.
That said, the headline efficiency number is a company claim, not an outside lab result. RNGD matches RTX PRO 6000 throughput at the same card count and serves twice as many users per kW of power , and outlets covering the raise have repeated that figure while cautioning that both figures are company claims, and the comparison point is a workstation-class Nvidia part, not its top data center accelerators.
Founded in 2017, FuriosaAI began mass-producing its RNGD inference chips in January , so the hardware behind this valuation jump has been shipping for roughly ten months.
Production plans back up the scale the new money is meant to fund. FuriosaAI’s RNGD inference accelerator is in mass production: it will produce 20,000 cards this year, and has secured the memory and fab capacity to more than double that in 2027.
Some of that money is already locked into future supply rather than sitting as a cushion. Readers who have followed how etching costs for advanced chip designs have climbed will recognize why locking in fab time counts as real progress.
The limits that still separate a funding round from a finished rival
A fourfold valuation jump is not the same as proof that FuriosaAI can dent Nvidia’s position in American data centers. One analysis of the round was blunt, noting the sharp repricing shows investors are willing to fund credible challengers to Nvidia, though the headline scale alone does not prove FuriosaAI can compete at data-centre scale.
That is the honest complication sitting underneath the headline number. Every confirmed deployment so far, Samsung SDS, LG Group, and the Korean AI startup Upstage, sits inside one country’s technology ecosystem, where LG Group, the parent company of LG Electronics, is using FuriosaAI’s chips to power its EXAONE models, while Upstage is using the chips to power live search summaries for millions of users, and Samsung SDS offers access to the company’s silicon through its commercial neural processing units-as-a-service offering.
Breaking into American or European data centers is a different sales problem entirely, one built on existing software stacks and years of vendor trust.
Even the company’s own summary of where things stand is cautious rather than triumphant. One industry writeup noted the release discloses no revenue, backlog or chip pricing, so the valuation rests on deployments and volume targets, and the efficiency claims are the company’s own, with no named third-party benchmark.
The next chip in the pipeline carries its own risk, too. FuriosaAI will use the money to make more RNGD chips, improve its software and build its third-generation chip, using a 2nm, multi-die design with HBM4 memory, an ambitious jump that will take years to deliver.
What this milestone means for the chip market watching from outside Korea
FuriosaAI’s story lands at a moment when every major buyer of AI hardware is hunting for ways to cut electricity costs without cutting capacity. Furiosa’s pitch is power efficiency: electricity and cooling limit how many accelerators a facility can host, so a lower-wattage chip can fit more capacity inside the same power budget. That is exactly the kind of pitch that gets a meeting, even without a flagship American contract yet.
That gap between funding momentum and proven global adoption is where the real test now sits. The money, the fab slots and the shipping hardware are all real and documented.
What remains unproven is whether a Korean challenger can turn efficient inference into orders from the data center operators who matter most outside its home market. Those figures are the vendor’s own, measured against a workstation-class Nvidia card rather than the data center parts that dominate large clusters, and Nvidia remains the reference point for the whole inference market, with its software stack the harder obstacle.
For now, FuriosaAI’s rise is a reminder that the AI hardware race is not only about who trains the biggest model. It is increasingly about who can answer billions of everyday queries for the least electricity, and that fight has only just begun.
