South Korean chip startup FuriosaAI has raised $620 million in a Series D round that would value the company at roughly $2.9 billion, according to SiliconANGLE and GamesBeat. The size of the cheque matters less than the thesis behind it: as power shortages slow data center expansion, performance per watt is becoming a commercial argument nearly as decisive as raw speed.
🔑 Key Takeaways
- $620 million Series D, including $280 million from South Korea’s National Growth Fund.
- Post-money valuation of about $2.9 billion in the main reports; one sector database puts it at $3 to 4 billion.
- RNGD chip built on TSMC’s 5nm process, with output targeted at 20,000 units in 2026 and over 40,000 in 2027.
- Production customers include LG Group, Upstage and Samsung SDS.
A Series D Backed by the South Korean State
On October 28, 2026, FuriosaAI announced it had raised $620 million, or roughly 820 billion won, in a Series D round. The money is meant to fund three priorities: scaling production of its RNGD inference accelerators, strengthening its software ecosystem and accelerating development of its third-generation chip, Gen 3. Founded in Seoul in 2017 by June Paik, a former engineer at AMD and Samsung, the company positions itself as a semiconductor firm focused on energy-efficient AI accelerators for inference in data centers.
The round is anchored by a public investor. The National Growth Fund of South Korea injected $280 million and becomes the lead investor, as part of a government plan to strengthen the domestic semiconductor industry and the country’s AI compute infrastructure. The table below summarizes the breakdown as reported.
| Investor group | Amount | Named participants |
|---|---|---|
| National Growth Fund (South Korea) | $280M | Lead investor of the round |
| International investors | $110M | Infinitum, Cisco Investments, NeoFort Holding, Worldline Capital, QUDA Capital Partners, Jericho Capital |
| Existing shareholders | $125M | NAVER Ventures, DSC Investment, KDB Capital, Korea Development Bank ($22M), Nextrans, Kakao Investment |
| New Korean institutions | Not disclosed | United Asset Management ($30M), Export-Import Bank of Korea, Hanwha Asset Management |
The disclosed contributions from the National Growth Fund ($280 million), international investors ($110 million) and existing shareholders ($125 million) add up to $515 million. The remaining $105 million is not broken down in the available information, so the sources cannot fully reconcile the allocation of the $620 million announced.
On the international side, US-based Infinitum, which specializes in AI semiconductors, is the largest foreign investor. Cisco Investments, NeoFort Holding, Worldline Capital and Singapore’s QUDA Capital Partners also took part. The sources describe Charlie Foo as founder and managing partner of one of the participating funds; he is a former executive at Nvidia and Mellanox Technologies and now serves as senior adviser to FuriosaAI.
Existing shareholders contributed $125 million through new positions. Korea Development Bank added another $22 million. All three investors from the 2017 seed round, NAVER, DSC Investment and KDB Capital, re-invested. NAVER first invested through NAVER D2SF at the seed stage and returned in 2026 through NAVER Ventures, its Silicon Valley-based investment platform. Nextrans and Kakao Investment, which joined during a Series C transition round the previous year, increased their commitments.
The Export-Import Bank of Korea selected FuriosaAI for its first direct investment in a startup. United Asset Management Company committed $30 million, making it the largest new domestic investor after the National Growth Fund. Hanwha Asset Management also joined the round.
“This Series D marks an important milestone for FuriosaAI and reflects the confidence of investors around the world in our technology and in our ability to execute.”
June Paik, Co-founder and CEO, FuriosaAI
Valuation: Figures That Diverge
Valuation is the least stable number in this deal. Bloomberg reports that FuriosaAI’s valuation has quadrupled, while SiliconANGLE and GamesBeat put the post-money valuation at about $2.9 billion. The New Market Pitch sector database estimates $3 to 4 billion, but with only partial confidence: that figure is inferred from the amount raised rather than observed in a market price.
What the Sources Say
| Source | Valuation | Confidence |
|---|---|---|
| SiliconANGLE, GamesBeat | ~$2.9B post-money | Announced valuation |
| Bloomberg | Valuation quadrupled | Not specified in the source document |
| New Market Pitch | $3 to 4B | Partial (inferred from funding) |
Total funding raised to date shows a similar gap. New Market Pitch cites about $737 million with strong confidence on the amount, while Compute Compass lists $246 million for FuriosaAI. The sources do not clarify whether the difference reflects different accounting dates, different scopes or data predating the 2026 announcement.
Sector Comparables
- Etched: $40 to 50 billion, based on investment offers
- Cerebras: $43.1 billion, based on its public market capitalization
- Kunlunxin: $14.7 to 20 billion
- SambaNova: $11 billion
- Tenstorrent: $8 to 10 billion
Even at the low end, FuriosaAI sits well below Cerebras or Etched. That gap may leave room for upside if execution holds, but it also underlines that the value of a young chip company depends heavily on commercial proof.
A Chip Designed Around Tensor Contraction
FuriosaAI introduced its Tensor Contraction Processor (TCP) earlier in 2026. It is the second generation of inference chips built on the RNGD architecture. Where Nvidia GPUs rely on a rigid two-dimensional matrix for computation, TCP and RNGD chips use tensor contraction as the fundamental unit of compute, meaning the multiply-and-add operation at the core of neural networks.
By making this operation the primary compute unit and using deterministic execution directly in silicon, the architecture reduces data movement, one of the most energy-intensive parts of inference. FuriosaAI claims its chips consume less than half the power of Nvidia GPUs while remaining compatible with frameworks such as PyTorch and vLLM. That is the vendor’s claim, and the sources provided do not include an independent verification.
| Parameter | RNGD (per Compute Compass) | Analytical reading |
|---|---|---|
| Manufacturing process | TSMC 5nm | Advanced node, outsourced fabrication |
| Memory | Two HBM3 stacks | High bandwidth, useful for inference |
| Performance | 512 TOPS at INT8 | Metric to compare under identical conditions |
| Power draw | Less than half of Nvidia GPUs | FuriosaAI claim, not independently verified |
The market remains dominated by Nvidia GPUs, yet a growing number of data centers struggle to expand because they lack sufficient electrical capacity. Power grids and cooling systems are also under strain. FuriosaAI and its investors see this energy constraint as a potential advantage for chips designed to maximize the number of tokens processed per watt. In that context, the relevant metric is no longer only raw performance but the energy cost of each query.
“We are looking for the moment when a technology company moves from demonstrating its technology to demonstrating its business model. FuriosaAI is now at that stage, with its chip in volume production and deployed with customers.”
Yong Jung Park, General Partner, NAVER Ventures
Customer Deployments and the 2026-2027 Roadmap
FuriosaAI plans to produce 20,000 RNGD accelerators in 2026 and more than 40,000 units in 2027, more than doubling output. In parallel, the third-generation processor is being co-developed with Broadcom. The first deployments cited by the sources are:
- LG Group uses the chips for its EXAONE models.
- Upstage uses them to generate live research summaries for millions of users.
- Samsung SDS launched a NPU-as-a-service offering in July 2026, a cloud neural processing unit service built on FuriosaAI silicon.
“Taking an advanced AI chip from design to volume production is extremely difficult, and few companies have shown they can do it.”
John Yetimoglu, Director of Investments, Infinitum
These points need context. Moving from design to volume production is demanding for any advanced chip, and the target of 40,000 units in 2027 will have to be confirmed by actual shipments. Nvidia’s widely adopted software ecosystem remains the main structural obstacle.
Outlook and Scenarios
FuriosaAI’s round confirms that the AI inference market now attracts institutional and public capital at scale. The company’s bet is clear: make energy efficiency a buying criterion for data center operators at a time when electricity is becoming as binding a constraint as chip supply.
Three scenarios stand out. In the base case, FuriosaAI meets its target of 20,000 units in 2026 and ramps toward 40,000 units in 2027, with Korean customers serving as international showcases. In the bull case, expansion to North American customers and the delivery of Gen 3 with Broadcom strengthen its position against Nvidia. In the bear case, production delays, limited adoption outside Korea or real-world efficiency below the vendor’s claims would weigh on a valuation whose estimates already diverge widely. Tracking delivered volumes and independent benchmarks will be decisive.
Sources
- SiliconANGLE, October 28, 2026
- New Market Pitch, AI inference chip startups ranking
- Compute Compass, AI accelerators
- Bloomberg, October 28, 2026
- GamesBeat, FuriosaAI raises $620M
Cet article est publié à titre informatif et éducatif. Il ne constitue en aucun cas un conseil en investissement. Faites vos propres recherches (DYOR) avant toute décision.
