
Nvidia, SK Telecom and Rebellions plan 2028 Korean NPU systems
Rebellions plans 2028 NVLink Fusion systems pairing its R200 NPU with Nvidia GPUs. Nvidia confirms one Korean startup investment, plans three more, and SK Telecom prepares A.X K2 tests.
- By
- Nathan Brooks
- Filed
- Channel
- AI & Compute
- Read
- 3 min read
Rebellions expects to supply a combined Nvidia GPU and Rebellions NPU inference system from 2028 under a memorandum of understanding that Nvidia, SK Telecom and Rebellions signed Monday at Nvidia's AI Day event at COEX in Seoul. The three companies will build systems that place Korean NPUs inside Nvidia's rack-scale AI infrastructure.
An MOU is not a purchase order. The document sets no purchase volumes or financial terms, so the 2028 schedule remains a roadmap rather than committed revenue.
What does NVLink Fusion connect?
The design uses NVLink Fusion, Nvidia's program for attaching third-party processors to its high-speed interconnect and rack-scale infrastructure. Rebellions plans to connect its next-generation R200 NPU to NVLink Fusion and use Nvidia's MGX server design.
GPU and NPU systems would share rack space, power supplies and cooling. Nvidia has previously formalized NVLink Fusion cooperation with Marvell and MediaTek.
How will SK Telecom test the system?
SK Telecom will build a combined GPU-NPU system in its AI data center to test performance and stability. It also plans to connect the setup to its A.X K2 model and the South Korean government's “AI for All” public AI services.
The companies say the arrangement could reduce power use and total cost of ownership. MoneyToday noted that actual savings depend on system configuration and operation, a useful brake on procurement optimism.
What has Nvidia committed to Korean startups?
At the same-day forum co-hosted by South Korea's Ministry of Science and ICT and Nvidia, Nvidia discussed its Korean startup investments. Nvidia said its venture arm NVentures had already invested in Korean interconnect startup Point2 during the first half of the year, according to Seoul Economic Daily. The company also plans to invest in Motif Technologies, RLWRLD and Upstage.
Nvidia disclosed no amounts for the Point2 investment or the three planned investments. Seoul Economic Daily reported that Nvidia has not closed those three planned investments, leaving one completed venture commitment and three announced but unclosed plans.
What do the four startups build?
- Point2: Builds radio-frequency interconnects that move data between chips and racks. The company pitches them as an alternative to copper and optical links.
- Motif Technologies: Seoul AI lab that released Motif-3 with open weights.
- RLWRLD: Builds robotics foundation models for humanoid and industrial robots.
- Upstage: Sells Solar language models and document AI tools to businesses.
Those bets span the physical layer, foundation models and business software. Only Point2's investment is confirmed; the other three remain plans without disclosed amounts.
Why does the Korean government care?
South Korea's Ministry of Science and ICT announced the Nvidia, SK Telecom and Rebellions agreement at the same AI Day event. The ministry said it would try to extend similar cooperation to other Korean NPU companies.
More than 20 Korean startups met Nvidia executives and Vice Minister of Science and ICT Ryu Je-myung at the forum. Nvidia's Inception startup program counts more than 1,000 members based in Korea.
That pool gives the government room to broaden the rack design beyond one NPU supplier. It does not create orders, and the MOU still carries no volume commitment.
What changes first?
SK Telecom already works with Rebellions. In April, the operator signed an agreement with Arm and Rebellions to develop inference servers pairing Arm CPUs with Rebellions accelerators.
The Nvidia arrangement adds GPUs, NVLink Fusion and MGX racks to an existing Korean server relationship. If the 2028 system reaches commercial supply, the first test will be whether shared infrastructure preserves accelerator performance while delivering the claimed savings.
Until SK Telecom reports test results, neither capacity nor pricing can support shipment or revenue estimates. If shared racks confirm lower power use and total cost of ownership without a performance penalty, the ministry's effort to add other Korean NPU suppliers gains commercial weight; otherwise, the common-denominator strategy remains mostly architectural.
Source: mlq.ai
More from Nathan Brooks
Show full bio
Senior reporter covering people, launches and day-to-day practice in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.
758 articles
Related articles
jensen-huang-to-meet-samsung-and-sk-chiefs-in-new-york-0765ffb3
Jensen Huang to Meet Samsung and SK Chiefs in New York
nvidia-reportedly-pairs-samsung-with-tsmc-for-next-gen-ai-chips-51564863
Nvidia Reportedly Pairs Samsung With TSMC for Next-Gen AI Chips
los-alamos-powers-up-first-nvidia-vera-cpu-node-in-darwin-testbed-efc159c8
Los Alamos Powers Up First NVIDIA Vera CPU Node in Darwin Testbed
nvidias-nvhbm-sidesteps-jedec-in-the-hbm4-stack
Nvidia’s NVHBM Sidesteps JEDEC in the HBM4 Stack



