Nvidia bets on four Korean AI startups and plugs a Korean NPU into NVLink Fusion
At AI Day Seoul, Nvidia announced direct investments in Point2, Upstage, Motif and RealWorld, signed a deal to connect Rebellions' R200 chip to its NVLink Fusion interconnect for SK Telecom's AI data center, and saw RealWorld launch DexBench, a 15-task benchmark for robot hands.
📊 Scorecard: our Bullish call on $NVDA is a miss so far: from $178.67 to $168.89 at the latest close (-5.5%). Against SOXX the call is wrong: the stock is 5.4 points behind the index.
The News
Nvidia used AI Day Seoul on November 9 to plant a flag in Korea's AI industry — twice. First, it unexpectedly disclosed plans to invest directly in four South Korean AI startups at the Nvidia AI Startup Partnership Forum, a side event of AI Day Seoul at COEX. Point2 Technology, which builds radio-frequency interconnects that move data between AI chips and racks inside data centers, had already received funding from NVentures — Nvidia's venture arm — in the first half of 2026, and now gets a direct Nvidia investment as well. The three new planned investments are Upstage and Motif Technologies, which build large language models, and RealWorld, which builds robot foundation models for humanoids and industrial robots. The investment procedures for the three have not yet completed, and no amounts were disclosed: these are announced plans, not closed deals.
Second, Nvidia, SK Telecom and Rebellions signed a memorandum of understanding to connect Rebellions' next-generation R200 NPU to Nvidia's NVLink Fusion interconnect and jointly develop an integrated GPU-NPU system on Nvidia's MGX rack design. An NPU is a chip designed for inference — running a trained AI model — and the pitch is economics: assigning GPUs and NPUs to the tasks each does best can lower a data center's power consumption and total cost of ownership. The system is to be deployed at SK Telecom's AI data center to verify performance and stability, with supply targeted from 2028. It is the first time Nvidia has formalized NVLink Fusion cooperation with a Korean NPU company, following Marvell and MediaTek. Nvidia also released Dynamo deployment recipes for four Korean open models, it said in a separate announcement: SK Telecom's A.X K2, LG AI Research's K-EXAONE 2.0, Upstage's Solar Open 2 and Motif-3.
The robotics angle arrived the same day. RealWorld unveiled DexBench at AI Day Seoul, it announced the same day: a benchmark for humanoid dexterous manipulation — how well robot hands do fine, precise work — built on Nvidia's Isaac Lab-Arena and the Newton physics engine, with 15 tasks across four industry environments: retail, logistics, assembly, and hotel food and beverage. The tasks include connector insertion, screw fastening and pouring liquids, and RealWorld plans to release the code, environments, training data, baseline models and a leaderboard. DexBench is the first major release from the RLWRLD–Nvidia dexterity initiative launched in June. This is new day-2 AI Day material, distinct from the day-1 physical-AI collaborations published on November 9.
Nvidia is not just selling chips in Korea — it is buying into the companies that will use them.
Why It Matters
The deal matters because Nvidia is opening a door it has historically kept shut. NVLink is the high-speed link between Nvidia's own GPUs; NVLink Fusion, unveiled in May 2025, extends it to outside AI chips. Letting Rebellions' R200 — a Korean NPU — onto that interconnect, and building a joint GPU-NPU system on Nvidia's MGX rack design, turns a supplier relationship into a partnership. The pitch is inference economics: GPUs and NPUs are good at different things, and assigning each task to the chip that does it best can cut a data center's power use and total cost of ownership. SK Telecom plans to link its A.X K2 model and public AI services into the same framework.
The startup investments are the other half of the same strategy. Point2's radio-frequency interconnects move data between chips and racks as an alternative to copper and optical links; Upstage and Motif build the language models; RealWorld builds the robot brains. Nvidia is buying into every layer of Korea's AI stack at once — chips, interconnects, models and robots — while its own chips sit at the center of all of them. Rebellions and Point2 are private, so the public read-throughs are Nvidia and Marvell, which already builds for the NVLink Fusion ecosystem. The caveat, stated plainly: none of this is closed yet, and the R200 system is a 2028 target. The bet is real; the returns are not.
Analyst's Take
Our call is Bullish. Nvidia is doing something it rarely does: investing directly, not through its venture arm, in four startups at once — and opening its interconnect to an outside chipmaker's NPU in the same breath. The message to Korea's AI industry is that Nvidia wants to be inside it, not just selling to it. For Rebellions and Point2, both private, the public read-throughs are Nvidia itself and Marvell, which already plays in the NVLink Fusion ecosystem. For SK Telecom, whose AI data center will host the first GPU-NPU system, the deal is a step toward the 'full-stack sovereign AI' framework it plans to build around its A.X K2 model.
What to watch: whether the plans close, and whether the R200 system works. The investment procedures have not completed, no sizes were disclosed, the MOU sets no volumes or financial terms, and 2028 is a supply target, not a delivery — this is a foothold, not a booking. DexBench is the most concrete piece: a real benchmark with real tasks, and a leaderboard that will let the industry measure robot dexterity against a common standard. If DexBench becomes the standard test for robot hands, RealWorld — and Nvidia's bet on it — gains influence over how the whole physical-AI field is judged. That is the long game here, and it is worth more than any single investment.
Are you bullish or bearish on this story?
By The Numbers
FAQ
What did Nvidia actually announce in Korea?
Two linked things at AI Day Seoul on November 9: direct investments in four Korean AI startups — Point2, Upstage, Motif and RealWorld — and a memorandum of understanding with SK Telecom and Rebellions to connect Rebellions' R200 NPU to Nvidia's NVLink Fusion interconnect.
What is the R200 and NVLink Fusion deal?
Rebellions will connect its next-generation NPU — a chip designed for AI inference, the work of running a trained model — to Nvidia's high-speed NVLink Fusion interconnect, and jointly develop an integrated GPU-NPU system on Nvidia's MGX rack design for deployment at SK Telecom's AI data center, to verify performance and stability. Rebellions expects to supply the system from 2028.
Who are the four startups?
Point2 Technology builds radio-frequency interconnects that move data between AI chips and racks inside data centers. Upstage builds Solar large language models. Motif Technologies builds foundation models. RealWorld builds robot foundation models for humanoids and industrial robots.
What is DexBench?
A benchmark for humanoid dexterous manipulation — how well robot hands do fine work — built on Nvidia's Isaac Lab-Arena and the Newton physics engine. It has 15 tasks across four industry environments: retail, logistics, assembly, and hotel food and beverage, including connector insertion, screw fastening and pouring liquids.
Are the investments closed deals?
No. The investment procedures for the three new companies have not yet completed, and no investment sizes or timing were disclosed. These are announced plans, not closed deals. The MOU also sets no volumes or financial terms, and 2028 is the supply target, not a delivery.
👁
Source: Seoul Economic Daily