NVIDIA is reportedly discussing a partnership, an investment, or possibly an acquisition with Rebellions, a Korean AI chip design company. The talks are at an early stage and may not lead to anything. Still, once you recall the pattern NVIDIA has repeated over the past year, taking the technology and the engineers without buying the company, the choice of partner starts to look deliberate.
Take the technology and the people, leave the company standing
According to the reports, NVIDIA CEO Jensen Huang met Rebellions co-founder and CEO Sunghyun Park this week at NVIDIA's Santa Clara headquarters in California, with a technical partnership, an investment, or an acquisition on the table. People familiar with the matter described the discussions as preliminary and said they may not result in a transaction. NVIDIA has not commented, and a representative for Rebellions declined to respond.
What stands out is that this would be the fourth deal of the same shape. Over the past year, rather than buying rival chip designers and model builders outright, NVIDIA has paid for non-exclusive licences, hired large numbers of their engineers, taken a stake, and left the legal entity operating.
The most recent example is Poolside. NVIDIA agreed to pay 6 billion USD (about 950 billion yen) for the company's model factory and to hire 109 of its staff. Before that it paid Groq 20 billion USD (about 3.2 trillion yen) for a non-exclusive licence and absorbed much of its engineering team. NVIDIA says this was not an acquisition, and Groq does continue to run its cloud business independently. Rivals are using the same playbook: AMD acquired Taalas, a startup that etches models directly into silicon, in August.
*1 USD = 159 JPY
Rebellions is built purely for inference
Rebellions was founded in 2020 and is based in Bundang, south of Seoul. It designs neural processing units (NPUs) for data centres, and the distinguishing point is that they target inference rather than training.
Inference is the work of running a model once it exists. Each query needs less compute than training does, but the count grows with every user, so the total far exceeds training. Whether that work can be run cheaply and quickly goes straight to the economics of any AI service.
The company has raised roughly 850 million USD (about 135 billion yen) from backers including SK hynix, Samsung Ventures and Arm, and it also holds a direct investment from the Korean government. Its most recent valuation was around 2.3 billion USD (about 370 billion yen). Its products have been shipping for three years: the Atom and Atom Max entered mass production in 2023, and customers in Japan, Saudi Arabia and the United States have reportedly deployed them. In 2024 the company merged with Sapeon Korea, a chip business SK Telecom spun out in 2016, and last year it partnered with Marvell on hardware for sovereign AI projects.
NVIDIA already sells an inference chip of its own
These talks read as more than a simple investment story because NVIDIA has already turned the Groq licence into a product of its own. The NVIDIA Groq 3, unveiled at GTC in March, is a language processing unit (LPU), and the company has said it will introduce a new LPU architecture every year alongside its GPU cadence. It ships in the second half of this year in liquid-cooled racks holding 256 LPUs, with 128GB of on-chip memory and 640 terabytes per second of scale-up bandwidth.
Rebellions designs parts in exactly that class. If a licensing arrangement came together, NVIDIA would be taking on a second inference architecture.
NVIDIA itself is careful about the economics of the category. Ian Buck, who runs the company's data centre business, explained at GTC that the LPU is optimised strictly for extreme low-latency token generation and can reach thousands of tokens per second, but that many chips are needed to get there, so tokens per second per chip is quite low. That is the head of the business acknowledging the trade-off against the GPU line that pays the bills.
Regulators, industrial policy, and the listing Rebellions wanted
A full acquisition would face 2 obstacles. The first is US competition law. NVIDIA's share of the chips used to train frontier models is large enough to invite scrutiny, and sizeable technology acquisitions generally need clearance from the Department of Justice. The licence-and-hire structure has the practical advantage of avoiding that review entirely.
The second is Korea. Seoul treats advanced semiconductors as strategic national assets, and Samsung and SK hynix work closely with government investment plans. SK hynix is a Rebellions shareholder, and it also backs Etched, whose valuation doubled in 7 months to 10.3 billion USD (about 1.64 trillion yen).
Rebellions, for its part, has been talking about going public rather than selling. Chief financial officer Sungkyue Shin said a year ago that an IPO was the company's master plan. In March it raised more than 400 million USD (about 64 billion yen) in a round it described as pre-IPO, and used the occasion to launch 2 rack-scale platforms, RebelRack and RebelPod. No listing date has been disclosed. A licence fee from NVIDIA would be a way to reach cash without a listing, which is the trade both Groq and Poolside took.
Very little is settled at this point. No structure, price or stake has been reported, and neither company has said anything on the record. Separately, NVIDIA spent this week rejecting a different chip report, stating that no China-specific LPU exists in its roadmap.
Summary
The NVIDIA and Rebellions discussions reflect 2 shifts at once: the centre of gravity in AI silicon moving from training toward inference, and NVIDIA settling into a pattern of absorbing rival technology without buying the company. Rebellions has a track record in inference-only NPUs and has been aiming at a public listing. A deal would make this the fourth case of its kind, but nothing about the structure or the price has been decided, and regulation and Korean industrial policy remain open questions.
