Jensen Huang spent part of this week in a conference room at Nvidia's Santa Clara headquarters with a man whose stated ambition is to break Nvidia's monopoly. Bloomberg reported Friday that Huang met Sunghyun Park, co-founder and chief executive of the Korean AI chip designer Rebellions, to discuss a possible tie-up — a technical partnership, an equity investment, or an outright acquisition. People familiar with the discussions, who asked not to be identified because the information is not public, told the news agency the talks are preliminary and may not produce a deal. Nvidia did not respond to Bloomberg's questions. A representative for Rebellions declined to comment.
It is the fourth arrangement of roughly this shape that Nvidia has pursued in under a year, and the pattern is now legible enough to predict. Over the past twelve months the company has spent heavily on access to rival inference designs without buying the companies that made them: $20 billion for a nonexclusive license to Groq's architecture plus most of its engineering staff, and, earlier this week, $6 billion to license Poolside's model factory and hire 109 of its people. Nvidia insisted the Groq arrangement was not an acquisition, because Groq continues to operate independently and still runs its own cloud business.
What Rebellions brings
Founded in 2020 and headquartered in Bundang, south of Seoul, Rebellions designs neural processing units for data centers, tuned specifically for inference rather than training. Its Atom and Atom Max NPUs entered mass production in 2023, and DataCenterDynamics reported that customers in Japan, Saudi Arabia and the United States have deployed them. It merged with Sapeon Korea in 2024, partnered with Marvell last year on sovereign AI hardware, and in June acquired the inference-optimization startup SqueezeBits to bring serving software in-house.
The capital table is the part Nvidia would have studied hardest. Rebellions has raised roughly $850 million from SK Hynix, Samsung Ventures and Arm, and holds a direct investment from the Korean government — it was named the first direct investment of the National Growth Fund in March, part of Seoul's explicit effort to cultivate a "K-Nvidia." Its most recent valuation was around $2.3 billion, set in a March round of more than $400 million that the company described as pre-IPO and used to launch two rack-scale platforms, RebelRack and RebelPod. Chief financial officer Sungkyue Shin has said a public listing is the "master plan," with JPMorgan appointed global lead underwriter.
Park has not been coy about the target. Speaking at Seoul Forum 2026 in May, in remarks translated from Korean by Seoul Economic Daily, he argued that leadership in AI silicon will shift from GPUs to inference-optimized NPUs the way Intel's open ecosystem once unseated IBM. He cited Gartner and Bloomberg projections that the inference chip market will grow from $39 billion in 2024 to more than $475 billion by 2030, and claimed Rebellions parts deliver five to seven times better price efficiency than competing systems. "We respect the versatility and flexibility of Nvidia's ecosystem, but Rebellions has focused exclusively on inference efficiency," he said. His closing line was blunter: "Rebellions' goal and journey is to confidently stand up to the giant called Nvidia and advance into the global market within the next two to three years."
In an official English-language statement in June, announcing the SqueezeBits deal, Park framed the strategy in ecosystem terms: "We believe Korea's AI infrastructure ecosystem opens up new possibilities when technical capability and exceptional talent come together across the boundaries of individual companies."
Why It Matters
Nvidia already sells the product Rebellions makes. It turned the Groq license into the Nvidia Groq 3 language processing unit, unveiled at GTC in March and shipping in the second half of this year in liquid-cooled racks of 256 LPUs with 128GB of on-chip memory and 640 terabytes per second of scale-up bandwidth. The company said it would add a new LPU architecture annually alongside its GPU cadence. Licensing a second inference architecture after productizing the first suggests Nvidia views this category as one it must own outright rather than merely compete in.
The economics are less flattering than the deal sizes imply. Ian Buck, who runs Nvidia's data center business, described the trade-off at GTC while discussing Nvidia's own chip: "The LPU is optimized strictly for that extreme low-latency token generation, offering token rates in the thousands of tokens per second. The trade-off, of course, is that you need many chips in order to get that kind of performance. And the economics, or the tokens per second per chip, is actually quite low." Nor has the first deal aged well for the seller: Groq, whose architecture Nvidia valued at $20 billion, closed $350 million at a $3.5 billion valuation this month — down by nearly half — with Nvidia joining the round.
A purchase would also be structurally hard. Nvidia's share of frontier training silicon is large enough to invite antitrust scrutiny, and a sizeable acquisition would need US Department of Justice clearance; the license-and-hire structure sidesteps that review entirely. Seoul is the second obstacle. Korea treats advanced semiconductors as strategic national assets, Samsung and SK Hynix coordinate closely with the government on investment, and SK Hynix is both a Rebellions shareholder and a backer of Etched, whose valuation doubled to $10.3 billion in seven months.
What to Watch
No structure, price or stake has been reported, and that absence is the story: whether this becomes a license, a minority investment or a purchase determines which regulators get a say. Watch whether Rebellions' 2027 listing plan survives contact with Nvidia cash — a license would deliver capital without a public offering, the trade Groq and Poolside both accepted. Watch Seoul's response, given that the government is a direct investor in a company it has branded the national champion. And watch whether Nvidia's next annual LPU refresh carries any Rebellions DNA — that would be the clearest evidence that Huang is not buying a competitor so much as consolidating a roadmap.
“Rebellions' goal and journey is to confidently stand up to the giant called Nvidia and advance into the global market within the next two to three years.”— Sunghyun Park, Co-founder and CEO, Rebellions