Nvidia CEO Jensen Huang is pursuing early-stage talks with Korean AI chip designer Rebellions about a technical partnership, investment, or possible acquisition. Huang met with Rebellions co-founder and CEO Sunghyun Park at Nvidia's Santa Clara headquarters this week to discuss a potential tie-up, according to people familiar with the matter. The discussions are preliminary and may not result in a deal, but they reveal a deliberate pattern in how Nvidia is consolidating its dominance in artificial intelligence hardware.

What Is Rebellions and Why Does Nvidia Want It?

Rebellions, founded in 2020 and based in Bundang, South Korea, specializes in neural processing units (NPUs) designed specifically for AI inference. Inference is the computational work of running an AI model after it has been trained, like when you ask ChatGPT a question and it generates a response. Unlike training, which requires enormous computing power upfront, inference happens millions of times at scale but with lower computational demands per individual query.

The company has already shipped products for three years. Its Atom and Atom Max NPUs entered mass production in 2023, and customers in Japan, Saudi Arabia, and the United States have deployed them. Rebellions has raised roughly $850 million from major investors including SK Hynix, Samsung Ventures, and Arm Holdings, plus direct backing from the Korean government. Its most recent valuation was around $2.3 billion.

How Is Nvidia Structuring These Deals Without Full Acquisitions?

This would be Nvidia's fourth arrangement of this type in under a year. Rather than buying companies outright, Nvidia has developed a distinctive playbook: secure a nonexclusive license to the company's technology, hire most of its engineering talent, take a strategic stake, and let the company continue operating independently. This structure sidesteps the regulatory scrutiny that large acquisitions typically trigger from the US Department of Justice and other authorities.

The most prominent example is Groq, an AI chip startup. In late 2025, Nvidia paid $20 billion for a nonexclusive license to Groq's technology and absorbed most of its engineering team. Nvidia insisted it was not an acquisition because Groq continues to operate separately and runs its own cloud business independently. More recently, Nvidia agreed to pay Poolside $6 billion for its model factory and hire 109 of its staff.

Why This Strategy Makes Business Sense for Nvidia

Nvidia faces mounting antitrust pressure. The company holds such a dominant share of the semiconductors used to train frontier AI models that regulators are watching closely. Large acquisitions typically require clearance from authorities like the US Department of Justice, which could delay or block deals. By licensing technology and hiring talent instead of acquiring the company, Nvidia avoids triggering formal antitrust reviews while still gaining access to the engineering expertise and intellectual property it needs.

The Rebellions situation involves an additional layer of complexity: Korea treats advanced semiconductors as strategic national assets. Samsung Electronics and SK Hynix work closely with the Korean government on investment projects, and SK Hynix is itself a Rebellions shareholder. A full acquisition might invite regulatory scrutiny from Seoul, making the license-and-hire structure even more attractive.

What Happened to Groq After Nvidia's $20 Billion Deal?

The Groq case offers a cautionary tale about the long-term value of these arrangements. Nvidia valued its license to Groq at $20 billion. Groq then raised $650 million in subsequent funding for what remained of the company. This month, Groq closed a $350 million funding round at a $3.5 billion valuation, with Nvidia participating in the round. That represents a valuation decline of nearly half in a single funding round. A company Nvidia paid $20 billion to license is now worth $3.5 billion as a standalone business.

The reason becomes clear when you look at what Nvidia did with the license. Nvidia turned the Groq technology into its own inference chip, the Nvidia Groq 3 language processing unit (LPU), which it unveiled at GTC in March. The company said it would add a new LPU architecture every year alongside its annual graphics processing unit (GPU) cadence. The chip ships in the second half of this year in liquid-cooled racks holding 256 LPUs, with 128 gigabytes of on-chip memory and 640 terabytes per second of bandwidth.

What Are the Trade-Offs of Nvidia's Inference Chip Strategy?

Ian Buck, who runs Nvidia's data centre business, described the fundamental challenge at GTC. He explained that the LPU is optimized strictly for extremely low-latency token generation, offering token rates in the thousands of tokens per second. A token is a small unit of text, roughly equivalent to a word or part of a word. The trade-off, Buck noted, is that you need many chips to achieve that performance, and the economics per individual chip are actually quite poor.

This is significant because it comes from Nvidia's own data centre head describing the per-chip economics of the inference category. Rebellions designs the same class of part. If Nvidia licenses Rebellions' technology, it would be productizing a second inference architecture after successfully turning Groq's design into its own chip.

How to Understand Nvidia's Broader Investment Strategy

• Licensing Over Acquisition: Nvidia pays for access to rival chip designs and engineering talent without acquiring the companies outright, avoiding regulatory review and allowing the original company to continue operating independently.
• Regulatory Navigation: By structuring deals as licenses and hiring arrangements rather than acquisitions, Nvidia sidesteps antitrust scrutiny from the US Department of Justice and foreign governments that view semiconductors as strategic assets.
• Technology Productization: Nvidia takes licensed designs, integrates them into its own product roadmap, and ships them at scale, while the original company's valuation often declines as Nvidia captures the market opportunity.
• Global Expansion: Nvidia targets startups in strategic regions like South Korea, where government backing and partnerships with major manufacturers like SK Hynix add credibility and geopolitical significance to the deal.

What Does Rebellions Want From This Deal?

Rebellions has been pointing toward an initial public offering (IPO) rather than an exit. Chief Financial Officer Sungkyue Shin said a year ago that going public was the "master plan." In March, the company raised more than $400 million in a round it described as pre-IPO and used the moment to launch two rack-scale platforms called RebelRack and RebelPod. No listing date has been disclosed.

An Nvidia license would give Rebellions cash without requiring a public listing, which is the trade that both Groq and Poolside took. For a company backed by the Korean government and major manufacturers, a partnership with Nvidia could provide validation and resources while preserving independence and the option to go public later.

No deal structure, price, or stake has been reported. Whether this would be a license, a minority investment, or a purchase remains unclear. Neither Nvidia nor Rebellions has commented on the record, and neither has said what a partnership would cover if it stopped short of an investment.

What Else Is Nvidia Doing in Chip Design?

Separately, Nvidia spent this week denying a different chip report. Asked about a claim that it had built a China-specific LPU, a Nvidia spokesperson said the reporting "is incorrect," adding: "We have no LPU sales in the China market today, and no China-specific LPU product in our roadmap". The denial suggests that Nvidia's inference chip strategy is focused on global markets rather than region-specific designs.

The Rebellions talks also come as memory becomes increasingly central to AI scaling. Micron CEO Sanjay Mehrotra stated that "memory is no longer a component in a system; memory is the strategic infrastructure for AI". Nvidia CEO Jensen Huang backed Micron's $10 billion research lab initiative, noting that memory reinvention is "one of the great challenges of the AI era". This suggests that Nvidia's strategy extends beyond chips to the entire infrastructure stack required to train and run AI models at scale.

Sanjay Mehrotra