Nvidia in Talks With Rebellions AI: What We Know

Nvidia and Rebellions are in early talks

Bloomberg, as relayed by The Next Web, reported on 21 August 2026 that Nvidia is in early talks with Rebellions, a Korean chip designer, about a technical partnership, an investment or possibly an acquisition. Jensen Huang met Rebellions co-founder and chief executive Sunghyun Park at Nvidia's Santa Clara headquarters that week, according to people familiar with the matter. They asked not to be named because the information is not public.

What has been reported about the talks

Item	Detail
Reported by	Bloomberg, 21 August 2026
Options under discussion	Technical partnership, investment, or acquisition
Who met	Jensen Huang (Nvidia) and Sunghyun Park (Rebellions)
Where	Nvidia's Santa Clara headquarters
Stage	Preliminary; may not produce a deal
On the record	Nvidia did not respond; Rebellions declined to comment
Not reported	Structure, price and stake
Status on 11 October 2026	No agreement and no collapse reported

Beyond the meeting and the stage of the talks, no terms have been reported. Nvidia did not answer the questions put to it, and a Rebellions representative declined to comment.

EE Times, in a piece published after the Bloomberg report, described Nvidia as pushing further into inference chips. That is the outlet's reading, not a stated Nvidia plan.

Rebellions AI funding, backers and the $2.3bn valuation

Rebellions has raised roughly $850m from SK Hynix, Samsung Ventures and Arm, and The Next Web gives its most recent valuation as around $2.3bn. It was founded in 2020 and is based in Bundang, south of Seoul. It designs neural processing units (NPUs, chips built for AI workloads) for data centers, tuned for inference.

Inference is the work of running an AI service after the model has been trained. Each query takes less compute than training does, but across all users the total is far larger.

The company also holds a direct investment from the Korean government. In September 2025 Rebellions announced a Series C round, with Arm and Samsung Ventures taking part, according to Tech Times via Zaikei.

Rebellions timeline from its own site

Date	Event
September 2020	Founded in Korea
June 2022	ATOM chip taped out (design finished and sent to manufacturing)
May 2023	ATOM first shipped, to kt cloud
November 2024	REBEL chip taped out
December 2024	Merger with SK Sapeon
February 2025	Rebellions Japan established, the first overseas office
August 2025	Saudi Arabian subsidiary established

Is Rebellions public? A pre-IPO round, no listing date

Not as of the latest reports: Rebellions raised more than $400m in March 2026 in a round it called pre-IPO, and no listing date has been disclosed. Tech Times via Zaikei reports that the round was led by Mirae Asset Financial Group and the Korea National Growth Fund.

Chief financial officer Sungkyue Shin said a year before the 21 August report that going public was the "master plan". The company launched two rack-scale products with the March round, RebelRack and RebelPod. The Next Web adds its own view that a license from Nvidia would give the company cash without a listing.

Where Rebellions chips are deployed

ATOM and the higher-performance ATOM-Max entered mass production in 2023, and The Next Web cites DataCenterDynamics as reporting that customers in Japan, Saudi Arabia and the United States have deployed them. EE Times lists Korea Telecom's NPU-as-a-service among four deployments and says the chips also run in Saudi Arabia's sovereign AI infrastructure.

In Korea, SK Telecom said in September 2026 that Rebellions NPUs serve four of its services, including call summaries, and handle more than 14 million AI requests a day. In Japan, Tomen Devices started a trial with the AI-infrastructure operator Tomorrow Net in July 2026, testing throughput, response time, power efficiency and cost under realistic load.

In September 2026 Rebellions announced a partnership with the Japanese AI platform company ai&, which is starting to deploy RebelRack in a Tokyo data center and aims to scale to 100 or more units. That is a plan, not a completed rollout.

Rebel100 pairs 144GB of HBM3e with Samsung's 4nm process

Rebel100 is Rebellions' data-center inference chip, and it carries 144GB of HBM3e on a Samsung 4nm process. HBM3e is high-bandwidth memory, stacked memory chips that sit next to the processor. Rebellions publishes the specification on its home page.

Rebel100 published specification

Property	Value
Compute (dense)	2 PFLOPS (FP8), 1 PFLOPS (FP16)
On-chip SRAM	512MB at 192TB/s
HBM3e	144GB at 4.8TB/s
UCIe interface	4TB/s (3 × 1TB/s channels per chiplet)
Chip-to-chip	1.6TB/s Ethernet (I/O die due 1Q 2027)
Process	Samsung 4nm

Tech Times via Zaikei says Rebel100 links four chiplets (small chips inside one package) over UCIe, an industry standard for that connection. The software supports PyTorch, vLLM and Triton. Zaikei describes Rebellions' design view as follows: for large language model inference, how fast the model's weights can be read from memory affects throughput as much as raw compute does.

EE Times says the design is memory-centric. Its compute chiplets carry large on-chip SRAM for the memory-heavy decode stage, while prefill runs elsewhere. EE Times adds that Rebellions co-designs custom HBM with Samsung and SK Hynix.

A RebelRack holds 32 Rebel100 cards, with a stated 64 PFLOPS of FP8 compute and a maximum power draw of 28kW. Those are the maker's specifications. Jon Peddie Research praises the design but says independent performance comparisons are still scarce.

Nvidia's Groq deal: a $20bn license, a $3.5bn valuation

The Next Web describes the pattern in Nvidia's recent deals: buy a license, hire the engineers, take a stake, and leave the company standing. The Next Web calls this structure license-and-hire. If the Rebellions talks produce a deal, it would be the fourth arrangement of this shape in under a year.

Nvidia's license-and-hire arrangements and the Rebellions talks

Company	Structure	Reported terms
Groq	Non-exclusive license	$20bn; most of the engineering talent absorbed; Groq kept operating
Enfabrica	Similar arrangement	Reported by The Decoder
Poolside	License plus hiring	$6bn for the model factory; 109 staff hired
Rebellions	Under discussion	Partnership, investment or acquisition; undecided

For Groq, the reports cover what happened next. Nvidia paid $20bn for a non-exclusive license and took most of Groq's engineers. Nvidia said it was not an acquisition because Groq continues to run its cloud business independently.

Groq then raised $650m for what remained. In August 2026 it closed $350m at a $3.5bn valuation, with Nvidia joining the round. Silicon Republic noted that the valuation fell by nearly half at that round.

Nvidia unveiled the Nvidia Groq 3 language processing unit (LPU, an inference chip) at its GTC conference in March 2026 and said it ships in the second half of the year. The Next Web says Rebellions designs the same class of part. The outlet draws that parallel itself, and it is not a forecast for Rebellions.

Poolside and Enfabrica used the same structure

In the same week as the Rebellions report, Nvidia agreed to pay $6bn to Poolside for its model factory and to hire 109 of its staff. NVIDIA poolside deal: $6bn license, not a buyout covers those terms, and The Decoder reported the Enfabrica arrangement. The Next Web adds that rivals are moving the same way, citing AMD's purchase in August of Taalas, a startup that etches models into silicon.

Antitrust review and Seoul could complicate a deal

The Next Web points to two possible complications: US antitrust review and the Korean government's view of advanced chips. Bloomberg noted that Nvidia's share of frontier-training chips is large enough that it has to watch for antitrust scrutiny, and sizeable technology acquisitions usually need US Department of Justice clearance. The Next Web's view is that the license-and-hire structure sidesteps the review a purchase would trigger.

Seoul is the second issue. The Next Web says the Korean government treats advanced semiconductors as strategic national assets, and Samsung and SK Hynix work closely with it on investment projects. SK Hynix is also a Rebellions shareholder.

What is still not established

As of 11 October 2026, no agreement had been reported, and the sources leave four points open. They are:

• Bloomberg's sources said the talks may not lead to a deal.
• No structure, price or stake has been reported.
• Neither company has commented on the record.
• Neither has said what a partnership would cover if it stopped short of an investment.

Nor had a collapse of the talks been reported by that date. Whether a deal would be a license, a minority investment or a purchase is the question the sources did not answer.

Each report cited here is a web page that can change after publication, so saving the whole page, for example as Markdown, lets you check a figure later. The Rebellions pages also show different text by language.

To turn a page URL into plain text with its headings and tables, try the tool that converts a web page into Markdown. The reports cited above are good pages to start with.