Nvidia circles Rebellions because the low-power inference tier is not optional

Nvidia is in preliminary discussions with Rebellions, the South Korean AI chip startup, over a possible technical partnership, an investment, or an outright acquisition, according to a Bloomberg report summarised by DataCenterDynamics ^[1][2]. Nvidia CEO Jensen Huang and Rebellions co-founder and CEO Sunghyun Park met at Nvidia's Santa Clara headquarters this week, though the report described the deliberations as preliminary and not guaranteed to produce a transaction ^[3][4]. A Rebellions representative declined to comment to Bloomberg ^[5].

What Rebellions has is narrower and more specific than a general-purpose accelerator business. Founded in 2020, it builds AI inference accelerators pitched on energy efficiency and low-latency performance ^[6], and began mass-producing its Atom and Atom Max NPUs for running large language models in the data center in 2023 ^[7]. The company says those chips have been deployed by customers in Japan, Saudi Arabia and the US, and that it wants to expand in the US, Europe and Asia-Pacific specifically to support sovereign AI infrastructure programmes ^[8]. It has been consolidating: a June 2024 merger with Sapeon Korea, which was spun out of SK Telecom in 2016 ^[9], and a July 2025 partnership with Marvell aimed at APAC and the Middle East ^[10]. In March 2026 it raised $400 million in a pre-IPO round and announced two platforms, RebelRack and RebelPod, with no IPO date disclosed ^[11].

Nvidia already has a low-latency answer, and it has been candid about the cost of it. It announced a $20 billion license agreement with Groq in January 2026 and unveiled the resulting Nvidia Groq 3 LPU at GTC two months later ^[12][13]. Nvidia data center head Ian Buck said in March that the LPU "is optimized strictly for that extreme low-latency token generation, offering token rates in the thousands of tokens per second," with the trade-off 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" ^[14]. The Groq 3 is due in the second half of 2026 in liquid-cooled LPX racks holding 256 LPUs, 128GB of on-chip SRAM and 640 TBps of scale-up bandwidth ^[15] - about 512MB of SRAM per chip ^[2]. LPU shipments are expected to begin in Q3 2026, and Nvidia says it will now ship a new LPU architecture annually alongside its annual GPU cadence ^[16][17].

That is the gap Rebellions fills. Nvidia's own framing puts the LPU at the extreme latency end with poor per-chip economics ^[14], which leaves the energy-constrained, cost-sensitive inference tier to NPU vendors selling exactly what Rebellions sells ^[6]. Buying or funding one of them also acquires the customer list that comes with it, in the same sovereign accounts Nvidia wants. For scale, the Groq license alone was 50 times the size of Rebellions' entire pre-IPO round ^[1].

The timing sits inside noisier news. The Information reported this week that Nvidia had built a China-specific LPU and planned to ship small batches before year-end with orders already placed ^[18]; Nvidia denied it, saying "We have no LPU sales in the China market today, and no China-specific LPU product in our roadmap" ^[19]. After its Q1 2027 results in May, Nvidia said it was booking no anticipated compute revenue from China in its second-quarter projections ^[20].

Watch which of the three structures emerges: a partnership leaves Rebellions' IPO path intact, an acquisition ends it and puts Nvidia's brand on a Korean NPU line. Watch the Marvell agreement ^[10], which an acquisition would have to survive or unwind, and watch whether Q3 2026 LPU shipments arrive on schedule ^[16].