Rebellions

Rebellions is a South Korean fabless semiconductor company that designs inference-focused AI chips, accelerator systems, and supporting software. Sunghyun Park and four co-founders established the company in 2020. Rebellions designs its processors while relying on outside manufacturing partners, including Samsung, rather than operating its own semiconductor fabrication plants.^[1][2]

The company first commercialized its ATOM data-center neural processing unit. In December 2024 it completed a merger with SAPEON Korea, retained the Rebellions name, and kept Park as chief executive. Its later REBEL-Quad design, now marketed as Rebel100, moved the product line toward chiplet-based accelerators for large-model inference.^[3][4]

Key facts

Scroll sideways for more →

Founding and merger with SAPEON

Rebellions was founded as an inference-chip company rather than a manufacturer of general-purpose processors. TechCrunch reported that Park, a former Morgan Stanley quantitative developer, founded it with four colleagues. The company's first announced processor was ION, an edge-oriented design that was still in qualification testing and had no commercial customer when TechCrunch reviewed the business in January 2024. ATOM became its first data-center product with a reported customer installation.^[2]

Rebellions' own timeline dates ATOM's tape-out to June 2022 and its first delivery to KT Cloud to May 2023. That sequence is consistent with TechCrunch's report that KT installed ATOM in its cloud NPU infrastructure that May. The distinction matters because a tape-out records completion of a chip design for fabrication, whereas delivery documents a later step toward commercial use.^[1][2]

Rebellions and SAPEON Korea completed their merger on December 2, 2024, about six months after announcing an integration plan. SAPEON had been an AI-processor affiliate of SK Telecom. Both the company announcement and Yonhap reported that the combined organization would use the Rebellions name and that Park would lead it. Yonhap gave an estimated value of 1.3 trillion won for the merged entity at the time; that dated estimate was not a market capitalization or a current valuation.^[3][4]

The merger added SAPEON personnel, intellectual property, and relationships with SK Group companies to Rebellions. The company's announcement framed the combination as a way to expand its chiplet program and overseas activity. Those strategic intentions should be separated from delivered products: the merger itself was completed, but its listed market-expansion goals were forward-looking.^[3]

ATOM product line

ATOM is an inference AI accelerator fabricated on Samsung's 5 nm process and built around GDDR6 memory. Rebellions described the device as a latency-oriented machine-learning system-on-chip. Its product brief lists peak rates of 32 TFLOPS at FP16 and 128 TOPS at INT8, while TechCrunch reported that the product targeted data centers and language models up to about 7 billion parameters.^[2][5][7]

Company engineers presented the silicon in the ISSCC 2024 paper "ATOMUS: A 5 nm 32TFLOPS/128TOPS ML System-on-Chip for Latency Critical Applications." The conference record and paper title establish that an implemented chip was described in an academic venue, not merely a product roadmap. The authors and measurements nevertheless came from Rebellions. The paper is primary technical evidence, not an independent reproduction or a neutral comparison with competing hardware.^[5][6]

ATOM's clearest independently reported early deployment was KT's May 2023 installation. Later Rebellions releases said ATOM and ATOM-Max entered mass production and were deployed with customers in Japan, Saudi Arabia, and the United States, and that the processors powered a large Korean commercial service. The reviewed independent sources did not provide customer-by-customer counts or audit that broader deployment scale, so those regional claims remain attributed to the company.^[2][11]

An August 18, 2026 Korean-language company release added a more specific public-sector example. Rebellions said ATOM-Max cards and servers received an innovative-product designation that allows direct public procurement and that deliveries had begun to Gyeongsangnam-do and Ulsan for an AI CCTV project. It also identified SK Telecom's A.dot call-summary service as a commercial use. The designation and reported deliveries are more concrete than a planned memorandum, but the source does not independently establish shipment volume, uptime, service quality, or comparative performance.^[15]

REBEL-Quad and Rebel100

Rebellions began developing its next generation with Samsung in 2023, aiming at larger generative-model workloads than ATOM could address. At Hot Chips 2025 the company publicly introduced REBEL-Quad, a package built from four homogeneous compute chiplets linked by UCIe-Advanced. In April 2026 Rebellions explicitly described Rebel100 as the product formerly known as REBEL-Quad, establishing that these names refer to the same main processor lineage rather than two unrelated chips.^[2][8][16]

The current Rebel100 page advertises 1,024 TFLOPS of FP16 peak compute, 2,048 TFLOPS of FP8 peak compute, 144 GB of high-bandwidth memory with 4.8 TB/s of bandwidth, two PCIe 5.0 x16 interfaces, and power consumption up to 600 W. Rebellions positions the design for large language models and mixture-of-experts workloads. These are vendor specifications and theoretical rates. They do not by themselves establish application throughput, latency, energy efficiency, cost per token, or parity with a particular Nvidia product.^[8][9]

ServeTheHome observed the processor running on a development board at Hot Chips 2025. It reported a live Llama 3.3 70B demonstration at an average 35.5 milliseconds per output token and photographed the four-chiplet package and PCIe card. This provides independent evidence that working hardware was shown publicly. It was not a standardized benchmark: the report did not supply all model, quantization, batching, prompt, power, software, or comparison conditions, and it expressly identified the system as a development board.^[10]

The company says its software supports PyTorch 2.x, vLLM, and Triton. In March 2026 it also announced RebelRack and RebelPOD, rack and clustered systems based on Rebel100, and described them as available that day. That availability language comes from Rebellions; it is not an independently verified shipment count or evidence that every advertised model and framework configuration had reached production maturity.^[9][13]

An April 2026 memorandum with SK Telecom and Arm illustrates the same boundary between current hardware and future work. The parties said they planned to combine Arm's AGI CPU with a Rebel100-based RebelCard, validate the system in an SK Telecom data center, and then consider broader commercial deployment. The announcement confirms a named collaboration and validation plan, not completed Rebel100 deployment under that program.^[16]

Funding, software, and commercial expansion

Rebellions raised capital in increasingly large rounds as it moved from ATOM toward system-level products. TechCrunch reported a $124 million Series B in January 2024, bringing its total at that point to about $210 million and assigning a roughly $658 million post-money valuation. The company then announced a $250 million Series C in September 2025 at a $1.4 billion valuation, with Arm and Samsung-related investors among the participants.^[2][11]

On March 30, 2026, Rebellions announced a $400 million pre-IPO round led by Mirae Asset Financial Group and the Korea National Growth Fund. Reuters reported the same company-disclosed total of $850 million raised and an approximately $2.34 billion valuation, and said the government-backed fund supplied 250 billion won. The valuation belongs to that financing event. It should not be read as a continuously quoted market value, a sale price, or evidence of subsequent revenue.^[12][13]

Rebellions said in June 2026 that it acquired SqueezeBits, a Korean company specializing in model compression and inference optimization. According to the announcement, the companies had worked together since 2024 on software for Rebellions processors and on community events around vLLM. The acquisition expanded Rebellions' stated scope from processor design toward hardware, optimization, and serving software. No purchase price or other transaction terms were disclosed in the cited release.^[14]

The commercial record therefore has several evidence levels. KT's installation is independently reported; the public-sector deliveries and named A.dot use are specific but company-reported; the Japan, Saudi Arabia, and United States deployments are broad vendor statements; and the SK Telecom-Arm work was still a planned validation. Demonstrations, framework compatibility, public-procurement eligibility, and memoranda of understanding should not be counted as equivalent to paid production deployments.^{[2][11][15][16]}

Reported Nvidia talks in August 2026

On August 21, 2026, Bloomberg News reported that Nvidia and Rebellions were in early discussions about possible forms of collaboration. Citing people familiar with information that was not public, Bloomberg listed a technical partnership, an investment, or an acquisition as possibilities. It also reported that Nvidia CEO Jensen Huang and Park had discussed a potential tie-up at Nvidia's Santa Clara headquarters during that week.^[17][18]

The report established only that preliminary talks had been described by anonymous sources. It did not confirm that the companies selected any structure or reached an agreement. In particular, it was not evidence of a partnership, investment, acquisition, signed term sheet, board approval, transaction valuation, regulatory filing, closing date, or completed deal.^[17]

A cutoff review of Rebellions' English and Korean newsroom indexes and Nvidia's official newsroom and blog channels on August 22 found no first-party announcement confirming the reported discussions or a transaction. The absence of an announcement is time-bounded and could change after the cutoff. Until a primary statement or filing appears, Bloomberg's report should remain attributed as preliminary private-source reporting rather than incorporated into Rebellions' completed corporate history.^[17][19][20]

Evaluating specifications and claims

Rebellions has published unusually detailed processor specifications, conference papers, and public demonstrations for a private startup, but those materials answer different questions. The ATOMUS paper documents a company-authored chip design and measured test setup. The Rebel100 product page states theoretical peak capabilities. ServeTheHome independently observed working REBEL-Quad hardware, but its single live demonstration was not a controlled product comparison.^{[5][6][8][10]}

Peak TOPS or TFLOPS values are not interchangeable with model-serving performance. Useful comparisons require the same model, precision, quantization, batch size, input and output lengths, latency target, software versions, power boundary, and availability state. Rebellions' claims of GPU-class performance, better performance per watt, or lower acquisition cost were not independently established by the sources reviewed and are therefore not treated as measured facts here.^[7][8][13]

Funding and deployment claims also require labels. Reuters corroborated that the company announced the March financing and reported the disclosed investors and amounts, but it did not turn the round valuation into a public-market price. Likewise, a delivered card, a product marked available by its vendor, a development-board demo, and an announced validation plan are separate milestones. Preserving those distinctions gives a more accurate account of Rebellions than combining them into a single claim of market adoption.^{[10][12][13][16]}

As of August 22, 2026, public evidence supports a company that had shipped an earlier data-center accelerator, completed a domestic merger, shown working next-generation silicon, raised substantial private capital, and expanded into inference software. It does not independently establish the full scale of the company's current deployments or the comparative economics of Rebel100, and it does not confirm any transaction with Nvidia.^{[2][3][10][12][14][17]}

References

1. ^Rebellions. "About." Accessed August 22, 2026. rebellions.ai/...about
2. ^Kate Park. "Rebellions lands $124M to develop its new AI Rebel chip with Samsung." TechCrunch, January 29, 2024. techcrunch.com/...hip-rebel-in-collab-with-samsung
3. ^Rebellions. "Rebellions and SAPEON Korea Complete Merger, Launching Korea's First AI Chip Unicorn." December 2, 2024. rebellions.ai/...hing-koreas-first-ai-chip-unicorn
4. ^Kim Na-young. "AI chipmakers Rebellions, Sapeon Korea complete merger." Yonhap News Agency, December 2, 2024. en.yna.co.kr/...AEN20241202003000320
5. ^Rebellions. "Korean AI Chipmaker Rebellions Unveils Mass-Production-Ready 'ATOM' at ISSCC 2024, setting new standards for AI acceleration." March 21, 2024. rebellions.ai/...new-standards-for-ai-acceleration
6. ^Changhyeon Yu et al. "ATOMUS: A 5 nm 32TFLOPS/128TOPS ML System-on-Chip for Latency Critical Applications." 2024 IEEE International Solid-State Circuits Conference, February 2024. doi.org/...ISSCC49657.2024.10454509
7. ^Rebellions. "ATOM 5nm Versatile Inference SoC" product brief, version 3.1. rebellions.ai/...lions_ATOMProduct_Brief_v.3.1.pdf
8. ^Rebellions. "Rebellions Debuts REBEL-Quad at Hot Chips 2025, breaking AI's Energy Tax with High-Performance Chiplet Innovation." August 27, 2025. rebellions.ai/...gh-performance-chiplet-innovation
9. ^Rebellions. "Rebel100." Accessed August 22, 2026. rebellions.ai/...rebel100
10. ^Patrick Kennedy. "Rebellions REBEL-Quad UCIe and 144GB HBM3E Accelerator at Hot Chips 2025." ServeTheHome, August 24, 2025. servethehome.com/...-accelerator-at-hot-chips-2025
11. ^Rebellions. "Rebellions Raises $250 Million to Advance the Next Generation AI Infrastructure, Backed by Arm and Samsung." September 30, 2025. rebellions.ai/...ructure-backed-by-arm-and-samsung
12. ^Heekyong Yang. "South Korea's AI chip startup Rebellions raises $400 million in latest funding round." Reuters, March 30, 2026. uk.finance.yahoo.com/...-ai-chip-startup-130414250
13. ^Rebellions. "Rebellions Closes $400 Million Pre-IPO and Launches RebelRack and RebelPOD to Accelerate Global Expansion." March 30, 2026. rebellions.ai/...od-to-accelerate-global-expansion
14. ^Rebellions. "Rebellions Acquires AI Inference Optimization Leader SqueezeBits to Deliver Full-Stack AI Infrastructure." June 30, 2026. rebellions.ai/...ns_squeezebits_acquisition_260630
15. ^Rebellions. "ATOM-Max cards and servers designated as excellent R&D innovative products." August 18, 2026. kr.rebellions.ai/...procurement_innovation_product
16. ^Rebellions. "Rebellions Collaborates with SK Telecom and Arm Targeting Sovereign AI and Telecom Infrastructure." April 10, 2026. rebellions.ai/...ign-ai-and-telecom-infrastructure
17. ^Yoolim Lee. "Nvidia in Talks With Chip Startup Rebellions for Potential Deal." Bloomberg News, August 21, 2026. news.bloomberglaw.com/...llions-for-potential-deal
18. ^Bloomberg Business. "Nvidia Corp. is in early discussions with the Korean AI chip designer Rebellions." X, August 21, 2026. x.com/...2090711775247278091
19. ^Rebellions. "Newsroom." Accessed August 22, 2026. rebellions.ai/...newsroom
20. ^Nvidia. "Nvidia Newsroom." Accessed August 22, 2026. nvidianews.nvidia.com