Rebellions suits readers tracking alternatives to Nvidia for running trained models rather than training them. Sunghyun Park leads it as CEO. Its first customer for ATOM was kt cloud, and SK Telecom has run its A.X K1 model on RebelServer. Expect hardware specs and partner news, not a consumer product.
Rebellions is a South Korean fabless semiconductor company founded in 2020 that designs AI accelerators for inference. Its chips are ATOM and Rebel100, sold alone or as RebelServer, RebelRack and RebelPOD systems. It merged with SK Sapeon and has offices in Seongnam, Santa Clara, Tokyo and Riyadh.
Founded: 2020 · HQ: Seongnam, Gyeonggi-do, South Korea · CEO: Sunghyun Park · Funding: $850M total raised across a Series B, a Series C and a pre-IPO round (latest: Mar 2026) · Valuation: About $2.34B (pre-IPO round, Mar 2026)
About Rebellions
Rebellions was founded in September 2020 in Korea by a team that set out to build inference-first silicon instead of adapting general-purpose hardware. Sunghyun Park is co-founder and CEO. The company is a fabless chip designer: it designs AI accelerators and outsources their manufacturing. Its Korean headquarters sits in Bundang-gu, Seongnam, with offices in Santa Clara (Rebellions America Inc.), Tokyo, and Riyadh, which serves as the MENA regional headquarters. The company also says it has set up an entity in Taiwan.
The product line runs from silicon to full racks. ATOM, a GDDR6-based inference accelerator, taped out in June 2022 and first shipped in May 2023 to kt cloud, which put it into a live data center. REBEL, taped out in November 2024, is described by Rebellions as the first UCIe-Advanced AI chiplet and pairs 144GB of HBM3E with scalable silicon. Rebel100, the product built on it, is a four-chiplet SoC rated by the vendor at 1,024 TFLOPS in FP16 and 2,048 TFLOPS in FP8, with 144GB of HBM3E at 4.8TB/s and power use of up to 600W. It supports PyTorch 2.x, vLLM and Triton natively. Above the chip sit RebelServer, RebelRack and RebelPOD, plus the Rebellions Inference Platform software layer.
Recent moves cluster around system-level products and overseas expansion. In March 2026 the company announced RebelRack and RebelPOD, described as AI infrastructure platforms, alongside its latest funding. In July 2026 it announced it had run SK Telecom's sovereign LLM A.X K1 on RebelServer. In September 2026 it announced a partnership with ai& to bring energy-efficient inference infrastructure to Japan.
Rebellions merged with SK Sapeon in December 2024, which the company says unified Korea's AI semiconductor capabilities and deepened its supply chain through SK hynix's HBM position. It opened Rebellions Japan in February 2025 and a Saudi subsidiary in August 2025, following a strategic investment from Aramco. According to TechCrunch, the company closed a $124 million Series B in 2024, a $250 million Series C in November 2025, and a $400 million pre-IPO round in March 2026 led by Mirae Asset Financial Group and the Korea National Growth Fund. Total funding stands at $850 million at a valuation of about $2.34 billion, and an IPO is planned, though the company declined to give timing. The about page names Korelya Capital, DGDV and SK Telecom among its backers.
Its customers and partners skew toward telecom operators, cloud providers and sovereign AI programs. kt cloud was the first ATOM customer, and SK Telecom runs its own model on RebelServer. Chief Business Officer Marshall Choy leads the push into the US, where the company says it plans to court cloud providers, government agencies, telecom operators and neoclouds.
The competitive field is Nvidia first. Rebellions pitches energy efficiency and inference cost per token rather than training, which is also the pitch of Groq in the US. Its home advantage is Korea: local models such as Naver's HyperCLOVA X, Upstage's Solar Pro 4 and LG's K-EXAONE 2.0 are natural workloads for a domestic accelerator, and Korean policy favors homegrown compute. The tradeoff is a software stack and developer base far smaller than Nvidia's CUDA ecosystem.
Rebellions publishes its SDK documentation at docs.rbln.ai and a Model Zoo for developers. It lists no compliance certifications on its public pages, so none are recorded here. Readers comparing it with other providers can start with Nvidia, Groq, Naver and Upstage, or with the Korean models HyperCLOVA X, Solar Pro 4 and K-EXAONE 2.0.
Mission
Purpose-engineered AI accelerators for efficient, high-performance inference at scale.
Products
- Rebel100 (AI inference accelerator (chiplet SoC)): https://rebellions.ai/rebellions-product/rebel100/
- RebelServer (Inference server): https://rebellions.ai/rebellions-product/rebelserver/
- RebelRack (Rack-scale inference infrastructure): https://rebellions.ai/rebellions-product/rebelrack/
- RebelPOD (Production inference cluster unit): https://rebellions.ai/rebellions-product/rebelpod/
- Rebellions Inference Platform (Inference software platform): https://rebellions.ai/rebellions-product/rebellions-inference-platform/
Links
Frequently Asked Questions
How much funding has Rebellions raised?
TechCrunch reports total funding of $850 million as of the latest round, announced in March 2026. About $650 million of that arrived in the six months before the announcement, across a Series C and a pre-IPO round. The company said its valuation sits near $2.34 billion.
What products does Rebellions make in 2026?
Rebellions sells its newest accelerator and the ATOM chip before it, plus RebelServer, RebelRack and RebelPOD for system-scale deployments. It also offers the Rebellions Inference Platform software and an SDK at docs.rbln.ai. It publishes no list prices; buyers contact sales.
Does Rebellions publish security or compliance certifications?
Its public pages list no compliance certifications and no trust center. Rebellions sells hardware, so certifications that apply to SaaS vendors are less relevant than export and supply-chain terms. Ask the sales team for any attestations your procurement needs.
Who are the main competitors of Rebellions?
Nvidia is the default choice for inference and wins on software maturity and developer tooling. Groq competes on low-latency inference, while FuriosaAI is the other Korean inference chip designer. Rebellions pitches energy efficiency and local-market fit, but it loses on ecosystem size.
How do you start using Rebellions hardware?
Contact the sales team through the Let's Talk form on rebellions.ai, since the products are sold to data centers and enterprises. Developers can read the SDK documentation at docs.rbln.ai and browse the Model Zoo first. The newest chip supports PyTorch 2.x, vLLM and Triton.