EE Times原文发布 08-24 16:07Radar 收录 08-24 16:46查看原文 ↗

英伟达推理战略转向触及韩国Rebellions

Nvidia’s Inference Pivot Reaches Rebellions in Korea

中文摘要

据报道，英伟达正与韩国AI推理芯片初创公司Rebellions洽谈技术合作、投资或收购事宜。该消息发布于某科技媒体。此动向表明英伟达在推理领域的布局延伸至韩国市场，可能影响相关AI芯片供应链格局。

原文深度解读

本文报道了英伟达（Nvidia）与韩国AI推理芯片初创公司Rebellions的潜在合作、投资或收购洽谈。文章涉及英伟达在推理领域的战略转向，以及Rebellions的芯片架构、HBM合作、产品部署和融资情况。关键数字包括Groq交易金额200亿美元、Rebellions融资约8.5亿美元、估值约23亿美元、政府投资1.66亿美元、REBEL-Quad平台参数等。该动向可能影响AI推理芯片供应链，尤其是HBM供应和韩国AI生态。

英伟达与Groq达成非排他性技术许可协议，并吸收其大部分工程人才，交易金额为200亿美元（2025年底）。
Rebellions的芯片专注于内存中心推理，其计算chiplet携带大量片上SRAM以处理内存密集的解码阶段，预填充在其他地方运行。
Rebellions与三星和SK海力士建立了定制HBM协同设计关系，以获得结构性内存供应优势，这是其他纯无晶圆厂公司无法复制的。
Rebellions成立于2020年，五年内从创立到量产，其ATOM和ATOM-Max芯片运行SK Telecom的韩国最大商业AI服务，并部署于沙特主权AI基础设施。
Rebellions的第二代推理平台REBEL-Quad于2025年8月发布，基于三星4nm工艺，集成四个计算chiplet，采用UCIe-Advanced互连，配备144GB HBM3E，在300W功耗内提供1 POPS FP16算力。
Rebellions已融资约8.5亿美元，投资者包括三星、SK海力士、Arm、SK Telecom、NTT DoCoMo和沙特阿美的Wa’ed Ventures，估值约23亿美元，并计划2027年上半年在韩国IPO。

供应链影响

若英伟达与Rebellions达成合作或收购，可能增强英伟达在推理芯片领域的竞争力，并可能影响其与Groq的现有关系。
Rebellions与三星和SK海力士的HBM合作可能使英伟达获得更稳定的HBM供应，但取决于交易结构是否保留这些合作关系。
潜在交易可能加深英伟达在韩国AI价值链的嵌入，可能影响三星和SK海力士的客户结构，但取决于监管审批和交易条款。
若交易引发反垄断审查，可能延迟或改变交易形式，进而影响相关芯片供应链的整合节奏。
Rebellions的IPO计划可能因潜在交易而调整，可能影响其融资和供应链扩张计划，取决于最终协议。

系统解读，仅作为判断线索，不构成备货、出货、涨价或投资建议。

正文内容 · AI翻译

中文英文

据报道，AI训练巨头英伟达在与Groq达成一项非传统交易后，再次进军推理芯片领域，该交易使英伟达获得了非排他性技术许可，同时吸收了其大部分工程人才。据彭博社报道，英伟达正在与韩国推理初创公司Rebellions进行早期讨论，涉及技术合作、投资，甚至可能是收购。

与Groq一样，Rebellions的芯片专注于内存中心推理；其计算chiplet携带大量片上SRAM，以处理内存密集的解码阶段，而预填充在其他地方运行。这与Groq基于SRAM的LPU模型形成对比，英伟达在2025年底以200亿美元收购了该模型。

但Rebellions的内存中心推理还有更多内容：它与三星和SK海力士建立了定制HBM协同设计关系，以获得结构性内存供应优势，这是其他纯无晶圆厂公司无法复制的。换句话说，凭借这些HBM供应合作伙伴关系，Rebellions可以获得独立AI芯片供应商（没有战略投资者关系）无法保证的内存分配。

此外，这家总部位于韩国首尔的推理初创公司在短短五年内从创立走向量产。其ATOM和ATOM-Max芯片运行SK Telecom的韩国最大商业AI服务，为处理韩国特定服务（如通话摘要）的专有AI助手提供支持。这些推理芯片也已部署在沙特阿拉伯的主权AI基础设施中。

作者：Arduino团队 2026年8月19日

作者：Raphi Zadicario，Lumissil Microsystems产品经理兼首席架构师 2026年8月17日

作者：Isaac Lopez，OmniScale Media总裁 & Charity Plata，SC26通信主席 2026年8月13日

英伟达在用于训练AI模型（如OpenAI和Anthropic的模型）的芯片中占据主导份额，目前正处于向推理领域的战略转向中，该领域尚未完全定型。此外，与Rebellions的潜在交易可能进一步将英伟达嵌入韩国的AI价值链，该价值链目前由内存巨头三星和SK海力士主导。

据报道，Rebellions联合创始人兼CEO朴成贤在加利福尼亚州圣克拉拉的英伟达总部会见了黄仁勋。（来源：Rebellions）

Rebellions的推理价值主张

Rebellions成立于2020年疫情期间，专注于数据中心推理工作负载的神经处理单元（NPU）。其第一代推理芯片——ATOM及其高性能变体ATOM-Max——于2023年进入量产，并获得了四次部署，包括韩国电信的NPU即服务基础设施。

Rebellions的第二代推理平台REBEL-Quad于2025年8月在三星代工的4纳米工艺节点上发布。它集成了四个计算chiplet，采用UCIe-Advanced互连，配备144 GB HBM3E，并在300瓦功耗内提供1 POPS的FP16算力。其旗舰产品Rebel100 NPU在AI推理场景中提供了显著的能效。

基于chiplet的设计实现了超可扩展性，同时不牺牲延迟或一致性。（来源：Rebellions）

AI推理对内存中心架构的大赌注带来了更高的效率和更低的成本。然而，除了Rebellions与Groq和Cerebras相似的SRAM密集型架构外，其软件重点也使其成为更完整的系统公司。这与其它AI芯片供应商形成鲜明对比，后者让开发人员自行解决如何在其芯片上运行AI。

其芯片构建在由Kubernetes驱动的云原生堆栈上，该堆栈是开源的，适用于大多数流行的AI开发框架，包括PyTorch、Hugging Face和vLLM推理引擎。使用无分支的开源软件为云开发人员提供了相同的体验。

英伟达的推理转向

Rebellions联合创始人兼CEO朴成贤在美国度过11年后返回韩国，他在MIT获得了硕士和博士学位，并曾在英特尔、SpaceX和摩根士丹利工作。他将Rebellions视为一家云原生推理公司，旨在与英伟达和AMD等AI芯片竞争对手竞争。

“即使我们进入与英伟达相同的擂台并被击败，我也想挥出一拳，”朴最近表示。Rebellions这个名字象征着对英伟达主导的AI秩序的叛乱；然而，与英伟达的联络可能从根本上改变其作为英伟达挑战者的叙事。

这家无晶圆厂芯片公司迄今已融资约8.5亿美元，其投资者包括三星、SK海力士、Arm、SK Telecom、日本NTT DoCoMo和沙特阿美的Wa’ed Ventures。Rebellions最近估值约23亿美元，并正在筹备可能在2027年上半年进行的韩国IPO。

它是韩国最受关注的AI初创公司之一，也是韩国政府“K-Nvidia”倡议的关键受益者，该倡议旨在构建具有全球竞争力的AI基础设施生态系统。事实上，Rebellions最近从政府的韩国国家增长基金获得了1.66亿美元的首笔直接投资。

韩国政府将半导体视为关键战略资产。因此，类似收购的安排也可能招致韩国的监管审查。同样，英伟达在AI训练芯片领域的主导地位意味着，即使是类似Groq的交易也可能在美国引发反垄断审查。

英伟达和Rebellions均拒绝对彭博社的报道发表评论，该报道指出谈判仍处于早期阶段，交易可能不会实现。但有一点是明确的：英伟达被认为相对较弱的推理市场正在达到临界点。在这个技术十字路口，英伟达正在进行战略转向，以拯救其AI帝国。

另见：

Groq：英伟达对AI推理的200亿美元赌注

为什么英伟达的AI帝国在2026年面临清算

Rebellions押注内存中心架构，权衡IPO选项

AI推理、AI训练、数据中心、并购、半导体、SRAM

英伟达、Rebellions

AI training kingpin Nvidia is reportedly making another foray into inference silicon after engaging Groq in an unconventional deal that secured Nvidia a nonexclusive technology license while absorbing most of its engineering talent. According to Bloomberg, Nvidia is in early discussions with the Korean inference upstart Rebellions for a technical partnership, an investment, or perhaps even an acquisition.

Like Groq, Rebellions’ chips specialize in memory-centric inference; its compute chiplets carry significant on-chip SRAM to handle the memory-intensive decode stage, while prefill runs elsewhere. That draws a comparison to Groq’s SRAM-based LPU model, which Nvidia acquired for $20 billion in late 2025.

But there’s more to Rebellions’ memory-centric inference: It has forged custom HBM co-design relationships with Samsung and SK Hynix to secure structural memory supply advantages that other pure-play fabless firms cannot replicate. In other words, with these HBM supply partnerships, Rebellions can secure memory allocation that independent AI chip vendors without strategic investor relationships cannot guarantee.

What’s more, this Seoul, South Korea-based inference startup has moved from founding to mass production in just five years. Its ATOM and ATOM-Max chips run Korea’s largest commercial AI service at SK Telecom, powering a proprietary AI assistant that handles Korea-specific services such as call summarization. These inference chips have also been deployed in Saudi Arabia’s sovereign AI infrastructure.

By Arduino Team 08.19.2026

By Raphi Zadicario, Product Manager and Chief Architect, Lumissil Microsystems 08.17.2026

By Isaac Lopez, President, OmniScale Media & Charity Plata, Communications Chair, SC26 08.13.2026

Nvidia—which holds a dominant share of chips used to train AI models such as those from OpenAI and Anthropic—is currently in the midst of a strategic pivot to the inference landscape, which is not fully settled yet. Moreover, a potential deal with Rebellions could further embed Nvidia in South Korea’s AI value chain, currently dominated by memory powerhouses Samsung and SK Hynix.

Rebellions co-founder and CEO Sung-hyun Park reportedly met Jensen Huang at Nvidia’s headquarters in Santa Clara, California. (Source: Rebellions)

Rebellions’ inference value proposition

Rebellions, founded during the pandemic in 2020, specializes in neural processing units (NPUs) for inference workloads in data centers. Its first-generation inference silicon—ATOM and its higher-performance variant ATOM-Max—entered mass production in 2023 and secured four deployments, including Korea Telecom’s NPU-as-a-service infrastructure.

Rebellions’ second-generation inference platform, REBEL-Quad, was launched in August 2025 on Samsung Foundry’s 4-nm process node. It integrates four compute chiplets using the UCIe-Advanced interconnect, carries 144 GB of HBM3E, and delivers 1 POPS of FP16 compute within a 300-W envelope. Its flagship product Rebel100 NPU delivers significant energy efficiency in AI inference scenarios.

The chiplet-based design enables hyper-scalability without compromising latency or coherence. (Source: Rebellions)

AI inference’s big bet on memory-centric architectures leads to greater efficiencies and lower costs. However, beyond Rebellions’ SRAM-heavy architecture, which parallels Groq and Cerebras, its software focus also makes it a more complete systems company. That’s in stark contrast to other AI silicon suppliers that leave it to developers to figure out how to run AI on their chips.

Its silicon is built on a cloud-native stack powered by Kubernetes, which is open-source and works with most popular AI developer frameworks, including PyTorch, Hugging Face, and the vLLM inference engine. Using open-source software without forks provides an identical experience for cloud developers.

Nvidia’s inference pivot

Rebellions co-founder and CEO Sung-hyun Park returned to South Korea after spending 11 years in the U.S., where he earned a master’s and a Ph.D. from MIT and worked at Intel, SpaceX, and Morgan Stanley. He sees Rebellions as a cloud-native inference outfit aiming to compete with AI silicon rivals such as Nvidia and AMD.

“Even if we step into the same ring as Nvidia and get beaten to death, I want to throw a punch,” Park said recently. The name Rebellions signifies a revolt against the AI order dominated by Nvidia; however, a liaison with Nvidia could fundamentally change the company’s narrative as an Nvidia challenger.

The fabless chip firm has raised roughly $850 million so far, and its investors include Samsung, SK Hynix, Arm, SK Telecom, Japan’s NTT DoCoMo, and Saudi Aramco’s Wa’ed Ventures. Rebellions, recently valued at around $2.3 billion, is also preparing for a potential Korean IPO in the first half of 2027.

It’s one of South Korea’s most closely watched AI startups and a key beneficiary of the South Korean government’s “K-Nvidia” initiative aiming to build a globally competitive AI infrastructure ecosystem. In fact, Rebellions has recently secured the first direct investment of $166 million from the government’s Korea National Growth Fund.

The South Korean government views semiconductors as critical strategic assets. So, an acquisition-like arrangement may also invite regulatory scrutiny in Korea. Likewise, Nvidia’s dominant position in AI training silicon means that even a Groq-like deal could trigger antitrust reviews in the U.S.

Both Nvidia and Rebellions have declined to comment on the Bloomberg story, which noted that negotiations are still in the early stages, and the deal may not materialize. But one thing is clear: The inference market, where Nvidia is seen as relatively weak, is reaching a tipping point. Here, at this technology crossroads, Nvidia is making a strategic pivot to save its AI empire.