While we are fond of compute engines, networking chips, and storage devices here at The Next Platform, what ultimately matters is how system architects take these components and weave them into systems that can do very specific and real things that have economic, cultural, and biological consequences. Some system architects like to integrate best of breed components on their own to get some level of co-design control, there are others who just want to buy a complete system that they know they can scale up as their jobs grow and want to get an as quickly as possible.

The next phase in the expansion of South Korean AI chip startup Rebellions AI is all about catering to the system buyers, which is fitting given that the company has aspirations of selling lots of AI accelerators and complete systems outside of its indigenous market. This is particularly true in those sovereignties that do not have indigenous AI accelerators in development or production – which means most countries in the world – or who want to hedge their bets by using a mix of AI accelerators. Given the high cost and scarce supply of systems based on Nvidia engines, everyone is also looking for a scalable architecture that won’t break the balance sheet.

That expansion by Rebellions AI into bringing full systems to market – with the help of ODMs and very likely a few OEMs – outside of Korean is going to be funded in part by a $400 million Series D funding round that Rebellions AI just landed. The company is calling this its “pre-IPO” funding round, which means we will see Rebellions AI go public sooner rather than later. No dragging this out over Series E, F, and G as we have seen some startups do. It will be interesting to see when and where Rebellions AI goes public; the company is not talking about that yet.

Rebellions AI just completed its Series C funding in September 2025, which is not that long ago, with the $250 million investment being made by Arm, the CPU intellectual property company that has just launched its AGI CPU for datacenter servers and that has not, thus far, been tempted to make its own GPU or XPU for AI acceleration. Which is good for AI startups like Rebellions AI, which need to have tight coupling with an Arm-based CPU as well as an X86 host option, which mirrors what Nvidia does in its system designs.

In this Series D round, Mirae Asset Financial Group, which is located in the capital city of Seoul and which does asset and wealth management, investment banking, and life insurance, with $730 billion in assets under management, is the lead investor. Because of disclosure laws in South Korea, we know that it kicked in around $199 million in this round; Mirae was an investor in the Series B round of funding for Rebellions AI two years ago. The South Korean National Growth Fund, which back in March set up a massive technology funding effort called the “K-Nvidia Nurturing Project,” plans to spend $99.5 billion over the next five years, with about a tenth of that being allocated to stakes in AI-related startups. The K-Nvidia effort’s first funding went to Rebellions AI, and it comes in at around $166 million. The Korea Development Bank kicked in another $33 million, and existing investors did the rest of check writing.

Total funding since Rebellions AI was founded in September 2020 comes to more than $850 million. Things are moving fast, and its valuation is currently around $2.34 billion. The company now has in excess of 300 employees and is expanding internationally to chase opportunities outside of Korea.

To be specific, this means getting Rebel100 compute engines, RebelRack systems, and full pods of Rebel machines (called RebelPod, naturally) into the hands of clouds, neoclouds, telco companies, other service providers, and AI labs in the United States, where the appetite for cheap AI inference is at a fever pitch right now.

We did a deep dive on the Rebel compute engine architecture back in late December 2025, and it is significant that Korean memory makers Samsung and SK Hynix are investors in Rebellions AI, that Samsung is etching the Rebel compute engines, and that Arm is also an investor. Up until now, the Rebel systems have used AMD Epyc X86 processors as their host processors, but the company was at the Arm AGI CPU launch two weeks ago and is an enthusiastic supporter of Arm’s homegrown, AI-tailored server CPU for RebelRack and RebelPod systems in the near term for those who want to have an Arm CPU in the host like Nvidia and many hyperscalers and clouds do for the AI systems.

Here’s the idea:

We look forward to an explanation of that around 2X better TPS per watt note showing the Arm AGI CPU versus the AMD Epyc CPU.

Here’s a demo that shows the two compute engines working on an AI workflow, which was shown at the Arm Everywhere event back on March 24:

“Arm's an investor in us, and that continues that theme of strategic relationships that go beyond just investment – deals that have strategic aspects to them,” Marshall Choy, chief business officer of Rebellions, tells The Next Platform. “Let's be honest, in this AI world today, there is lots of capital being brought into the mix. We are focused on what else we can you achieve through those partnerships. And having multiple points of engagement, I think, is much more useful, not only for ourselves, but for whoever we are partnering with in the ecosystem.”

The RebelServer, as you can see above, has eight Rebel100 AI inference engines in the box, supported by a pair of CPUs. This 2 x 8 design is bog standard in commercial-grade, air-cooled AI systems, even today. (The rackscale machines sold by Nvidia are still pretty exotic for all but the most advanced hyperscalers, cloud builders, and AI model builders.)

The RebelRack has four of these nodes, for a total of 32 accelerators, and they are spaced out enough that the rack can still be cooled by datacenter chilled air. (Higher density will presumably require liquid cooling.) Most organizations do not have datacenters with liquid cooling, so this is a good place to start. The more important idea, says Choy, is that 48 hours after machines hit the floor, they are hooked into data and are doing inference.

The RebelRack system brings 16 petaflops of AI inference oomph at FP8 precision and 8 petaflops at FP16 precision. Each Rebel100 has 512 MB of SRAM and 144 GB of HBM3E stacked memory with 4.8 TB/sec of memory bandwidth, so the RebelRack has 4 GB of SRAM in total, with 4.5 TB of aggregate HBM3E memory and 153.6 TB/sec of aggregate HBM bandwidth. Each server node has 400 GB/sec of Ethernet bandwidth (eight ports at 400 Gb/sec), and the Rebel100 cards are hooked to each other and to the host CPUs with a PCI-Express 5.0 all-to-all interconnect. The typical power draw of this RebelRack is 5 kilowatts, with a maximum power draw of 7 kilowatts.

This picture of the RebelRacks shows them at a higher density:

With this picture of the RebelPod, you can't see inside the racks, so it is hard to say how many machines are in each rack:

RebelPod takes a bunch of RebelRacks and interlinks them with an 800 GB/sec Ethernet back-end network, with a 25 GB/sec front-end network linking out to the data residing in other systems. The RebelPods come in configurations of two, four, six, eight, and sixteen racks, ranging from 64 to 1,024 accelerators in a single system image for running very large inference workloads.

We look forward to seeing how this machinery stacks up on real-world AI inference workloads. Rebellions AI says its architecture was designed for massive scale, and with the latest round of funding, it has the chance to prove it.