NVIDIA, Rebellions and SK Telecom Bet on Heterogeneous AI: GPUs Meet Korean NPUs via NVLink Fusion

Nothing here is sponsored — no affiliate links on this page. This breaking article is not the daily buyer-intent post. Facts below come from the sources listed at the end, checked 9 November 2026. We have not tested any of this hardware hands-on, and nothing here is a hands-on finding.

NVIDIA, Rebellions and SK Telecom announced today that they are building Korea's sovereign AI infrastructure together — not on NVIDIA GPUs alone, but on a mix of NVIDIA GPUs and Korean-designed NPUs, connected inside the same rack. The deal was signed at NVIDIA AI Day Seoul on Monday, 9 November 2026, and it is backed by Korea's Ministry of Science and ICT.

The headline is simple. The interesting part is the plumbing: Rebellions' next-generation R200 neural processing units will connect to NVIDIA's platform through NVLink Fusion, inside NVIDIA's MGX rack architecture, and SK Telecom plans to deploy those mixed systems in its AI data centers. I have not tested any of this hardware — nobody outside the three companies has — and this article is built from the announcement itself, checked 9 November 2026. Where a claim is the companies', I'll say so.

What was actually announced

This is a collaboration under a memorandum of understanding, not a product launch. The three parties agreed to:

• Connect Rebellions' R200 NPUs to NVIDIA's platform via NVLink Fusion, and build rack-scale systems on the NVIDIA MGX rack architecture.
• Deploy those systems in SK Telecom's AI data centers, drawing on SKT's data center operations, its proprietary models (including A.X K2), and its public-facing AI services under Korea's government-backed "AI for All" project.
• Develop a disaggregated inference approach: NVIDIA GPUs handle prefill, Rebellions NPUs handle decode — each stage of generating an AI response assigned to the processor architecture best suited to it.

The Ministry of Science and ICT (MSIT) is supporting the collaboration, and Korea's Deputy Prime Minister and Minister of Science and ICT, Bae Kyung-hoon, framed it as Korean-developed NPUs entering the computing infrastructure market alongside NVIDIA. Jensen Huang's line in the release: "Sovereign AI gives nations the power to turn their language, culture and expertise into intelligence."

One date to hold onto: R200 systems are expected to be ready for deployment in 2028. That is two years away. Nothing here ships today.

What NVLink Fusion actually does

NVLink Fusion is the part of this story that matters most to developers and infrastructure builders, because it is the mechanism that makes the rest possible.

Until recently, NVLink was NVIDIA's own interconnect — the fast link between NVIDIA GPUs inside a server or rack. NVLink Fusion extends that platform outward: it lets chipmakers connect their own custom processors — what NVIDIA calls XPUs — into NVIDIA's rack-scale infrastructure, using NVIDIA's high-bandwidth, low-latency scale-up networking.

In practical terms, that means a Korean NPU no longer has to live in a separate, walled-off system with its own networking, its own rack design, and its own supply chain. It can plug into the same MGX rack architecture that NVIDIA's GPUs use, share the rack footprint, power delivery and cooling, and talk to the GPUs over NVLink. The MGX ecosystem also gives Rebellions access to an established supply chain of original design manufacturers and component suppliers — which, for a startup shipping its first rack-scale system, is arguably as valuable as the interconnect itself.

This is NVIDIA's "vertically integrated, horizontally open" pitch: NVIDIA keeps control of the platform, and everyone else's silicon is welcome inside it. Whether you read that as openness or as a very clever moat depends on where you sit. Both readings are honest.

Why a GPU + NPU mix matters

The obvious question: why not just use more GPUs? The answer is in the workload split the companies describe.

Running a large language model has two very different phases. Prefill — processing your prompt — is compute-heavy and parallel: a job GPUs are built for. Decode — generating tokens one at a time — is memory-bandwidth-bound and latency-sensitive, and it is where a purpose-built inference chip can win on throughput per watt and cost per token. Rebellions' whole pitch is inference efficiency: its flagship Rebel100 is built specifically for inference, and the R200 is, per the announcement, optimized for low-latency, energy-efficient inference on frontier models, including multitrillion-parameter mixture-of-experts models.

So the planned split — GPUs for prefill, NPUs for decode — is not a marketing diagram. It is the textbook case for heterogeneous computing: use each chip for the phase it is best at, and the system as a whole gets cheaper and faster than either chip alone. If it works at rack scale, the savings compound across a whole data center: fewer watts per token, more tokens per rack.

That "if" is doing real work, though. Disaggregated inference across two different chip architectures in the same rack is a systems problem, not just a chip problem — scheduling, memory movement between stages, failure handling, and the software stack that has to treat two vendors' silicon as one system. The announcement says the companies "plan to develop" this approach. That is the right verb. It is not built yet.

What is real now vs what is still future

Let me draw the line plainly, because launch-day coverage rarely does:

• Real now: the agreement itself, the government backing, and the technology pieces that already exist — NVLink Fusion as a platform, the MGX rack architecture, Rebellions' Rebel100 in commercial deployments, and SK Telecom's operating data centers and models.
• Still future: the R200 chip (a next-generation product), the mixed GPU+NPU rack systems, the disaggregated prefill/decode inference pipeline, and the deployments — all aimed at 2028.
• Not in the announcement at all: performance numbers, power figures, prices, or any benchmark. There is no claim here you can test against a GPU-only system yet, and no independent evaluation exists. Any article that tells you how much faster or cheaper this will be is inventing numbers.

What it means for sovereign AI beyond Korea

The Korea angle is specific, but the pattern is general — and that is why this announcement is worth your attention even if you never touch a Korean data center.

Sovereign AI, as NVIDIA sells it, means a country running its own models, on its own infrastructure, tuned to its own language and priorities. Every country attempting it hits the same wall: building the whole stack yourself — chips, interconnect, racks, models, data centers — is impossibly expensive, and buying it all from one vendor leaves you dependent on that vendor. NVLink Fusion is NVIDIA's answer to that tension: your chips, our rack. Korea gets to say its silicon is inside the system; NVIDIA keeps the system itself.

If this works in Korea, expect the template to travel. Other countries with domestic chip ambitions — and there are many — will be watching whether a startup NPU can genuinely share a rack with NVIDIA GPUs and deliver the promised efficiency. Rebellions is backed by an interesting list — Aramco, Arm, KT, Mirae Asset, Samsung, SK Hynix and SK Telecom — which tells you this is not a one-country science project; it is a bet that Korean inference silicon becomes an export.

The risk, stated plainly, is lock-in by another name. A rack built on MGX and NVLink Fusion is still NVIDIA's platform, and the software and interconnect standards are NVIDIA's. Sovereignty over the chip is not sovereignty over the stack. Countries making this trade should know exactly which layer they control and which layer they rent.

The catch, stated plainly

• Nothing ships until 2028. This is a memorandum of understanding and a development plan. The R200 does not exist as a deployable product yet, and the mixed systems are to be built.
• No numbers. No performance, efficiency, power or price figures were published. The efficiency case for the GPU+NPU split is a design argument, not a measured result.
• Disaggregated inference is hard. Splitting prefill and decode across two architectures in one rack is a real systems-engineering challenge. The announcement commits to developing the approach, not to having solved it.
• It is still NVIDIA's platform. NVLink Fusion opens the rack to third-party silicon, but the interconnect, the rack architecture and the ecosystem are NVIDIA's. That is the trade Korea is making, eyes open.
• No hands-on testing — mine or anyone independent's. This article is built from the companies' announcement, checked 9 November 2026. Nothing here is a hands-on finding.

The bottom line is a measured one. This is not a chip launch. It is a structural announcement: the first time a national sovereign-AI program has publicly bet on heterogeneous computing — NVIDIA GPUs and a domestic NPU, in the same rack, on the same interconnect — with government backing and a 2028 deployment target.

The idea is sound. Prefill on GPUs, decode on specialized NPUs, is the right split on paper, and NVLink Fusion is the first credible mechanism for making it happen inside NVIDIA's rack-scale world rather than outside it. But the proof is two years away, and until there are systems, benchmarks and watts-per-token figures, the honest position is interested, not convinced.

If you build AI infrastructure, the thing to watch is not the press release — it is whether Rebellions ships an R200 system that runs in an MGX rack, and what the measured numbers look like when it does. Until then, this is a bet worth tracking, and Korea just made it in public.

Sources: GlobeNewswire / NVIDIA press release, 9 November 2026, "NVIDIA, Rebellions and SK Telecom Announce Collaboration to Advance Sovereign AI Infrastructure With NVIDIA NVLink Fusion" (TradingView mirror); Yonhap News, 9 November 2026 (Yonhap). All capability, performance and deployment claims are the companies', stated as such. No hands-on testing was performed for this article. This breaking article is not the daily buyer-intent post, and it contains no affiliate links.