NVIDIA, Rebellions and SK Telecom have announced a collaboration using NVIDIA NVLink Fusion to connect Korean-designed processors to NVIDIA's rack-scale platform for sovereign AI. For UK buyers weighing sovereign AI build or rent decisions, it's an early signal of a non-hyperscaler path worth tracking closely.
View the data behind this chart
| Phase | Starts (week) | Duration (weeks) |
|---|---|---|
| Collaboration announced | 0 | 4 |
| Design, validation & build | 4 | 70 |
| R200 deployment window | 74 | 30 |
What was actually announced
NVIDIA, Korean NPU designer Rebellions and SK Telecom announced plans to connect Rebellions' next-generation Rebel R200 neural processing units to NVIDIA's open infrastructure using NVLink Fusion, with SK Telecom planning to deploy the resulting systems across its AI data centre, proprietary models and public-facing services under Korea's "AI for All" project.
The arrangement is supported by Korea's Ministry of Science and ICT, and leans on NVIDIA's MGX rack architecture for system development. The MGX ecosystem is intended to provide access to an established supply chain of ODMs and component suppliers, while SK Telecom plans a disaggregated inference approach — NVIDIA GPUs handling prefill, Rebellions NPUs handling decode — planned to use shared MGX design elements, including rack footprint, power delivery and cooling.
Why this matters beyond Korea
The headline implication for UK infrastructure buyers is that NVLink Fusion opens a route for non-NVIDIA silicon to sit inside NVIDIA MGX-based rack designs and NVLink scale-up fabric, rather than forcing a binary choice between an all-NVIDIA stack and a fully separate vendor ecosystem.
That matters because it's a live example — albeit outside the UK — of a government-backed sovereign infrastructure programme using Korean silicon alongside NVIDIA infrastructure. Anyone building a UK sovereign AI index business case should note it as evidence that semi-custom architectures are moving from slideware to named, government-backed development programmes, even if the UK has no equivalent announcement yet.
The limits of the independence claim
NVLink Fusion is not a clean break from NVIDIA. It brings third-party silicon into NVIDIA's NVLink and MGX ecosystem while retaining dependence on NVIDIA interconnect, systems integration and relevant software compatibility — it doesn't replace them. NVIDIA's cited scale-out path includes ConnectX-8 SuperNICs, Spectrum-X Ethernet and Quantum-X800 InfiniBand, subject to the final system design.
UK buyers should separate two different kinds of independence: reducing reliance on buying NVIDIA GPUs specifically, versus reducing reliance on NVIDIA's interconnect, rack qualification process and software ecosystem. NVLink Fusion addresses the first far more than the second. Published figures — NVLink 6 offering 3.6 TB/s per XPU and 260 TB/s aggregate across a 72-accelerator domain, scaling toward 1,152 devices in future configurations — are described by NVIDIA as preliminary and subject to change, so procurement teams shouldn't treat them as locked specifications.
Capex and procurement implications
There is no public NVLink Fusion pricing, and NVIDIA states features, availability and specifications may change without notice. That means any UK capex comparison needs supplier quotations covering custom silicon, NVLink components, MGX racks, networking, power and cooling — not a headline accelerator price.
The announced model is semi-custom rather than a documented off-the-shelf sovereign appliance. NVIDIA identifies MediaTek, Marvell, Alchip Technologies, Astera Labs, Synopsys and Cadence as early ecosystem participants, with design services and solutions available through the partner ecosystem; no packaged UK sovereign appliance is identified in the cited material. That introduces chip-design, validation and integration work that can shift both budget and timeline versus simply financing an on-premise AI build around existing, qualified systems.
Power, timing and the 2028 reality check
Facility constraints remain the bigger near-term limiter for most UK projects. A separate UK AI-supercomputer plan involving 23,040 NVIDIA GB300 GPUs was reported by The Register at 50MW, expanding to 90MW, as a planned capacity figure rather than an independently verified operating load — a reminder that rack density, power delivery and cooling decide feasibility as much as chip choice does.
The Rebellions R200 systems referenced in this announcement are not expected to be ready for deployment until 2028. UK organisations needing capacity before then face a choice: deploy currently available NVIDIA systems or existing alternative accelerators today, or plan an interim build with custom-silicon integration layered in later, as a possible procurement path conditioned on workload, supplier availability and integration feasibility. Either way, the workload decision belongs in sizing exercises alongside tools like the AI GPU sizing calculator.
Where the near-term business case is strongest
Rebellions' R200 is described as optimized for inference, so the announcement does not provide evidence that it is intended to replace training clusters. That points UK buyers toward evaluating specialised XPU architectures for high-volume inference workloads specifically, where the discipline of procuring servers for AI inference already applies, rather than assuming any semi-custom silicon displaces GPU training clusters.
Sovereignty also doesn't guarantee lower cost: UK policy commentary notes that public-sector and sovereign buyers may have smaller budgets and may rely more on major-cloud-provider discounts. Teams evaluating this path alongside standard NVIDIA DGX systems or a conventional server configuration refresh should model tokens-per-joule, memory footprint and software-porting cost, not just peak interconnect bandwidth, before committing capex to a semi-custom route.
- 01NVIDIA Newsroom — NVIDIA, Rebellions and SK Telecom Announce Collaboration to Advance Sovereign AI Infrastructure With NVIDIA NVLink Fusion · 8 November 2026
- 02NVIDIA — NVLink Fusion product page · 8 November 2026
- 03NVIDIA — NVLink scale-up networking · 8 November 2026
- 04NVIDIA Investor Relations — NVIDIA Unveils NVLink Fusion for Industry to Build Semi-Custom AI Infrastructure · 19 May 2025
- 05NVIDIA Developer Blog — NVIDIA NVLink: The Scale-Up Network for AI Factories · 8 November 2026
- 06The Register — UK AI supercomputer power capacity report · 19 October 2026
- 07Dark Reading — UK tech sovereignty AI policy · 8 November 2026
- 08nvidia.com