The disaggregated split is the only architectural detail the document commits to. GPUs handle Prefill, computing the internal states from the input prompt. Rebellions' R200 NPUs handle Decode, generating the response token by token.

The constraint runs one direction. Third-party ASICs require a Nvidia-CPU in the rack, which makes the open-platform framing a dependency rather than a choice. Rebellions is architecturally locked to NVIDIA's interconnect regardless of how the partnership is described.

No die size, no memory bandwidth, no interconnect bandwidth, no token-per-second figure appears anywhere in the document. The stated goals are higher inference throughput and better energy efficiency, and the document stops there.

The rack itself is built on NVIDIA's MGX reference architecture, a modular platform for standardized server and rack design. SK Telecom will connect the systems to its own AI models and services, and the Korean government supports the collaboration.

NVLink Fusion is drawing a growing list of adopters beyond this deal. AWS plans to integrate its Trainium4 chips, D-Matrix is building Raptor-XPUs, and MediaTek, Marvell, Fujitsu, and Qualcomm each have integration plans in various stages.

The draft repeated the 2028-as-procurement point and the missing-specs point three times across overlapping paragraphs. The rewrite consolidates each into a single paragraph with a distinct angle: the date, the architectural dependency, the absence of numbers, the rack and government context, and the broader NVLink Fusion ecosystem. The lede now leads with the date and immediately reframes it as a commitment rather than a spec, which the draft buried in its second paragraph.