NVIDIA announced on November 8, 2026 a collaboration with Rebellions, a Korean inference-chip company, and SK Telecom, supported by Korea's Ministry of Science and ICT. Rebellions will use NVLink Fusion to connect its next-generation Rebel R200 neural processing units to NVIDIA's platform, and build racks on NVIDIA's MGX architecture. SK Telecom plans to deploy them in its AI data centre for its own models, including A.X K2, and services under Korea's "AI for All" project.
Why split one request across two kinds of chip
The technical core is one sentence in the release: the companies plan "using NVIDIA GPUs for prefill and Rebellions NPUs for decode to improve inference throughput and power efficiency."
The two phases of serving a language model stress hardware differently. Prefill reads the whole prompt at once and is mostly dense matrix arithmetic, which GPUs do well. Decode produces one token at a time per request, rereading the model weights and the growing key-value cache each step, so it is usually limited by memory bandwidth rather than arithmetic. Disaggregated serving already places the two phases on separate GPU pools. This plan goes further and puts them on different silicon from different vendors, with the key-value cache moving between them over NVLink.
The appeal is that a decode chip can trade arithmetic for memory bandwidth and lower power. The cost is engineering: the cache must cross the interconnect fast enough that the handoff does not erase the gain, and two software stacks must agree on its layout. Rebellions says it targets "multitrillion-parameter mixture-of-experts models", where that cache traffic is largest.
What the announcement does not say
- No numbers. The release gives no throughput, latency, power or cost figures, for the R200 or the combined rack.
- A long timeline. It says R200 systems "are expected to be ready for deployment in 2028". When we searched on November 10 we found no independent coverage of the R200 itself.
- Who controls the stack. NVLink Fusion lets other companies' chips join NVIDIA's scale-up network. The sovereign hardware still sits inside an NVIDIA rack design and interconnect. Korea's minister also said cooperation would cover "securing GPU supply".
What remains uncertain. We read NVIDIA's release and earlier reporting on Rebellions' chip design, and could not independently confirm the R200's specifications. Whether a cross-vendor handoff of the cache beats an all-GPU split at equal power is exactly the measurement that is missing.
Analysis: Prefill and decode are separate costs, and it pays to measure them separately long before 2028. If Geetha's team at Belgavi's Book Store served a long-context help-desk model, logging time to first token apart from tokens per second would show which phase dominates the bill. A workload of long return-policy prompts and short answers is prefill-heavy, while a chatty assistant is decode-heavy, and the right hardware differs. Keep the serving layer able to place the two phases separately, because that is where the industry is heading.
Sources and evidence
NVIDIA, Rebellions and SK Telecom plan racks where NVIDIA GPUs handle prefill and Rebellions R200 NPUs handle decode, linked by NVLink Fusion, with systems expected in 2028. The release gives no performance, power or cost figures. Measure prefill and decode separately on your own traffic now, since they favour different hardware.