On November 8, at the eve of Nvidia AI Day Seoul 2026, Nvidia, Rebellions and SK Telecom announced a collaboration that sounds like diplomacy and reads, in the technical details, like a very specific engineering bet. Rebellions, a Korean AI chip designer, will attach its next-generation Rebel R200 neural processing units to Nvidia's NVLink interconnect using NVLink Fusion, package them into Nvidia MGX rack designs, and run them inside SK Telecom's AI data center. Deployment target: 2028.
The line most coverage will pull is Jensen Huang's: "Sovereign AI gives nations the power to turn their language, culture and expertise into intelligence." The line that matters more for what your inference bill looks like in 2028 is buried a few paragraphs later. The companies plan disaggregated inference that assigns "NVIDIA GPUs for prefill and Rebellions NPUs for decode." That is not a vague co-existence plan. It is a claim that two different chip architectures should each run the phase of a large language model request they are physically better suited for, in the same rack, talking over the same interconnect.
Context: What Was Actually Signed
The agreement is a memorandum of understanding, supported by Korea's Ministry of Science and ICT, with three concrete technical commitments per Nvidia's press release:
- Rebellions connects R200 NPUs to NVLink scale-up networking through NVLink Fusion, the program Nvidia first unveiled at Computex in May 2025 that lets third-party silicon join NVLink-based systems.
- Systems are built to the Nvidia MGX rack architecture and reference designs, which means shared rack footprint, power delivery and cooling with Nvidia GPU systems, plus access to MGX's supply chain of ODMs and component vendors.
- SK Telecom validates and deploys the mixed GPU-NPU racks in its AI data center, wiring them to its own models (including A.X K2) and public services under Korea's "AI for All" project.
Korea's Seoul Economic Daily reports the memory angle the release understates: Rebellions' current R100 NPU carries four HBM3E stacks for 144 GB capacity and 4.8 TB/s of bandwidth, with Samsung supplying that HBM3E, and the industry expects the R200 to carry HBM3E or the next-generation HBM4, with SK hynix in play as an additional supplier. MoneyToday adds that the ministry ran an AI Startup Partnership Forum the same day with 22 Korean companies, including Upstage, TwelveLabs, FuriosaAI and DeepX, and that this follows earlier work connecting Korean NPUs to AMD platforms.
The Engineering Core: Prefill and Decode Are Different Workloads
Why would a GPU company hand its interconnect to an inference specialist? Because the two phases of LLM serving stress silicon in opposite ways, and running them on one chip type wastes money in a way every inference operator recognizes.
| Phase | What happens | Bottleneck | Hardware preference |
|---|---|---|---|
| Prefill | The whole prompt is processed in parallel to build the KV cache | Compute (matrix multiply throughput, FLOPs) | GPUs: strong tensor cores, high clocks |
| Decode | One token at a time, each step re-reading weights and KV cache | Memory bandwidth, bytes moved per token | Purpose-built NPUs: max bandwidth per dollar, less raw FLOPs needed |
This is why Nvidia's own open-serving stack, Dynamo, already treats the phases separately, and why the press release's plan to put prefill on GPUs and decode on R200s is a coherent division rather than a marketing arrangement. Decode is bandwidth-bound: for a model with $W$ bytes of active weights per token, each generated token requires moving roughly $W$ bytes plus the KV cache read, regardless of how many FLOPs you could theoretically throw at it. An NPU that trades away training-oriented features and peak FLOPs for cheap HBM bandwidth can generate tokens at a far better cost per token than a flagship GPU, as long as the KV cache produced during prefill can be shipped to it fast enough.
That "as long as" is exactly what NVLink Fusion is for. The transfer of KV cache state between prefill and decode workers is the tax on disaggregation, and it scales with prompt length and context. A PCIe-class link makes that tax expensive enough to erase the NPU's savings at long contexts; NVLink-class scale-up bandwidth keeps the penalty small inside a rack. Nvidia is effectively saying: if custom silicon wants efficient inference economics, it should pay the interconnect toll in our currency, sit in our racks, and draw power and cooling to our spec.
Rebellions' CEO framed the same bet in economic terms at the signing, per Seoul Economic Daily: "What we ultimately want is to become an exporter of tokens, the units in which AI generates text," he said, positioning the R200 as a step toward "a standard for producing cost-effective tokens."
NVLink Fusion Is a Moat Widened by Digging It Deeper
The strategic read: NVLink Fusion looks like openness, but every participant it attracts makes the rack standard stickier. The pattern mirrors what Nvidia already runs on the software side. At the same Seoul event, Nvidia released Dynamo-optimized recipes for Korean open models (A.X K2, K-EXAONE 2.0, Solar Open 2, Motif-3), claiming up to 2x out-of-the-box performance on B200, and applied-research VP Jonathan Cohen used the keynote to argue token usage is "doubling every month," with an internal model-routing tool called Switchyard reportedly holding an 80% task success rate at roughly one third the cost. Nvidia is expanding from selling accelerators into owning routing, deployment and now the interconnect terms on which rivals' accelerators may join.
For custom-silicon buyers, the calculus has two columns. Join NVLink Fusion and you inherit MGX power, cooling and ODM supply chains, validated rack designs, and a fast path to production, at the price of designing to Nvidia's interface and reinforcing its system-level lock-in. Stay outside and you keep architectural independence but must rebuild the ecosystem MGX provides, which is the chronic failure mode of inference ASICs we have covered before: Microsoft's Maia 300 and Meta's MTIA 450 both chase cost per token behind their own firewalls, and OpenAI's Jalapeño is benchmarked against Blackwell but deployed only where OpenAI runs it. Rebellions is the first Korean chipmaker to try the third path: win by being the best decode chip inside Nvidia's house rather than a castle of its own.
The numbers worth watching, all from opened sources:
| Item | Value | Source status |
|---|---|---|
| Rebellions R100 memory | 4x HBM3E, 144 GB, 4.8 TB/s | Reported (Seoul Economic Daily, citing sources) |
| R100 HBM supplier | Samsung Electronics | Reported, unnamed sources |
| R200 memory choice | HBM3E or HBM4, supplier open | Industry expectation, not disclosed |
| Phase split | GPUs prefill, NPUs decode | Nvidia press release (plan, not benchmark) |
| Deployment target | 2028 | Nvidia press release |
| Switchyard cost claim | ~1/3 cost at 80% task success | Nvidia internal eval, vendor claim |
What This Means: Token Economics, Not Chip Economics
Three consequences follow if the 2028 timeline holds.
First, inference pricing may decouple from GPU pricing. A data center that buys decode capacity from bandwidth-optimized NPUs and prefill capacity from flagship GPUs sets its blended cost per generated token by a ratio operators control (the prefill-to-decode fleet mix), not by Nvidia's GPU list price. This is the same pressure that pushed DeepSeek to publish $3.80 GPU-hour efficiency targets: once serving efficiency is measured per token, the cheapest token source wins regardless of brand. Korea's government explicitly wants this outcome; a nation that can serve its own models cheaply is, in the phrase Rebellions' CEO chose, an exporter of tokens.
Second, watch what the KV cache transfer tax actually measures. The whole GPU-prefill, NPU-decode plan is falsifiable: publish prompt-length sweeps showing time-per-output-token under disaggregation versus GPU-only serving, and the win or loss is visible. No such numbers exist yet, and both vendors have incentives to keep them internal until the product ships. Until then, "improved throughput and power efficiency" is a vendor intention, not a result.
Third, MGX becoming a neutral rack standard changes procurement conversations well beyond Korea. If third-party NPUs can slot into MGX racks with validated power and cooling budgets, operators stop asking "can we physically deploy non-Nvidia accelerators" and start asking "what is the cost per token delta." That is a healthier question for buyers, and Nvidia is choosing to answer it rather than let AMD or custom-ASIC vendors frame it, because being the host of the neutral rack is its own kind of market share.
Outlook
The MOU is validation work now and revenue in 2028: SK Telecom validates on its infrastructure, Rebellions tapes out R200 systems on MGX reference designs, and the ministry has promised support "across a range of areas, including physical AI and securing GPU supply," per Deputy Prime Minister Bae Kyung-hoon. Two signals will tell you whether this becomes a template or a press release. The first is whether a second NPU maker outside Korea adopts NVLink Fusion on the same MGX terms, which turns a national program into an industry standard. The second is any published disaggregation benchmark at long context, because the KV transfer tax is the number that decides whether a mixed rack beats a pure one.
For builders, the practical takeaway is unglamorous and useful: start separating your prefill and decode capacity in your own metrics today. Track time-to-first-token and time-per-output-token as independent costs with different hardware answers. When heterogeneous racks arrive in 2028, the teams that already know their phase-level numbers will be able to price the offer instead of taking a vendor's word for it.