Confirmed 9 Nov 2026 · Compute A one-minute read

NVIDIA agreed to let a Korean chip do half the work inside its own racks.

NVIDIA, Rebellions and SK Telecom said they will link Rebellions' R200 processors into NVIDIA racks over NVLink Fusion, with NVIDIA chips handling the first stage of inference and Korean chips the second. Systems are expected in 2028.

Until now

NVIDIA sells the whole rack. Rival inference chips compete from outside it.

Now

A rival chip sits inside the rack, on NVIDIA's interconnect, doing the output half.

Why it matters

The split matters more than the partnership. Generating output is the part of inference that runs constantly and burns the most power, and it is now assigned to a Korean processor on NVIDIA's own fabric. NVIDIA keeps the interconnect and the rack. It gives away the volume half of the workload.

Established

Rebellions R200 processors will connect to NVIDIA's platform via NVLink Fusion.
The plan assigns prefill to NVIDIA chips and decode to Rebellions chips.
Korea's Ministry of Science and ICT supports the collaboration; the minister is quoted.

Not established

Any performance or efficiency figure. The release publishes none.
Any money, volume or purchase commitment. None is disclosed.
That it ships. Systems are expected in 2028, 2 years out.

The five-minute readThe record, the timeline, the weight and 6 sources

The record

We read the primary. The joint release, "NVIDIA, Rebellions and SK Telecom Announce Collaboration to Advance Sovereign AI Infrastructure With NVIDIA NVLink Fusion," sits on nvidianews.nvidia.com, which dates it November 8, 2026, and went out through GlobeNewswire dated November 9, 2026 at 12:00 AM Eastern Time, datelined Seoul. That is 05:00 Universal Time on November 9, and we date the card by that instant. The 3 parties will combine NVIDIA's platform, Rebellions' Korean-designed processors and SK Telecom's data centre, model and service capability. NVIDIA supplies NVLink Fusion, which extends its platform to custom processors, and the MGX rack architecture and reference designs, which give Rebellions an established supply chain of design manufacturers and component suppliers and let both kinds of system share rack footprint, power delivery and cooling. Rebellions will connect Rebel R200 neural processing units to the platform, through NVLink scale-up networking inside a rack. The release describes the R200 as optimised for low-latency, energy-efficient inference on frontier models, including mixture-of-experts models with multitrillion parameters, and publishes no performance figure for it. The companies plan disaggregated inference: NVIDIA graphics processors for prefill, Rebellions processors for decode, to improve throughput and power efficiency. SK Telecom plans to deploy the systems in its AI data centre and connect them to its own foundation models, including A.X K2, and to public services under Korea's AI for All project. Korea's Ministry of Science and ICT supports the collaboration. R200 systems are expected to be ready for deployment in 2028. Quoted: Jensen Huang, NVIDIA founder and chief executive, "Sovereign AI gives nations the power to turn their language, culture and expertise into intelligence"; Sunghyun Park, Rebellions cofounder and chief executive, "Today's announcement marks a turning point, opening a new chapter for sovereign AI"; Jung Jai-hun, SK Telecom president and chief executive, that it validates next-generation sovereign AI infrastructure; and Bae Kyung-hoon, Korea's Deputy Prime Minister and Minister of Science and ICT, that it enables Korean-developed processors to enter the computing infrastructure market alongside NVIDIA. Not in the release: money, volumes, purchase commitments, R200 specifications, benchmark results, data centre location, capacity, or any milestone before 2028.

How it got here

Jun 2025
SK Telecom begins piloting Rebellions processors for its AI services, the first step in a relationship that now reaches inside NVIDIA's rack.
Sep 22
Rebellions processors are reported running 4 SK Telecom AI services in production. The Korean chip is serving real traffic before this announcement, not after it.
Oct 12
NVIDIA gives away its next-generation rack designs, and Meta and Google buy its networking. Edition 049. The pattern is older than this card: open the rack, keep the fabric.
Nov 3
Nvidia tells cloud providers to sell most capacity to non-Chinese firms, the Financial Times reports. Edition 071. The same week it widens its Korean position.
Nov 9
At 05:00 Universal Time the 3 companies announce the NVLink Fusion collaboration, the prefill and decode split, and a 2028 deployment date, with Korea's science minister quoted.

Who has to react

Advanced Micro Devices and every other inference challenger now outside the rack; Samsung and SK hynix, who supply the memory decode depends on; Korea's Ministry of Science and ICT, which has staked a programme on domestic processors; SK Telecom's shareholders, funding a 2028 build; and the design manufacturers NVIDIA just introduced Rebellions to.

What to watch

Would strengthen itA second non-NVIDIA inference processor is announced on NVLink Fusion with a disaggregated prefill and decode split, or SK Telecom reports decode running on R200 in production, by December 31, 2027.
Would weaken itThe collaboration lapses, the 2028 date slips past 2029, or SK Telecom's AI data centre is built on NVIDIA processors alone, by December 31, 2028.
Would strengthen itKorea's Ministry of Science and ICT discloses funding, a purchase guarantee or a domestic-processor quota attached to this collaboration, by June 30, 2027.

Why it weighs this much

Signal strength

CostMedium

CommitmentMedium

ReversibilityModerate

PersistenceYear

Cost is medium. No money is disclosed and nothing is bought, but opening NVLink Fusion to a competitor's inference chip, and handing it decode, costs NVIDIA optionality it has never sold before, and Rebellions is committing a chip generation to someone else's interconnect. Commitment is medium: a 3 party joint release on NVIDIA's own newsroom with a national minister quoted is hard to walk back politically, but it contains no contract, no volume and no money. Reversibility is moderate, because a 2028 date leaves 2 years in which any party can quietly stop. Persistence is a year rather than structural: if the decode split works, it outlives the announcement; if it does not, nothing was signed.

The words

Prefill and decode: The 2 halves of answering a prompt. Prefill reads the input in one burst. Decode writes the answer one piece at a time, and is where most serving cost sits.
NVLink Fusion: NVIDIA's high-speed link, opened so other companies' chips can join its racks. Joining means designing to NVIDIA's interconnect.
Neural processing unit: A chip built only for running models, not training them. Narrower than a graphics processor, and cheaper per answer when it fits the job.
MGX: NVIDIA's reference rack design, covering footprint, power and cooling, which partners build to instead of designing a rack themselves.

Sources

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.