Future TechnologySK Group and NVIDIA's 500 Billion Dollar Deal Is About Memory as Much as AI Factories
Key takeaways
- SK Group and NVIDIA announced a 500 billion dollar multi-year strategic partnership covering AI factories and next-generation memory
- SK Hynix, part of SK Group, is one of only a few companies worldwide that can produce High Bandwidth Memory (HBM) at scale
- HBM stacks memory directly adjacent to GPU dies via through-silicon vias, dramatically increasing memory bandwidth
- Most LLM inference workloads are memory-bound rather than compute-bound, making HBM bandwidth the key performance constraint
The headline number from the SK Group and NVIDIA partnership announcement was 500 billion dollars, and understandably that number dominated the coverage. But buried in the deal structure is something arguably more important: next-generation memory, and why it is becoming the critical bottleneck in AI systems.
SK Group is the parent company of SK Hynix, which produces High Bandwidth Memory (HBM) chips. HBM is the memory that stacks directly alongside GPU dies, allowing data to move at extraordinary speeds without the energy penalty of going to conventional DRAM. Every serious AI accelerator, including NVIDIA's Blackwell and now Vera Rubin chips, depends on HBM. And SK Hynix is one of the very few companies in the world that can make it at scale.
Why Memory Is the Bottleneck Nobody Talks About Enough
The public conversation about AI hardware tends to focus on compute: how many FLOPS can a chip deliver, how many operations per second, how does it benchmark on standard models. But for inference workloads, which is where the vast majority of AI infrastructure spend is going, memory bandwidth is often the binding constraint rather than raw compute.
The reason is that large language models need to load enormous amounts of data into the processor's working memory continuously. If the memory cannot supply data fast enough, the compute units sit idle waiting. This is known as being memory-bound, and it is the state that most LLM inference runs in most of the time on most hardware.
HBM solves this by stacking memory dies directly on top of or adjacent to the compute die using through-silicon vias, creating a much shorter and wider data path than conventional memory architectures. HBM5, the generation that SK Hynix is ramping for next-generation systems, offers substantially higher bandwidth than HBM3e, the generation currently shipping in Blackwell systems.
What the Partnership Actually Covers
The 500 billion dollar figure covers a comprehensive multi-year partnership across two main pillars. The first is AI factory construction: SK Group will invest in and help build AI data centre infrastructure running NVIDIA systems, adding to the growing network of large-scale AI compute facilities being established across Asia and potentially elsewhere.
The second, and more technically interesting, pillar is next-generation memory collaboration. SK Hynix and NVIDIA are working together on HBM roadmap development, which means aligning memory specifications with GPU architecture requirements before either product ships. This kind of co-design relationship is relatively rare and reflects how tightly coupled the GPU and memory development cycles need to be for next-generation systems.
The Geopolitical Angle
It would be naive to discuss a deal of this scale without noting the geopolitical context. South Korea is home to Samsung and SK Hynix, the two dominant HBM producers. Taiwan is home to TSMC, which manufactures the silicon for most leading AI chips. The concentration of critical AI supply chain components in a small number of East Asian locations has been a persistent source of anxiety for Western governments.
From NVIDIA's perspective, deepening its relationship with SK Hynix is partly about securing supply, partly about getting preferential access to the next generation of memory technology, and partly about maintaining influence over how that technology develops. From SK Group's perspective, locking in NVIDIA as a committed long-term partner provides revenue visibility and technology access that competitors would struggle to match.
The Numbers Behind the Number
Five hundred billion dollars is a very large number, but it is worth contextualising. This is a multi-year partnership across multiple SK Group entities, not a single transaction. SK Group encompasses SK Hynix, SK Telecom, SK Innovation, and numerous other subsidiaries. The investment will be spread across AI infrastructure construction, research and development, and presumably some degree of equity or preferred commercial arrangements.
None of that diminishes the significance of the deal. Commitments of this scale send a clear signal about where both organisations expect AI infrastructure investment to flow over the coming decade. The memory angle in particular deserves more attention than it has received.