NVIDIA and SK Group sign a 500 billion dollar plus AI partnership
NVIDIA and South Korea's SK Group signed letters of intent for a comprehensive AI infrastructure partnership described as a 500 billion dollar plus initiative. The agreement has two legs: SK Telecom builds a 2-gigawatt-scale AI cloud, and SK hynix becomes a long-term partner on next-generation AI memory.
Factory and memory in the same agreement
According to NVIDIA's announcement dated July 24, 2026, the two sides signed letters of intent covering everything from AI factory construction to AI memory supply, in what the release describes as a 500 billion dollar plus initiative. SK Telecom will build a 2-gigawatt-scale AI cloud on NVIDIA's DSX platform, deploying Vera Rubin accelerated computing powered by SK hynix HBM4, with the first AI factory planned to come online in 2027.
The memory half is the quiet, decisive one
The second leg is a long-term collaboration between NVIDIA and SK hynix to codevelop and optimize next-generation AI memory, including HBM, aimed at large language model training and physical AI applications. The release also frames regional demand as a driver: the goal is large-scale infrastructure for sovereign, physical, agentic and enterprise AI services across the Asia-Pacific region.
SK Group Chairman Chey Tae-won says that by combining SK hynix's AI memory with SK Telecom's infrastructure capability, SK will work with NVIDIA to build a world-class AI factory. NVIDIA CEO Jensen Huang says the companies are building a new generation of AI factories to power Korea's next wave of growth. Infrastructure cooperation between the two is not new: a plan for a gigawatt-scale sovereign AI cloud with SK Telecom, with its first facility due in 2027, had already been announced; this agreement widens that frame to include memory supply.
The bottleneck in AI is the supply chain, not the model
What recent large announcements have in common is no longer models but capacity: power measured in gigawatts, accelerator counts and memory supply. That this deal includes the memory leg is precisely the signal; producing accelerators is not enough when high-bandwidth memory has to be locked in years ahead. Capacity plans pointing at dates like 2027 also show the industry does not consider the bottleneck solved in the short term.
The second point is geography. The phrase sovereign AI describes countries wanting compute capacity to sit inside their own borders. As that trend grows, which capacity exists in which country, where data is processed and where latency comes from all become part of commercial decisions.
What it means for businesses in Türkiye
An investment on this scale does not change a Turkish business's bill today; the effect is indirect and spread over time. Three practical conclusions still follow. First, because capacity plans run long, treat movement in model and cloud costs as a structural variable rather than a passing promotion: do not pin budget assumptions to a single price level.
Second, provider dependency. As capacity concentrates around certain manufacturers and regions, architectures locked to one provider carry more risk. Treating the model provider as a replaceable component leaves you with a manageable system without having to track news like this. Third, data residency: as the sovereign cloud trend strengthens, the question of which country processes your data will come up more often, both for data protection compliance and for customer trust. The work to do now is to turn which workload runs where into a written architectural decision.
The UNALSOFT take
When we pass news of this scale to clients we reduce it to one question: what does this change in our architecture? Most of the time the answer is nothing in the short term, and cost plus dependency in the long term. In our Agentic AI builds the model provider is a component chosen in configuration, so movement on capacity and pricing means changing a setting rather than rewriting the system.
Capacity news comes and goes, architecture stays.
Let's review provider dependency in how you use AI.