The AI Infrastructure Race Just Went Nuclear

July 2026 will be remembered as the month the world’s second-largest economy went all-in on AI infrastructure — and every other major power noticed.

Yesterday, South Korea’s President Lee Jae-myung hosted Jensen Huang in San Francisco alongside South Korea’s biggest conglomerates. The result: a $500 billion-plus partnership between Nvidia and SK Group spanning AI data centers, next-generation memory chips, and a 2-gigawatt facility powered by Nvidia’s Vera Rubin chips and SK Hynix’s HBM4 memory, coming online in 2027.

Let that number sink in. Two gigawatts. That’s enough power to run roughly 1.5 million homes, devoted entirely to training and running AI models.

This Isn’t Just About Korea

Look at the global picture taking shape this month:

South Korea announced a $1 trillion+ 10-year push covering semiconductors, physical AI, and data centers — anchored by Samsung and SK Hynix committing $518 billion to new fabrication plants, with the goal of 15 gigawatts of AI data center capacity by 2035.
The US has been on a chip investment spree since CHIPS Act, with Micron alone committing $250 billion through 2035.
China is reportedly letting Alibaba, ByteDance, and DeepSeek buy up to 200,000 Nvidia H200 chips — a significant carve-out from export restrictions — while simultaneously pushing its own domestic silicon.
Meta is building its first Canadian data center in Alberta and its own custom AI chip goes into production in September.

Every major power is making the same calculation: AI infrastructure is national infrastructure. Control the compute, control the intelligence.

The Memory Problem Nobody Talks About

What’s often missing from these announcements is the memory angle. GPUs get the headlines, but you can’t run them without high-bandwidth memory (HBM). SK Hynix is already Nvidia’s largest memory supplier, and this deal locks in a multi-year commitment to co-develop HBM specifically for AI training, agents, and physical AI applications.

The bottleneck isn’t just chips anymore. It’s the memory stack, the power delivery, the cooling, the land. These deals are really about securing the entire supply chain — from silicon wafer to completed data center.

The Price War Is Infrastructure’s Shadow

All of this investment is happening against a backdrop of dramatic cost deflation on the model side. Grok 4.5, GPT-5.6 (Sol/Terra/Luna), and Muse Spark 1.1 all launched within 24 hours of each other in early July. Output token costs have dropped to $4–$6 per million tokens, down from $25–$50 for legacy flagships just eighteen months ago.

This is the strategic logic: build the infrastructure cheap enough that everyone has to use it. If you control the compute layer, it matters less which model runs on top of it.

What This Means for Builders

If you’re building AI products, the implications are concrete:

Inference is getting cheap fast. The economics that made AI prohibitive for high-volume applications are dissolving. Features that were “too expensive to ship” six months ago are now viable.
Sovereign AI is real. Countries are building their own AI stacks not just for pride, but for economic and security reasons. Expect regional model variants, local data requirements, and regulatory fragmentation to increase.
The memory and power layer is where the real competition is. If you can’t get HBM, you can’t train. SK Hynix and Samsung are now as strategically important as Nvidia in the AI supply chain.
Physical AI is arriving. Nvidia, SK Group, and Doosan are explicitly targeting robotics and industrial AI alongside the cloud compute story. The AI that’s coming won’t just live in data centers — it’s heading to factories, logistics networks, and infrastructure.

The Bottom Line

We’ve moved past the “which model is best” phase into something more structural: who controls the compute, memory, and energy infrastructure of AI. That’s a geopolitical arms race with real economic consequences, and it’s accelerating.

July 2026 marked the moment major nations stopped treating AI as a technology sector and started treating it as critical national infrastructure. Everything else — the model releases, the price wars, the governance debates — flows from that.

The race is on.