Half a trillion dollars.

Let that sink in. That's roughly the GDP of Vietnam. More than three times Nvidia's full-year 2025 revenue. Half of all global AI startup funding in 2025 combined. On July 25, at an AI summit in San Francisco, Nvidia and South Korea's SK Group jointly announced a comprehensive AI partnership valued at over $500 billion — top to bottom, from memory chips to GPUs to data centers, two giants are welding the entire AI infrastructure stack together.

This isn't an ordinary business deal. It's a landmark inflection point in the AI compute arms race: the competitive dimension has shifted from “whose GPU is faster” to “who can lock down the entire supply chain.” When a chip company and a national-level conglomerate bind together at $500 billion depth, the rules of AI infrastructure have fundamentally changed.

Where Does the $500B Go? Two Layers of Binding, Three Pillars

Jensen Huang personally explained the composition of the $500 billion at the launch event. This isn't a simple procurement contract — it's a mutually binding, full-stack strategic alliance.

The first layer of binding is memory chips. Nvidia is establishing a long-term partnership with SK Hynix to lock in supply of next-generation HBM4 high-bandwidth memory. SK Hynix will ensure sustained HBM4 supply for Nvidia's next-gen GPUs, and the two will co-develop specialized high-bandwidth memory for AI training, AI agents, and physical AI. Put simply: Nvidia is pre-buying all of SK Hynix's best memory capacity for years to come. Everyone else? Get in line — if there's any left.

How aggressive is this move? HBM is the chokepoint of AI training — no matter how powerful your GPU, if memory can't feed it data fast enough, it's wasted. The HBM market is essentially controlled by three players: SK Hynix, Samsung, and Micron, with SK Hynix leading on HBM3 and HBM4 technology. By deeply binding with SK Hynix, Nvidia has essentially built a private highway for itself at the critical memory bottleneck. Competitors get the country roads.

The second layer is data centers. SK Telecom will build a 2-gigawatt (GW) AI data center in South Korea using Nvidia Vera Rubin chips and SK Hynix HBM4 memory. 2GW — that's roughly the daily electricity needs of 1.5 million households, roughly the output of two large nuclear power plants. The first facility is expected to go online in 2027.

Notice the mutuality here: the $500 billion includes both Nvidia's memory chip purchases from SK Hynix and SK Group's supercomputer and GPU purchases from Nvidia. You buy my chips, I buy your compute — each side is the other's biggest customer and supplier, bound so tightly neither can walk away.

The alliance's three pillars are clear: SK Hynix handles memory, Nvidia handles compute, SK Telecom handles data centers. Memory — chips — data centers, the three most critical links in AI infrastructure, strung together on a $500 billion rope.

Why South Korea? Why Now?

Huang said at the event: “South Korea is entering its golden era. The country's semiconductor industry and industrial sector are both thriving, with the complete capability to build AI infrastructure for the world.”

This isn't just flattery. South Korea's position in the AI infrastructure era is a bit like Saudi Arabia in the 20th century — you might not like it, but you can't live without its resources.

South Korea's core resource is memory chips. SK Hynix and Samsung together control over 80% of the global HBM market. In an era where AI training is nearly insatiable for VRAM bandwidth, HBM capacity and technology gaps have become a scarcer bottleneck than GPUs themselves. You can buy GPUs if you have money, but you can't necessarily get the HBM that goes with them — and a GPU without HBM is just expensive dead weight.

The timing of the $500 billion announcement also has an important context: AMD is storming Nvidia's moat with 2nm GPUs. Multiple outlets including InfoQ report that AMD's latest 2nm GPU has garnered collective backing from OpenAI, Meta, Microsoft and other giants. The AI chip market is evolving from “Nvidia dominates alone” to a “two-horse race.”

Nvidia's counter-strategy is clear: you challenge me at the chip level, I'll escalate the competition to the supply chain level. AMD can make a 2nm GPU, but can it lock down SK Hynix's HBM4 capacity? Can it build 2GW data centers with SK Telecom? Can it pull together a full-stack ecosystem from memory to data center? Not in the short term. With $500 billion, Nvidia is telling the market: AI compute competition isn't about whose chip is faster — it's about whose ecosystem is deeper and whose bindings are tighter.

There's also a geopolitical factor that can't be ignored. U.S. export controls on AI chips to China keep tightening, and Nvidia's high-end chips can't enter the Chinese market. This forces Nvidia to find new growth drivers and manufacturing bases. South Korea is a U.S. ally, has a complete semiconductor supply chain, and has the power and land resources for large-scale data center construction — practically the perfect partner.

The Compute Race Triple Jump: From Cards to Plants to Alliances

Looking back at the evolution of AI compute competition, you can see a clear escalation path — each level expands the dimension of competition and raises the barriers.

Phase one was the “buying cards” era (2023–2024). Back then, AI companies competed on who could buy more H100s. OpenAI bought tens of thousands, Anthropic bought tens of thousands — everyone was competing on procurement capacity and relationships with Nvidia. Compute = GPU count. Simple and brutal.

Phase two was the “locking plants” era (2025–first half of 2026). AI companies realized buying chips wasn't enough — data centers were scarce, power was scarce, memory was scarce, network bandwidth was scarce. So Anthropic signed a $19 billion, 20-year data center deal. Google committed 5GW of compute to Anthropic. Microsoft launched Frontier Company dedicated to AI deployment. The compute race escalated from “buying chips” to “building infrastructure.”

Phase three is now — the “alliance era”. Nvidia and SK Group's $500B deal marks a turning point: compute competition is no longer between individual companies, but between entire supply chain alliances. On one side, the Korea-U.S. alliance of Nvidia + SK Hynix + SK Telecom. On the other, the AMD + Samsung + cloud vendors combination. And then there's China's self-developed ecosystem as a third pole. The future won't be decided by single-chip performance, but by the coordination efficiency, cost advantage, and supply security of the entire chain.

This trend is bad news for smaller AI companies. Before, with a few hundred million in funding you could buy thousands of GPUs and train a large model. Now, without tens of billions, without full-chain partners, without national-level support, you can't even get a ticket to the game. The number of players in frontier LLMs will keep shrinking and concentrating.

China's Position: Shortcomings and Breakouts

Where does China's AI industry stand in this full-chain binding compute race?

Objectively speaking, the gaps are obvious. HBM is basically import-dependent. High-end GPUs are blocked. Advanced-node chip manufacturing capacity is limited. A QbitAI report titled “AI's Most Embarrassing Shortcoming, Chinese Academy of Sciences Steps In” addresses exactly these hard bottom-layer problems in memory and compute chips.

But there are positive signals too. Jensen Huang has publicly defended Chinese open-source AI models, and nearly 200 Silicon Valley companies have signed a letter to the White House urging against banning U.S. companies from calling Chinese open-source models like Kimi K3 and Qwen 3.8. What does this mean? It means the quality and ecosystem of Chinese AI models have reached a level that can't be ignored — you can restrict hardware exports, but you can't restrict the flow of open-source code, and you can't restrict developers from voting with their feet.

China's AI industry is forging a unique path: hardware independence, software openness. The hard bones of compute chips, HBM, and advanced processes need to be gnawed through, but at the same time, China's open-source LLMs, AI application ecosystem, and vertical industry deployment are already at the global forefront. It's like the “Qualcomm chips + Android system” dynamic in the mobile era — you control high-end chips, but I control the application ecosystem and user scale, and ultimately it's not clear who has more leverage.

Another angle worth watching is power and land. The ultimate constraint of AI data centers isn't chips — it's electricity. A 2GW data center needs enormous power supply and land resources. China's Guizhou, Inner Mongolia, Ningxia, and Gansu have abundant clean energy and cheap electricity, giving them unique advantages in compute infrastructure construction. SK Group is building 2GW of data centers in South Korea; China is building data centers of equal or even larger scale. The endgame of compute competition might not depend on who has the best chip, but on who has the cheapest electricity, the most land, and the fastest construction speed.

Endgame: AI Infrastructure Is Going National

Putting the $500B deal in the bigger picture, a clear trend is emerging: AI infrastructure is moving from enterprise-scale to national-scale.

The Nvidia-SK partnership is, on the surface, a commercial alliance between two companies. In substance, it's a national-level binding of U.S. AI computing power with Korean semiconductor manufacturing. Korean memory chips + American compute architecture + Korean land and power = the world's most powerful AI infrastructure alliance. The opponent of this alliance isn't a single company — it's an entire country's AI industrial system.

What does this mean for the industry?

First, the “national team” era of compute is arriving. Frontier AI compute of the future won't be in the hands of startups, or even single tech giants — it will be in the hands of “nation + consortium + tech giant” alliances. That's how it is in Korea, in the U.S., and inevitably in China too.

Second, supply chain security matters more than performance. In a $500B partnership, performance is just the baseline. What really determines victory is supply security — can you reliably get HBM? Can you sustain access to power? Will you get choked by export controls? The resilience and security of the supply chain matter far more than whether a single chip is 10% faster or slower.

Third, the geography of AI will be redefined. In the past internet era, tech hubs were Silicon Valley, Beijing, Shenzhen — wherever there were people, ideas, and capital, that's where the center was. The AI infrastructure era is different. The center shifts to places with power, land, and chip manufacturing capacity. South Korea, Taiwan, Guizhou, Inner Mongolia — places that don't sound so “high-tech”will instead become strategic strongholds of the AI era.

Five hundred billion dollars isn't just a number. It's the price of admission to an era. When AI infrastructure competition rises to the level of national alliances, every player must reassess their position. You can choose not to play — but the price might be complete irrelevance in the next era.

See you tomorrow.