Nvidia's $500 Billion Memory Grab Redraws the AI Supply Chain

The AI industry has spent two years obsessing over model parameters, training compute, and inference benchmarks. But this week, Nvidia made clear what the real bottleneck has become: memory. Not the kind your laptop needs, but High Bandwidth Memory — the ultra-fast, vertically stacked DRAM that feeds data to GPUs fast enough to keep trillion-parameter models running.

On Friday, Nvidia announced a deal with South Korea's SK Hynix that could be worth $500 billion over multiple years. It is the largest semiconductor supply agreement in history, and it signals a fundamental shift in how the AI industry thinks about strategic resources.

What Happened

The agreement, announced at an AI summit in San Francisco alongside South Korean President Lee Jae Myung, locks Nvidia into a long-term supply of HBM memory from SK Hynix — the undisputed global leader in HBM production. The deal includes the construction of large-scale data centers expected to come online in 2027, with SK Hynix affiliate SK Telecom building a cloud business on Nvidia's next-generation Vera Rubin systems.

The scale is hard to overstate. Nvidia is targeting 2 gigawatts of power capacity for these facilities — enough to run hundreds of thousands of GPUs simultaneously. For context, 2GW is roughly the output of a large nuclear reactor, or about 1.5 times the peak power draw of all of Las Vegas.

In parallel, Samsung Electronics signed a separate $200 billion memorandum of understanding with Broadcom to expand collaboration across memory and foundry technologies. And Nvidia committed an additional $1 billion investment into Naver, a Korean cloud company building 200MW of data center capacity around Nvidia GPUs.

"The expansion will include a co-develop opportunity for us on the next-generation SK Hynix AI memory, and this will help us secure a stable supply of HBM memory," said Raj Mirpuri, Nvidia's enterprise vice president, on a call with reporters.

Why Memory Became AI's Critical Path

For the first two years of the generative AI boom, the conversation focused on GPUs — who could fab the most chips, fastest. But as models grew from billions to trillions of parameters, a second bottleneck emerged: memory bandwidth.

Modern AI training runs are memory-bound, not compute-bound. A single H100 GPU can perform 1,979 teraflops of FP8 computation, but can only feed data into those compute units at 3.35 terabytes per second. For large models, the GPU spends more time waiting for data than computing on it. Every new generation of GPU — from Hopper to Blackwell to Vera Rubin — demands proportionally more HBM.

SK Hynix controls roughly 50% of the global HBM market and is the sole supplier of HBM3E, the latest generation used in Nvidia's Blackwell chips. Samsung, the only other major producer, has struggled with qualification yields. Micron, the American contender, remains a distant third.

Nvidia's move effectively locks up the majority of the world's advanced memory supply for years. Anyone building AI infrastructure — hyperscalers, startups, national governments — will either buy through Nvidia's ecosystem or face HBM scarcity.

The Geopolitical Dimension

The deal is as much about geopolitics as semiconductors. South Korean President Lee Jae Myung's presence at the San Francisco summit was no coincidence. The US and South Korea are deepening a technology alliance at the exact moment Washington is weighing restrictions on Chinese AI models and chips.

SK Hynix listed on the Nasdaq earlier this month, opening a direct channel for American capital to flow into Korean semiconductor manufacturing. The Samsung-Broadcom deal adds another layer: it ties American chip design to Korean fabrication at a scale that makes decoupling economically unthinkable.

This is a deliberate counterweight to China's own semiconductor ambitions. While the US debates banning Chinese open-weight AI models, it is simultaneously locking in the physical supply chain that makes frontier AI possible at all. The message is clear: you can build models, but you cannot build the infrastructure without US-allied memory.

What It Means for AI Builders

For the startups, research labs, and enterprise teams actually building on AI, the implications are mixed.

On one hand, secured HBM supply means Nvidia can ship more GPUs, more predictably. The global memory shortage that threatened to constrain Blackwell and Rubin production now has a release valve. Cloud providers and enterprises with Nvidia allocations will benefit from more stable pricing and availability.

On the other hand, concentrated supply creates concentrated risk. If Nvidia controls the GPU and locks up the memory, the AI stack becomes a monoculture. Alternative chipmakers — AMD, Intel, Cerebras, Groq — will compete for the remaining HBM capacity. Open-weight model builders who want to run on their own hardware may find themselves priced out of memory altogether.

There is also a pricing question. Nvidia's margins on data center GPUs already exceed 70%. When the company controls both the chip design and the critical input supply, the pricing power flows entirely in one direction. The $500 billion figure is what Nvidia pays SK Hynix; what customers pay Nvidia for the resulting systems will be a multiple of that.

The deal also signals that massive AI infrastructure is moving beyond hyperscalers. Foreign governments and conglomerates are now direct participants. South Korea is effectively co-investing in infrastructure that will serve as a national AI capability. Expect Japan, the EU, and Middle Eastern sovereign wealth funds to pursue similar arrangements — each trying to lock in compute before the capacity is fully allocated.

The Road to 2 Gigawatts

The 2GW figure is the most consequential number in this announcement. It tells us the AI industry is planning for an order-of-magnitude increase in compute deployment. Current hyperscale data centers typically operate in the 100-300MW range. A single 2GW campus — if built as one facility — would be the largest data center ever constructed by a factor of five.

SK Hynix's Nasdaq listing provides the financial vehicle: American capital markets will fund Korean manufacturing capacity, which in turn supplies American-designed chips, which power global AI workloads. The loop is closed.

For builders, the takeaway is straightforward: compute is not getting cheaper in absolute terms, but it is becoming more available. The question is whether that availability flows broadly through the ecosystem or concentrates in the hands of the largest players. Nvidia's memory grab suggests the answer, at least for now, is the latter.

The AI industry just got its equivalent of Standard Oil locking up the railroads. What happens next depends on whether anyone can build an alternative path to market.