SK Hynix plans to spend $38 billion building two new memory-chip plants, a major bet on future demand for the hardware that powers AI systems. In this episode, we break down why the AI race is not only about software. It is also a massive manufacturing story, where the companies building the underlying chips may hold enormous leverage. Support the Show • Amazon picks: https://amzn.to/4fUGHEm • Amazon picks: https://amzn.to/4pUHt8Q • Amazon picks: https://amzn.to/45a4MC9 • Need a website or app? https://hireawebteam.com • Support us on Patreon: https://www.patreon.com/c/StageZeroNews?view_as=patron&amp;vanity=StageZeroNews • Subscribe on Supercast: https://clubelon.supercast.com/ Disclosure: Amazon links may earn a commission from qualifying purchases.

SK Hynix Is Betting $38 Billion on the AI Chip Boom

Elon Musk Podcast

SK Hynix is spending $38 billion, about $54 trillion to build two memory chip plants in South Korea. It's just everyone focuses entirely on the processors, right? The GPUs doing the actual math for artificial intelligence. But advanced memory is the actual bottleneck here. I mean, the most powerful AI chips sit completely idle if they don't have the memory bandwidth to feed them data fast enough.

Yeah, exactly. And roughly two-thirds of that capital is earmarked for a facility in Yungen, which is dedicated to high bandwidth memory and next generation DRAM. Then the remaining third goes to a facility in Tianju focused on NAND storage chips. So looking at that split, I mean, as AI models become larger and way more demanding, is the ultimate constraint for the industry going to be compute, memory, advanced packaging, or just something else entirely?

Well, you absolutely need HBM and DRAM because they act as the immediate working memory for AI servers. They're the components handling the calculations happening in real time. Right. But the training data itself has to live somewhere permanently, which is why that Tianjun-NAND expansion matters so much. You just cannot build these models without the raw storage capacity to hold trillions of parameters.

Think of the GPU as a chef working at a counter. The high bandwidth memory is the countertop space right in front of them. And the NAND storage is the massive walk-in freezer down the hall. And if the chef is the fastest in the world but only has a tiny cutting board, their speed doesn't really matter. They spend all their time waiting for ingredients to be brought from the freezer. Exactly.

SK Hynix Is Betting $38 Billion on the AI Chip Boom

Read the full transcript

The candid version is on a podcast,
not the earnings call.

SK Hynix Is Betting $38 Billion on the AI Chip Boom

Read the full transcript

The candid version is on a podcast, not the earnings call.

The candid version is on a podcast,
not the earnings call.