Samsung & SK Hynix $590B Bet On AI Memory Needs
$590B in new chip factories, by Samsung & SK Hynix. They have 80% market share in high bandwidth memory (used in all AI servers), and prices will increase 40–50%/quarter until 2027.
Samsung and SK Hynix, the top 2 semiconductor companies in South Korea, unveiled a massive $590 billion investment in new semiconductor plants on June 29th, with support from the new South Korean administration led by President Lee Jae Myung. The plan involves constructing four new plants in South Korea’s southwest worth 800 trillion won, a state-of-the-art packaging facility worth 81 trillion won, and 30 trillion won dedicated to R&D on next-gen chips over the next 15 years.
This is an investment in AI. High bandwidth memory (HBM) is the special type of memory used in all AI servers, stacked between the GPUs and CPUs. Every H100 GPU and every custom AI ASIC needs huge amounts of high bandwidth memory to operate. As AI compute grows, so does the need for high bandwidth memory, and the ongoing AI boom has created a massive shortage of HBM.
What analysts project for memory prices
According to Jefferies equity research, memory prices will increase 40 to 50 percent in Q3 2026, followed by another 30 to 40 percent in Q4 2026, and then another 40 to 45 percent until 2027. Yes, you read that right: three years of consecutive price hikes due to AI driving up demand. We won’t see relief until about 2028, when roughly 15 to 20 percent of the new capacity comes online. Until then, expect the price of everything with memory to go up.
We’re starting to feel the impact already. Apple just increased the price of their 16” Macbook Pro by $300, bumped up the price of the 11” iPad Air from $599 to $749, and tacked on an extra $30 for the HomePod Mini. “Component shortages,” they said. While it’s true that the memory in your phone isn’t the same kind of high bandwidth memory used in AI servers, the same production lines make both. Building high bandwidth memory for AI servers means less capacity to build memory for your phone, and the AI boom is taking up a huge and ever-growing chunk of the industry’s growth projections.
Why this is a structural commitment, not a cyclical bet
This is a big deal. $590 billion is bigger than the GDP of many nations. A 15 year R&D effort at 30 trillion won isn’t a “just in case”, it’s a statement of faith that AI will continue to drive up demand for memory, and keep doing so for the foreseeable future. Samsung and SK Hynix currently hold around 80% of the market, and they’re making sure nobody else can catch up for decades.
This is also a message. When companies are sinking hundreds of billions into new factories that won’t be operational for years, they must believe the demand will still be there. The AI infrastructure story isn’t just a quarterly call, it’s a 15 year capital expenditure plan. It’s the strongest indication yet of the industry’s view on the longevity of the AI boom.
If you're pricing AI-powered products or services, build in some headroom on the cost side for the next twelve to eighteen months. AI API pricing tends to follow infrastructure costs with a delay, and with memory prices projected to keep climbing through 2027, the cost environment for AI inference is unlikely to get meaningfully cheaper in the near term. The steady price drops of 2023 and 2024 were a feature of excess capacity; that phase appears to be ending.
Source: the-decoder.com
Free course
Stop reading about AI — start building with it
The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.
Start free →▌ Related guides

Author
Evgenii Arsentev
PhD · Chief Executive Officer, digital health
Articles · Latest articles