Samsung, SK hynix, Micron bet $129B on AI memory shortage

Key Takeaways

• Samsung, SK hynix, and Micron will invest a combined $129 billion in 2024 to expand memory production
• Global memory market forecast to reach $803.9 billion, up 250% from current levels
• Industry analysts expect HBM and DRAM supplies to remain tight for years despite massive investments

The three companies that control the world's memory chip supply are betting $129 billion this year that AI demand will keep outpacing their ability to produce. Samsung Electronics, SK hynix, and Micron Technology are racing to build new fabrication plants, expand packaging capacity, and lock in equipment before competitors can secure the same.

The urgency comes from a simple math problem. According to World Semiconductor Trade Statistics, the global semiconductor market will grow 90% this year to $1.51 trillion. The memory segment alone, which these three companies dominate, will surge 250% to $803.9 billion. That growth rate signals demand far exceeding current production capacity.

Where is the $129 billion going?

Samsung is spreading its bets across South Korea. The company is building out memory manufacturing capacity through new projects in Pyeongtaek and Yongin in Gyeonggi Province, plus an integrated semiconductor hub in Gwangju-South Jeolla. These facilities will produce both high-bandwidth memory (HBM) for AI servers and conventional DRAM.

SK hynix, which currently holds an estimated 85% market share in HBM chips and supplies NVIDIA's AI accelerators, raised $26.5 billion through its recent Nasdaq listing. The company plans to use that capital to increase wafer production capacity and strengthen advanced chip packaging capabilities. Packaging has become a bottleneck because HBM requires stacking multiple memory chips vertically, a process more complex than traditional chip manufacturing.

“We announced that we would double our production capacity within five years, but every customer says that will not be enough.”

— Chey Tae-won, SK Group Chairman

Micron is taking a different geographic approach. The company announced plans to invest more than $250 billion in the United States by 2035. That includes new memory fabrication plants in New York and Idaho, plus an expansion of its existing Virginia facility. The investment aligns with US government incentives to bring semiconductor manufacturing back to American soil.

Why can't supply catch up with AI demand?

Building a memory fab takes three to four years from groundbreaking to first production. Equipment lead times add another layer of delay. ASML, which makes the lithography machines essential for advanced chip manufacturing, has a backlog stretching years into the future.

Meanwhile, demand keeps accelerating. Every major tech company is building AI infrastructure. NVIDIA's AI chips each require six to eight HBM stacks. Meta, Microsoft, Google, and Amazon are all ordering as many AI servers as they can get. OpenAI and Anthropic need compute capacity for training ever-larger models.

Park Seung-young, head of portfolio strategy at Hanwha Investment & Securities, framed the situation bluntly: "Supply matters more than demand in commodity memory. The key question is how fast capacity can grow."

Will the AI boom last long enough to justify these investments?

Some investors worry that AI spending may be peaking. The concern is that hyperscalers have over-ordered, and demand will normalize before these new fabs come online. That would leave memory makers with expensive overcapacity.

Industry executives disagree. According to the Korea Herald's reporting, analysts and executives expect supplies of high-bandwidth memory and conventional DRAM to remain tight for years. The reasoning: AI infrastructure buildout is still in early stages. Enterprise adoption of AI workloads is just beginning. And the models themselves keep getting larger, requiring more memory per training run and per inference query.

The bet these companies are making is that they will sell every chip they can produce for the foreseeable future. Their stock performance over the next five years will depend less on short-term pricing fluctuations than on whether they can add capacity faster than competitors.

What does this mean for AI infrastructure buyers?

If you're a company trying to build AI infrastructure, the supply picture won't improve dramatically in 2024 or 2025. Even with $129 billion in combined investment, new capacity takes years to materialize. Companies that need HBM-equipped servers will continue facing long lead times and premium pricing.

For cloud customers, this translates to continued high prices for AI compute instances. AWS, Azure, and Google Cloud are all constrained by the same memory shortage. Their pricing reflects the underlying scarcity of the chips they need to expand AI capacity.

Frequently Asked Questions

What is HBM and why does AI need it?

High Bandwidth Memory stacks memory chips vertically, enabling data transfer rates far exceeding traditional DRAM. AI accelerators like NVIDIA's H100 require HBM to feed data to their processors fast enough for training large language models.

Which company leads in HBM production?

SK hynix holds an estimated 85% market share in HBM chips and is NVIDIA's primary supplier. Samsung and Micron are investing heavily to close the gap.

When will AI memory chip shortages ease?

Industry analysts expect tight supply to continue for years. New fabrication plants take three to four years to build and ramp to full production.

How much will the global memory market grow?

World Semiconductor Trade Statistics forecasts the memory market will reach $803.9 billion, representing 250% growth from current levels.

Source: Tech-Economic Times / ET