The memory boom-and-bust cycle intensifies due to AI

What happened?

DRAM and NAND memory manufacturing is experiencing an unprecedented boom driven by artificial intelligence. The three major manufacturers — SK Hynix, Micron, and Samsung — have seen their revenues triple in the past year thanks to demand for HBM, DDR5, and NAND flash for GPU servers. According to The Register, SK Hynix and Micron have tripled their revenues, while Samsung has roughly doubled them. However, supply cannot keep pace: building new factories takes years, and prices remain high. According to IDC, no relief is expected until at least 2028.

This boom is due to the explosion in demand for AI infrastructure. Data centers training models like GPT-4 or Gemini require enormous amounts of high-bandwidth memory (HBM) and DDR5. HBM, in particular, is a critical component for NVIDIA, AMD, and Intel GPUs. The three major manufacturers have redirected their production capacity toward these high-margin products, leaving less supply for other segments like consumer electronics.

Why is it important?

This cycle affects the entire AI value chain. Startups and model developers, like OpenAI, rely on expensive infrastructure. With memory prices soaring, margins per token shrink, jeopardizing the economic viability of many projects. Historically, memory is a commodity with boom-and-bust cycles; the current boom could be the wildest yet, and the bust the deepest if AI demand doesn't materialize as expected.

For context, the memory market has experienced boom-and-bust cycles since the 1970s. In 2018-2019, oversupply drove DRAM prices to historic lows, causing multi-billion-dollar losses at Samsung and Micron. The current cycle is different because it's driven by structural AI demand, not PC or smartphone seasonality. However, volatility remains high. As one analyst cited by The Register noted: “If anticipated AI demand doesn't materialize, everyone loses, and memory manufacturers will find themselves at the bottom of a bust cycle that surpasses all previous ones.”

Additionally, market concentration is a risk factor. The three giants control over 95% of DRAM and NAND production. Any supply chain disruption, whether from natural disasters, geopolitical tensions, or manufacturing issues, can have a disproportionate impact on prices. For example, in 2021, an earthquake in Taiwan affected TSMC's production but also Micron and Nanya's memory factories, temporarily raising prices.

Consequences

• High prices until 2028: New production capacity won't be ready for at least three years, keeping prices elevated. According to IDC, additional supply from factories like SK Hynix's in Korea or Micron's in the US won't be operational until 2027-2028. Meanwhile, AI demand continues to grow at double-digit annual rates, suggesting the imbalance will persist.
• Pressure on AI startups: Infrastructure costs rise, making profitability difficult and accelerating the need for funding or consolidation. Companies like OpenAI, Anthropic, or Cohere spend hundreds of millions of dollars annually renting cloud servers, which in turn pay high memory prices. This could lead to greater concentration in the sector, where only the best-funded startups survive.
• Risk of overinvestment: If AI demand slows, memory manufacturers will face excess capacity and a devastating price drop. In June 2024, South Korean President Lee Jae-myung announced a $576 billion investment plan led by SK Hynix and Samsung to boost chip production. Micron also announced up to $3 billion in investment to strengthen the semiconductor supply chain in the US, in addition to expanding plants in Singapore, Taiwan, and Japan. If AI demand doesn't meet expectations, these investments could create an oversupply similar to 2018.
• Impact on consumers: Memory shortages raise prices for consumer electronics like smartphones and PCs. The Register notes that “you can't even buy a budget smartphone today” without the price being affected. PC and phone manufacturers have had to raise prices or reduce specifications to maintain margins. For example, mid-range smartphones now often have 6GB of RAM instead of 8GB, and budget laptops frequently come with 8GB instead of 16GB.

What should readers know?

The memory market is at a critical point. Investment in new factories is massive, but the time lag between investment and production is the biggest challenge. Building a DRAM or NAND flash wafer fab is not trivial: it requires financing, site selection, permits, installation of support systems like ultrapure water filtration and air handling, and then installation and validation of lithography and wafer transport equipment costing hundreds of millions of dollars. Even after everything is ready, it can take months for yields to become acceptable. This process often takes years even without delays. Therefore, any factory started today will take at least three years to become operational, and longer to reach full production.

Meanwhile, the AI industry must prepare for high costs for several years. The key will be whether AI demand holds or the cycle reverses abruptly. If AI adoption slows due to regulatory, safety reasons, or simply because use cases don't generate expected returns, memory manufacturers will face excess capacity and free-falling prices. Conversely, if demand continues to grow, prices could remain high even after 2028, as AI integrates into more sectors.

As one analyst cited by The Register said: 'If anticipated AI demand doesn't materialize, everyone loses, and memory manufacturers will find themselves at the bottom of a bust cycle that surpasses all previous ones.'

In summary, the memory boom is a short-term blessing for manufacturers but a double-edged sword for the AI ecosystem. Investors and companies must closely monitor demand trends and capacity investments to anticipate the next turn of the cycle.