Samsung, SK Hynix bet $2 trillion on AI chips. Is it too much?

Key Takeaways

• Samsung and SK Hynix committed 3,200 trillion won ($2.07 trillion) to expand South Korean chip production capacity
• Most new fab capacity won't come online until well into the next decade, creating timing risk if AI spending cools
• Both companies retain flexibility to adjust spending if oversupply signals emerge, drawing on lessons from past boom-and-bust cycles

Samsung Electronics and SK Hynix just made the largest semiconductor bet in history: 3,200 trillion won ($2.07 trillion) to double South Korea's memory chip production capacity within five years. President Lee Jae Myung responded with a deep bow. Analysts responded with a warning: chip plants take years to build, and much of this capacity won't arrive until the 2030s. If AI spending slows before then, both companies face a painful oversupply scenario they've lived through before.

What are Samsung and SK Hynix actually building?

The investment breaks into two pieces. First, an 800 trillion won chip cluster in southwestern South Korea. Second, accelerated construction at the existing Yongin semiconductor complex, where both companies will shorten the typical 7 to 12 year fab timeline by securing faster government approvals.

Samsung's share totals 2,100 trillion won through 2040. The company noted spending could be adjusted for market conditions. That caveat matters. In 2024, Samsung halted construction of its P5 chip plant in Pyeongtaek for nearly two years during a market downturn before resuming work late last year.

The strategic target is high-bandwidth memory, the specialized chips that power Nvidia's AI accelerators and sit at the center of the current shortage. SK Hynix dominates this market, supplying Nvidia directly. Samsung has struggled to catch up after quality issues delayed its HBM products.

Why the sudden shift from caution to expansion?

Both companies have spent decades learning hard lessons about overbuilding. SK Hynix nearly went bankrupt in 2001. Both firms posted significant losses in 2023 during the last memory downturn. That history bred conservatism in capacity planning.

Two months ago, SK Hynix Chairman Chey Tae-won was openly skeptical when lawmakers asked about building fabs in the southwest. "I'm not sure semiconductors are necessarily the field you have to go into," he told them.

Neither company explained what changed. But the timing coincides with ruling party lawmakers pushing to bring industrial investment to the southwest, their political base. The government promised to accelerate approvals in exchange for the commitment.

The financial incentive is real too. Memory chip prices nearly doubled in Q1 2025. AI hyperscalers and electronics makers including Apple are pressing for more supply. When demand is strong and margins are high, the pressure to build is enormous.

The timing problem analysts are flagging

Morningstar analyst Jing Jie Yu put the concern bluntly: "We see memory pricing remaining a function of demand and supply, and accelerating capex over the next decade further increases the risk of an oversupply longer term."

The memory chip boom depends on AI hyperscalers continuing to expand at current rates. That's a strong assumption to carry into the 2030s. Lee Jong-ho, a professor at Seoul National University's Department of Electrical and Computer Engineering, said the investments appeared rushed.

“It is the kind of investment that could determine a company's future. No one knows what the situation will look like three years from now. We need to respond quickly while demand is strong, but after that, demand is uncertain and decisions should be made cautiously.”

— Lee Jong-ho, Seoul National University

CW Chung at Nomura offered a different read. Building in the southwest could hedge uncertainty around the Yongin cluster. If permitting or infrastructure issues delay Yongin, capacity elsewhere provides insurance.

How past cycles shape the current bet

CLSA senior analyst Sanjeev Rana expects the companies to apply lessons from prior downturns. "A downturn in the memory industry is clearly a risk to the plan," he said. But "memory producers retain the flexibility to adjust their investment pace if signs of excess capacity emerge."

Samsung's language reinforces this. The 2,100 trillion won figure runs through 2040, with explicit room to scale back. The company already demonstrated that willingness by pausing Pyeongtaek construction during the 2023 slump.

The question is whether political commitments constrain that flexibility. When a president publicly celebrates your investment and a region's economy depends on your construction schedule, pulling back gets harder.

What this means for AI infrastructure buyers

If the buildout proceeds on schedule, HBM supply will loosen considerably by the late 2020s. That could ease the chip constraints that have limited AI infrastructure deployment and driven up costs for training and inference workloads.

In the near term, shortages persist. Apple and AI hyperscalers are competing for limited memory capacity. Companies planning major AI infrastructure purchases should expect tight supply and elevated pricing through at least 2026.

Frequently Asked Questions

How much are Samsung and SK Hynix investing in AI chips?

The two companies pledged 3,200 trillion won ($2.07 trillion) total, with Samsung committing 2,100 trillion won through 2040. This includes a new 800 trillion won chip cluster in southwestern South Korea.

When will the new chip capacity come online?

Most new capacity won't be available until well into the next decade. Typical fab construction takes 7 to 12 years, though the companies plan to accelerate timelines with faster government approvals.

What is high-bandwidth memory and why does it matter for AI?

High-bandwidth memory (HBM) is specialized memory that provides the data throughput AI processors need for training and inference. SK Hynix supplies HBM to Nvidia for its AI accelerators.

Could Samsung and SK Hynix scale back this investment?

Yes. Samsung explicitly stated spending could be adjusted for market conditions. Both companies have paused construction before during downturns, including Samsung's P5 plant halt in 2024.

Why are analysts concerned about oversupply?

Chip plants take years to build, and AI demand could slow before capacity arrives. The memory industry has a history of boom-and-bust cycles caused by overbuilding during demand peaks.

Source: Tech-Economic Times / ET