China's AI Boom Is a Silent Power Play for Samsung and SK Hynix

The Supply Squeeze No One Is Talking About

Global datacenter investment is about to double.

According to Delo Group, cumulative worldwide datacenter capex will surpass $3 trillion by 2030 — nearly twice the $1.7 trillion forecast just last January. Restructuring Energy puts China’s share of that expansion at roughly 60 gigawatts of capacity by decade’s end, up from 32GW at the end of last year. The United States built the first wave. China is building the second, and it is arriving faster than most forecasts accounted for.

This is not incremental demand. It is structural. And it changes the bargaining dynamics in a way that favors Samsung Electronics and SK Hynix — even though they are largely blocked from selling directly to the very buyers creating the shortage.

China’s Cloud Giants Are Spending Like It Is 2021 Again

Alibaba has stated its cloud and AI infrastructure investment over the next three years will exceed everything it spent in the previous decade combined. ByteDance is reviewing plans to pour $59 billion to $70 billion into datacenters and AI computing infrastructure this year alone — more than double the prior year.

That scale of capital deployment is exceptional for Chinese tech companies, which have operated under stricter regulatory oversight and tighter capital controls than their American counterparts. The direction, not the exact figures, is what matters: China’s four largest cloud service providers are committing to growth that mirrors the American hyperscaler cycle.

Counterpoint Research projects server DRAM and high-bandwidth memory (HBM) will account for 57% of total DRAM bit shipments in 2026, and 65% by revenue. This is the moment memory has been waiting for since the 2022–23 downturn.

The Efficiency Penalty That Makes China’s Problem Worse

Here is the detail English-language coverage rarely captures: Chinese AI accelerators like Huawei’s Ascend series are widely assessed as less compute-efficient than Nvidia’s offerings. That means a Chinese datacenter needs more chips — and proportionally more memory — to deliver the same amount of training or inference work. The gap is not marginal. It compounds the DRAM demand coming out of China relative to an equivalent American buildout.

If a Chinese cloud provider needs 20–30% more memory per petaflop than an American peer using Nvidia silicon, that gap shows up directly in the DRAM order book.

The Export Control Paradox That Benefits Samsung and SK Hynix

The United States placed HBM2 and above on its China export license list in late 2024, with a presumption of denial. Samsung and SK Hynix cannot freely sell their most advanced memory to Alibaba, ByteDance, or Baidu. That is the headline constraint everyone notes.

But the constraint works in their favor.

China’s domestic memory producer, Changxin Memory Technologies (CXMT), is absorbing some of the growing demand. CXMT recently raised 14.4 trillion won through a listing to expand capacity. On paper, that helps close the supply gap. In practice, industry sources estimate a three-year technology lag for DRAM and a five-year lag for HBM between CXMT and the Korean leaders.

Those gaps are not closing fast. Memory fabrication requires process know-how, yield management, and customer qualification cycles that cannot be shortcut with capital alone. CXMT is scaling volume, but the quality and performance ceiling remains well below Samsung and SK Hynix output for the foreseeable future.

The result is a market where Chinese demand flows indirectly into the global DRAM pool, tightening supply everywhere — including for Samsung and SK Hynix’s core customers in America, Europe, and elsewhere.

Long-Term Contracts Are the Real Weapon

Samsung is responding with a strategy that turns the shortage into leverage. The company is filling a significant portion of its production capacity with five-year long-term agreements (LTAs). This locks in pricing and volume commitments before the market can fully price in the supply crunch. It also raises the cost for any buyer trying to secure memory on the spot market during the shortfall.

SK Hynix operates on a similar logic, though with a different product mix tilted more heavily toward HBM for the American AI chip market.

An industry source told Chosunbiz that US-China datacenter investment occurring simultaneously makes the global DRAM shortage likely to persist longer than previously expected, and that CXMT’s capacity expansion, while real, is still far from sufficient in both volume and quality to meet China’s accelerating AI infrastructure needs.

Who Wins, Who Loses, and What Comes Next

Who wins: Samsung and SK Hynix. They get stronger pricing power on standard server DRAM from a demand surge they cannot directly serve. Their LTAs insulate revenue. Their American customers — the very companies driving the competing side of this demand — remain loyal because there is no alternative source of equivalent quality and volume.

Who loses: Buyers on the spot market. Mid-tier cloud providers without LTA coverage will face higher prices and longer lead times. CXMT gains capital and visibility but not competitive parity for years. American export controls achieve their stated goal of restricting advanced memory flow to China while inadvertently reinforcing the pricing power of the very companies those controls were designed to pressure.

What happens next: The DRAM cycle should remain elevated through at least 2027 as both American and Chinese datacenter builds run in parallel. Memory prices will reflect that. The question is whether CXMT’s expansion eventually narrows the gap enough to erode Korean pricing leverage — or whether the five-year HBM technology barrier proves durable enough to keep Samsung and SK Hynix on top regardless.

For now, the paradox is simple: China’s AI boom, constrained by American export controls, is making Korean memory makers more powerful, not less.