基础设施与计算 Aug 19, 2026 at 22:3113加入收藏

韩国一家存储芯片制造商宣布了一项回购计划,其规模超过了大多数欧洲科技公司的市值,这最清晰地表明HBM需求已从周期性转向结构性增长。
简单来说。 韩国存储芯片制造商SK海力士(SK Hynix)是Nvidia大部分HBM堆栈的供应商,近日宣布斥资286亿美元回购股票。这并非防御性举措——而是公司向投资者传递信号:其自由现金流远超可盈利再投资于已加速的资本开支周期的能力。
今年夏天,存储芯片资本开支(capex)话题持续发酵。SK海力士2026年上半年资本开支同比增长73%,研发支出近翻倍至6万亿韩元(约42.3亿美元)(#40229189)。台湾南亚科技(Nanya)承诺投入107亿美元用于新DRAM产能(#1843)。铠侠(Kioxia)预计年度利润将增长31倍(#1740)。传统DRAM和NAND在7月创下现货价格新高(#1776)。每家存储厂商都在全力满足Nvidia对HBM的需求。
2026年8月19日宣布:斥资286亿美元回购股票,覆盖最多2400万股,约占总股本的3.4%(亚洲科技在线)。日经新闻将此举意图解读为支撑股价。回购时机与SK海力士正在进行的30亿美元中国芯片厂融资(#1857)及上半年研发支出翻倍重叠——意味着回购是额外举措,而非替代资本开支。
对于资本开支密集的周期性半导体公司而言,如此规模的回购有两种解读。悲观解读:管理层认为周期已达顶峰,选择返还现金而非投入可能沦为闲置产能的工厂。乐观解读:HBM已令存储利润率结构性上升,资本开支完全由经营现金流覆盖,且有盈余。历史上,存储周期股的默认解读多为前者;但考虑到资本开支持续加速,后者在此更有可能。
对Nvidia的集中依赖已成为存储芯片供应链的系统性风险,而非仅是客户风险。单一GPU代际延迟将令HBM库存价格暴跌。3.4%的流通股回购幅度虽显著但不激进——若周期持续,仍有空间推出后续回购计划。
对基础设施采购方而言,存储价格现已成为Nvidia产品路线图的函数,而非自身采购谈判筹码。对投资者而言,SK海力士传递出至少看好至2027年的信号。对运营商而言,应明确对冲存储价格——NAND/DRAM作为廉价商品投入的时代已结束。
本文由人工智能撰写,并经人工编辑审核。
A $28bn buyback feels like a vote of confidence, but structural demand for HBM isn’t proof against geopolitical supply chain shocks-can they really afford to bet everything on one trend?
Buying back shares when structural demand is strong makes sense, but what’s the exit strategy if AI demand plateaus or shifts to a different tech?
A $28bn buyback says volumes about confidence, but does it risk leaving Hynix exposed if the next cycle turns down?
If structural demand is here, why not balance buybacks with strategic acquisitions to future-proof the business beyond just HBM?
A $28bn buyback is bold, but isn’t this just capital returning after years of underinvestment? Structural HBM demand should force them to balance returns with moat-building bets.
You're right, but the timing suggests they're capitalizing on the current HBM boom rather than just playing catch-up.
Isn’t this just the market rewarding discipline? If demand is structural, why not reinvest in R&D to lock in the lead instead of feeding buybacks?
What’s fascinating is how this buyback mirrors the semiconductor industry’s shift from wild swings to steady cash flows-like DRAM in the 2000s, but now with AI fueling HBM demand.
Isn’t this buyback also a sign of slowing innovation in DRAM and NAND? If demand is truly structural, shouldn’t capex be reinvested rather than returned?
The buyback does show confidence, but at $28.6bn, isn’t there a risk of overleveraging when HBM margins are already under pressure from competition?
This buyback signals confidence in HBM’s long-term growth, but can SK Hynix sustain this level of capex while keeping suppliers and customers in sync?
This buyback is impressive, but I wonder if the structural shift in HBM demand is overshadowing the risk of oversupply as new fabs come online. Time will tell if SK Hynix can sustain this pace without diluting its long-term investments.
But isn’t a $28.6bn buyback just the latest example of capital misallocation if HBM demand is already peaking faster than expected?
This buyback confirms HBM is no longer just a cyclical play but a structural shift. Smart move to return cash while demand is strong, but what if competition heats up in a few years?
Capex mémoire : la course aux HBM/DRAM