The world's best AI memory maker just bet $31 billion that the boom won't bust, even as investors wonder if everyone's building data centers nobody asked for.
The Summary
- SK Hynix committed at least $31 billion in capital spending this year, a record outlay despite quarterly profit jumping sixfold
- The spending war with Samsung intensifies precisely when questions about AI overinvestment are loudest
- Record profits weren't enough: investors expected more, revealing the gap between actual AI economics and market fever dreams
The Signal
SK Hynix posted record quarterly profit, up six times year-over-year, riding demand for high-bandwidth memory chips that power AI training. The company supplies the HBM chips that go into Nvidia's H100 and H200 GPUs. When OpenAI, Anthropic, or xAI buy compute, SK Hynix gets paid. That business model is printing money right now.
But the market shrugged. Investors had priced in even bigger numbers, a tell that expectations have detached from operational reality. The AI infrastructure buildout is real. The question is whether it's proportional to actual demand or whether hyperscalers are building capacity for a future that may arrive slower than the capex cycle.
"Record profit that disappoints the market is the signature move of a boom that's gone vertical."
The $31 billion spending plan is where this gets interesting. That's not maintenance capex. SK Hynix is spending to safeguard its lead over Samsung in the HBM market, the specialized memory that's become the chokepoint in AI compute. Samsung has been playing catch-up after SK Hynix locked in early contracts with Nvidia. Now both companies are in an arms race to expand production capacity.
Here's the setup for 2027 and beyond:
- SK Hynix bets $31B that AI compute demand keeps growing faster than supply
- Samsung will likely match or exceed that figure to reclaim market share
- Both are expanding capacity while fears about overinvestment in AI capacity grow louder
The timing matters. This spending decision comes as the AI infrastructure narrative splits into two camps. One side sees hyperscalers building the rails for an agent economy that will need 10x more compute than today's chatbots. The other side sees a classic overbuilding cycle where everyone adds capacity simultaneously, creating a glut just as demand plateaus.
The Implication
Watch the HBM supply-demand balance over the next two quarters. If SK Hynix can fill all that new capacity at current margins, the agent economy thesis holds. If utilization rates drop or pricing pressure appears, we're in the overbuilding phase of the cycle. Either way, the gap between record profit and disappointed investors tells you the market has priced in perfection. Anyone building AI infrastructure companies should assume memory costs stay high longer than the bulls expect, but also plan for what happens when the marginal cost of compute drops 80% in 2027.
The real winner here might be whoever figures out how to build useful AI products that don't require training runs measured in exaflops. The memory makers are betting on more of the same. The smart money is on someone making the same capability work with a tenth of the silicon.