Memory chips have historically been one of the most punishing places to invest. The economics are brutal: capacity takes years to build, demand is lumpy, and the gap between the two has historically swung between feast and famine with little warning. Boom-and-bust is not a flaw in the memory industry's design, it is the design.
Which is why the current conversation deserves scrutiny. Because what's being argued now is that this time is structurally different. And the data, at least on the surface, makes a compelling case.
What has actually changed
The demand profile for memory has shifted in a way that matters. AI inference, the process of running trained models to generate responses, and agentic AI systems require substantially more memory per compute cycle than traditional workloads. Hyperscalers are not buying memory to refresh consumer devices on a predictable replacement cycle. They are building out data centre infrastructure on a multi-year capital commitment, with urgency driven by competitive pressure rather than consumer demand.
That urgency is changing buyer behaviour in a concrete way. SK Hynix and Samsung both report that customers are now seeking three-to-five year contracts, against the usual quarterly agreements. That's not a pricing negotiation tactic, it's a signal that buyers are more concerned about securing supply than optimising cost. When a hyperscaler with billions in AI capex at stake is willing to lock in multi-year pricing to guarantee access, the balance of power in the supply chain has genuinely shifted.
Samsung's co-CEO called it an "unprecedented supercycle." SK Hynix described a "structural shift." Macquarie's analyst said we are in year three of an upcycle "with no end in sight." The supply crunch is not expected to ease before 2028.
The oligopoly structure reinforces this. Samsung, SK Hynix, and Micron jointly control roughly 90% of the DRAM market, a concentration built through decades of consolidation that wiped out most of the competition. With supply constrained and three players setting terms, margins have reached levels that would have seemed implausible a few years ago. SK Hynix reported a quarterly operating margin of 72%. Samsung's memory margin is estimated above 60%. Nomura projects prices could rise a further 50% in Q2, potentially pushing margins above 80%.
The valuation gap is the opportunity, and the risk
Here is the part of the thesis that I find most interesting. Despite the earnings trajectory, both Samsung and SK Hynix trade at less than six times projected earnings. TSMC trades at roughly 19x. Nvidia at 22x. The gap reflects the market's residual scepticism about whether memory's historical cyclicality has genuinely been broken, or merely deferred.
If the structural case holds: multi-year contracts, persistent AI-driven demand, a constrained oligopoly, that discount closes. Samsung's net profit this year is forecast at $151 billio n. SK Hynix at $115 billion. Both figures exceed TSMC's $81 billion forecast, yet both trade at a fraction of TSMC's multiple. The re-rating, if it comes, has room to run.
But the discount exists for a reason. The market is not being irrational in pricing in cycle risk.
Where I'd push back on the bull case
The structural demand argument is sound as far as it goes, but it rests on a chain of assumptions that compound quickly. AI data centre investment has to remain aggressive. Hyperscaler capex has to hold up even if AI monetisation takes longer than expected. And the geopolitical picture, specifically Chinese producers CXMT and YMTC using the current boom to build capacity and market share, has to remain manageable.
That last point deserves more attention than it gets. Tight supply in advanced memory is currently pushing some manufacturers to consider broader use of Chinese chips in consumer electronics. It's not a near-term threat to the high-end market, but it represents a long-term overhang on the lower end of the product stack. Kwon estimates Samsung and SK Hynix could more than double combined capacity by 2035, at which point new Chinese capacity and legacy chip oversupply could create a very different market structure.
Citi's analyst put it plainly: this cycle may be stronger and longer-lasting, but cycles won't disappear. Many investors still don't believe the industry has fundamentally changed. And it's worth sitting with that view rather than dismissing it.
My read
The memory supercycle is real, and the near-term setup is as strong as it has ever been. The combination of AI-driven demand, constrained supply, multi-year contract visibility, and oligopoly pricing power is not something this industry has had before simultaneously. The valuation discount is a genuine anomaly worth taking seriously.
But the investment case should be sized accordingly. This is not a set-and-forget structural compounder in the way that TSMC or ASML are. The upside is front-loaded into a window where demand growth outpaces supply expansion, and that window has a closing date, even if no one agrees on when. The positioning that makes sense here is leaning into the cycle while it has legs, not treating it as a permanent re-rating of the industry's multiple.
The key variable to watch is hyperscaler capex guidance. Amazon, Google, Microsoft, and Meta report this week. If any of them signal a moderation in AI infrastructure spend, memory is the first place that shows up.
