Fiat Elpis · Semiconductors note 02
The AI memory supercycle is a duration story, not a spot-price story
The strongest signal is not this quarter’s DRAM price. It is the widening gap between multi-year customer commitments and the time required to build qualified capacity.
From 300 recent @FiatElpis posts · expanded with primary-source research
My view is that the memory cycle is being misread as a familiar commodity spike. HBM, server DRAM and AI storage are creating a longer-duration demand claim while new supply remains slow, expensive and technically constrained.
01 · The information delta
Three producers are describing the same constraint
The useful information delta is the consistency of the language across the industry. Samsung said server DRAM, enterprise SSD and HBM demand should accelerate in the second half of 2026 while supply constraints persist. SK Hynix reported multi-year agreements with roughly ten major customers and began mass shipments of HBM4. Micron’s fiscal third-quarter release likewise tied record results to the strategic value of memory in the AI era.
This matters because three different product portfolios are pointing in the same direction. Conventional server DRAM is not being displaced by HBM; the two are complementary tiers inside the same system. HBM consumes advanced wafers, packaging capacity and engineering attention, which can tighten conventional DRAM even when bit output rises.
The historical reflex is to fade exceptional margins because commodity suppliers eventually overbuild. That reflex remains useful—but timing is the entire trade. A fab announced today does not become qualified, high-yield AI memory tomorrow. The current cycle can remain scarce long enough for earnings and balance sheets to change before the supply response arrives.
“The gap between supply and demand in the following year will become much wider.”
The original Fiat Elpis market note on X
02 · Signals to track
What makes this cycle structurally different
Demand is being reserved years ahead
Multi-year agreements, deposits and capacity reservations reduce the chance that today’s orders are only double-booking. They also transfer more planning risk to customers.
AI pulls on several memory pools at once
Training needs HBM; inference expands server DRAM and storage; larger context and agentic workloads increase the memory content per unit of compute.
Qualified capacity cannot be summoned
Cleanrooms, tools, advanced packaging and customer qualification make effective supply much slower than a headline capex budget implies.
The producers can self-fund the response
Record profitability and net-cash balance sheets reduce financing risk. The bearish question shifts from solvency to whether disciplined producers eventually overshoot.
The distinction
A supercycle does not mean prices rise forever. It means the shortage can outlast the market’s valuation horizon. The important window is 2027–2028: demand is contracted, supply projects are late-cycle, and AI memory content is still increasing.03 · What may be mispriced
The market is still pricing a short peak
The bearish case usually jumps from record margins to mean reversion without doing the intermediate work. It must identify which capacity comes online, at what yield, for which product, and against what customer commitments. Commodity DRAM additions are not automatically HBM supply; HBM capacity is not automatically qualified at the leading customer; and a nominal wafer-start number is not the same as saleable bits.
The better framework is to track the duration of the order book against the duration of the build. If committed demand extends beyond the point when incremental supply can be qualified, current earnings deserve more weight than a generic cycle chart suggests.
- Watch the share of HBM and server products in total wafer allocation, not only aggregate DRAM bits.
- Track customer deposits and long-term agreements for cancellations, price reopeners and genuine take-or-pay language.
- Separate announced fab capacity from installed tools, qualified yields and volume shipments.
- Treat Chinese commodity capacity as a 2027–2028 risk rather than an immediate substitute for leading HBM.
04 · What would change my mind
What would end the supercycle thesis
This is a duration thesis, so the invalidation signals must also be about duration:
- Major customers cut or cancel multi-year reservations rather than merely renegotiating price.
- HBM qualification broadens fast enough to remove the current supplier bottleneck.
- Conventional DRAM inventory rises while server demand and hyperscaler capex slow together.
- New Chinese capacity reaches high yields faster than expected and is allowed into global supply chains.
- Producer capex accelerates without matching deposits, contracts or end-demand growth.
Bottom line
Supply is the limiting variable
The cleanest version of the thesis is not that AI demand is infinite. It is that realizable demand is currently larger than the qualified supply the industry can deliver over the next several planning cycles.
That makes the memory complex a study in time. Spot prices will wobble, positioning will wash out and individual earnings will miss whispers. The core question is whether the supply curve can catch the contract curve before 2028. Today, the evidence says it probably cannot.
Sources & method
Primary sources, thesis separated from fact
- Samsung Electronics — second-quarter 2026 results
- SK Hynix — second-quarter 2026 financial results
- Micron — fiscal third-quarter 2026 results
- Kioxia — AI inference growth strategy
This note expands themes from the author’s recent X posts. Reported facts are linked to their sources; market interpretation is explicitly the author’s view. Market levels may change after publication.