The Memory Cycle vs the Logic Cycle
Memory sells a price; logic sells a socket. One clears continuously and adjusts through price, the other holds price and adjusts through volume — which is why they turn and swing so differently.
Investors speak of the semiconductor cycle in the singular, and the language does damage. Under the one label sit two economically distinct organisms: memory, where the product is an interchangeable commodity and the cycle expresses itself through price, and logic, where the product is a design win and the cycle expresses itself through volume. Treating them as one cycle produces systematic errors in both directions — reading a memory price collapse as a sector demand signal, or expecting logic order books to time a memory turn. They share fabs, physics, and an index; they do not share a machine.
Here is the framing we use: memory sells a price, logic sells a socket. A commodity memory bit from one producer substitutes freely for another's, so memory faces a single market-clearing price set by aggregate bit supply against aggregate bit demand — repriced continuously, with a live spot market underneath the contract market. Logic is the opposite construction: the product is designed into a customer's system, priced at the design win, and defended by the cost of switching once qualified. When demand weakens, memory ships roughly the same bits at collapsing prices, while logic ships fewer units at roughly the same prices. Same shock, two adjustment variables — and everything that follows about timing, amplitude, and what to watch flows from that single difference.
Two ways a price gets made
The distinction is worth grounding in the mechanics of price formation, because it is where the two cycles genuinely part. Memory pricing is made in a market: bits are graded, fungible, and traded, so every marginal shift in the supply-demand balance registers in price quickly, and no producer can opt out of the clearing level for long. Logic pricing is made in a negotiation that happens once per design cycle: the price is set when the socket is won, holds broadly across the product's life, and is insulated from spot conditions by qualification — a customer who has validated a part into a system does not swap it to save a few points mid-life, because requalification risk dwarfs the saving. Memory's price is therefore an output of the current balance; logic's price is an archive of a past negotiation. One tells you about conditions now, the other about bargaining power then.
Three layers of why memory swings hardest
The timing asymmetry deserves its own emphasis, because it is the most usable part of the framework. Memory turns first for a mechanical reason: continuous clearing means a shift in the aggregate hardware balance registers in memory pricing almost immediately, while logic metabolises the same shift slowly through order books, backlog, and scheduled deliveries. This makes memory pricing the closest thing the sector has to a real-time thermometer for the entire hardware complex — useful to an investor who owns no memory at all. Sterling's Embedded Intelligence reads it exactly that way: when Sterling sees memory pricing and logic narrative disagree, the working assumption is that memory is telling the truth earlier, because a daily-clearing market has no narrative to defend and a backlog does.
The illustration: two downturns, two different animals
Recent industry history offers clean, widely documented contrasts, stated here as history. In the 2018-2019 downturn, memory absorbed a severe pricing decline as datacenter and handset demand cooled against supply that had been committed in the boom years before — while much of the logic complex experienced an ordinary inventory correction: slower orders, unit softness, margin compression, and nothing resembling memory's collapse. The same shock, metabolised through two different machines, produced two different magnitudes of damage.
The pandemic-era cycle showed the other half of the mechanism. Memory entered its downturn among the earliest and deepest in the sector, with pricing falling far enough that producers took the step that historically marks a memory trough: announced output and capex cuts — deliberate supply withdrawal. That step matters because it exposes what actually ends a memory downturn. Demand rarely rescues a memory market on any schedule an investor can hold through; the floor is put in from the supply side, when prices near cash cost force even loaded-fab logic to give way. Logic downturns, by contrast, end through the channel — inventory digests, orders resume, units recover, prices having never really moved. Watching for supply cuts to call a logic bottom, or waiting for inventory normalisation to call a memory bottom, is using the right tool on the wrong machine.
The strongest case against this framework
The serious objection is that memory is becoming logic-like, and the framework describes an industry that is disappearing. High-bandwidth memory for accelerated computing is qualified into specific systems, sold on long-term agreements, differentiated by packaging and integration capability, and capacity for it is committed against named customers — socket economics, not spot economics. Meanwhile consolidation has left commodity memory in the hands of a few producers with long memories of what overbuilding costs, which should mean durable supply discipline and gentler cycles. On this view, the price-clearing description is a backward-looking caricature.
We accept the observations and would sharpen the framework rather than retire it. The framework was never about the product label; it is about the clearing mechanism. A memory product that is qualified, contracted, and differentiated genuinely does behave logic-like — while it stays scarce and differentiated. But memory product generations have a documented tendency to commoditise: competitors qualify, capacity catches up, and the premium product of one generation becomes the fungible product of the next, at which point the clearing mechanism reasserts itself. There is also a linkage the logic-like reading misses: premium and commodity memory draw on overlapping capacity, so allocation toward the premium line tightens the commodity balance, and disappointment in the premium line releases supply back into it. The two markets are connected vessels, not separate industries. The durable read is a spectrum position that migrates — ask of any memory product where it sits between socket and spot at this point in its generation, and assume migration toward spot as the generation matures. Consolidation, meanwhile, changes amplitude, not mechanism: discipline among few producers can soften the swings, but a single clearing price over a fixed-cost base is still the machine, and discipline is a choice that scarcity pricing has historically tested and broken.
How to apply this framework
- Use memory pricing as the hardware thermometer, whatever you own. A continuously clearing market aggregates every end market into one honest number. When memory pricing and logic narrative disagree, treat memory as the earlier truth — a spot market has no backlog to hide behind.
- Time memory off supply, not demand. Memory downturns have tended to end when producers withdraw supply — output cuts, capex cuts, utilisation cuts — as prices approach cash cost. The recovery signal is discipline, and the equities have tended to move on the deceleration of price declines, well before profits confirm.
- Time logic off the channel, not price. Logic prices are sticky, so revenue weakness means units, and units mean inventory. The digestion sequence — inventory days peaking, orders resuming — is the logic-side clock, and it runs slower than memory's.
- Classify any memory product by clearing mechanism, then assume migration. Qualified, contracted, customer-committed products sit at the socket end of the spectrum; fungible products sit at spot. Each generation drifts toward spot as competitors qualify and capacity catches up — position for where the product will clear, not where it clears at the moment of the pitch.