Memory chips: weaker guidance need not mean the end of AI. Why prices, inventories and the mix decide

Do not read weaker guidance in memory chips as the automatic end of the AI boom; separate HBM from ordinary DRAM and NAND, and watch price per bit, ASPs, bit shipments, fab utilisation and customer inventories.

A memory chip is like the fridge in AI’s Michelin-starred kitchen: the chef may be a genius, but without the right ingredients at the right moment the menu slows to a crawl.

Shares in the memory giants fell after softer guidance — in other words, after more cautious company commentary on future demand and margins. At the same time the report notes a jump in oil prices and the end of a winning streak for the Dow, so this is not an isolated problem in one corner of semiconductors but a broader shift in mood across cyclical assets.

Memory chips are not the glamorous poster of the AI revolution — that role usually goes to GPUs from NVIDIA (NVDA) or accelerators from AMD (AMD). Yet memory is often a more practical thermometer of the cycle than the processor itself. Why? Because DRAM and NAND are bought in bulk, and their cycle can be read through concrete gauges: spot prices versus contract prices, bit shipments — the volume of memory delivered, measured in capacity — ASPs, that is average selling prices, fab utilisation, wafer starts, meaning how many silicon wafers actually enter production, and inventories held by customers.

The less obvious point: weaker guidance from memory firms need not be saying “AI is over”. It may be saying something subtler: the market is starting to distinguish between scarce premium memory of the HBM type, which sits alongside AI accelerators, and more ordinary DRAM and NAND, where inventories, prices and producer discipline call the tune. It is like a restaurant that keeps selling out of truffles while the store cupboard is stuffed with ordinary flour — the kitchen is not dead, it is just that not every ingredient commands the same price.

This is precisely where the story differs from the general narrative about AI stocks, backlogs and hyperscaler capex. With memory it is not only about whether the big cloud customers keep building data centres. It is about the mix: how much of revenue and margin is driven by HBM and how much by commodity memory. If HBM is growing while ordinary DRAM or NAND faces pressure on contract prices, a single company can look like an AI winner and a victim of the classic inventory cycle at the same time.

On top of that, the jump in oil adds a second squeeze: more expensive energy can stoke inflation worries, complicate rate expectations and weigh on the valuations of growth stocks. The end of the Dow’s streak then suggests that part of the market is not simply digesting one company’s guidance, but testing whether the earlier optimism was a little too smooth.

Who it helps and who it hurts

A negative for memory makers such as Micron Technology (MU), Samsung Electronics (005930.KS) and SK hynix (000660.KS), because their sensitivity runs through ASPs, bit shipments and customer inventories. When prices per bit fall faster than shipped volumes rise, both revenue and margins can deteriorate. In NAND the same logic touches Western Digital (WDC), where shifts in storage prices quickly move margins.

Equipment suppliers to the chip industry such as Lam Research (LRCX), Applied Materials (AMAT) and ASML (ASML) can also come under pressure if more cautious guidance means memory firms defer capacity spending. Here it matters to watch not just orders in general, but utilisation and wafer starts specifically: when producers run their fabs at lower loadings or start fewer wafers, that is a different kind of pain from simply pushing a delivery back a quarter.

The impact is mixed for AI names such as NVIDIA (NVDA), AMD (AMD) and data-centre networking of the Arista Networks (ANET) variety. If the problem is mainly in ordinary memory, the core of AI infrastructure may prove more resilient. But if there are hints of a pause in capital spending by the hyperscalers, that is the largest cloud customers, the sensitivity spreads to networking gear, servers and power infrastructure too. For memory, though, the first question is more practical: is HBM’s share of the mix growing fast enough to outweigh the weaker commodity side?

A positive for energy: higher oil supports producers such as Exxon Mobil (XOM) and Chevron (CVX), and service firms of the SLB (SLB) type. A negative for fuel consumers, for instance airlines Delta Air Lines (DAL) and United Airlines (UAL), the logistics of FedEx (FDX) and parts of retail, because dearer fuel takes the oxygen out of both margins and household wallets.

With news like this it is not enough to read the word “guidance” as a red siren. A better checklist: 1) separate HBM from ordinary DRAM and NAND, 2) compare spot and contract prices — spot moves faster, contracts hurt the income statement more, 3) track ASPs and bit shipments together rather than separately, 4) look for commentary on utilisation, wafer starts and customer inventories, 5) set memory makers’ commentary against the orders coming from cloud companies, 6) watch whether oil is pushing up inflation expectations and bond yields, 7) check the breadth of the move — is it only memory falling, or the whole semiconductor chain?

“Soft guidance” is the corporate version of the line: “We have plenty of orders, but for next month I would rather not shop as if catering a wedding for 200.” In memory it is even more delicate, because a warehouse full of ordinary DRAM or NAND can crush prices even while luxury HBM for AI stays in demand. For the market it means investors are recalculating earnings expectations. For share prices it usually means pressure mainly where valuations were priced for a perfect future. For the ordinary person: if that caution spreads, it can point to a slower economy; if the problem is confined to one part of the chip world, it is more a resetting of expectations than a fire across the whole economy.

This article was written by QMA Brain (artificial intelligence) and may contain errors. It is descriptive analysis and educational context, not investment advice or a forecast.

Analytical and educational content — not investment advice. The author is not a registered investment adviser. Past performance is not a guide to future results.