Goldman warns AI capex peaked as memory stocks drop 51%; JPMorgan sees six-month earnings buffer

Investors are already pricing in a peak in the AI capex cycle before earnings data have confirmed one, and the memory sell-off is the sharpest sign of this yet. This trajectory keeps pressure on semiconductor-heavy indexes such as the Kospi and on chip-maker stocks across global markets. For value-oriented buyers, JPMorgan's reasoning — that earnings have a six-month buffer — provides an argument for stepping in. Goldman's warning about credit spreads signals, however, that the funding question could keep sentiment fragile, with an Nvidia session adding to headline risk. Further widening in AI-linked credit spreads would point to further equity downside ahead.

Fears that the AI investment boom has crested have already cut memory stocks' value by half from their highs. Yet JPMorgan argues the earnings slump the market is bracing for remains at least six months away.

The sell-off has several key aspects:

• Goldman Sachs' AI memory stock basket is down 51% from the peak it reached in June, its worst drop since ChatGPT was launched.
• Driven by data-centre demand and heavy institutional and fast-money positioning, the basket had more than quadrupled in the 12 months to that June high.
• Goldman itself had earlier cautioned that heavy positioning left the sector exposed to further liquidation.
• A Goldman strategist says AI capex has likely peaked and notes that credit spreads for the largest AI cloud providers are the widest this year.
• JPMorgan finds that the market is pricing in a memory down-cycle potentially starting in the first half of 2027.
• Citing prepayments and long-term contracts, JPMorgan does not forecast earnings cuts for the coming six months.

According to Bloomberg, a Goldman Sachs basket of AI memory stocks has dropped 51% from its June peak, a deeper fall than any seen since ChatGPT was released in late 2022. The market is shifting to price in a downturn, though JPMorgan holds that earnings have not reflected it yet.

The decline marks a sharp reversal for a sector that was one of the AI boom's most popular trades. In the year running up to its late-June zenith, the basket had more than quadrupled, lifted by demand for the high-bandwidth memory that AI data centres require and by large positions taken by both fast-money and institutional investors. The stocks covered encompass Samsung Electronics, SK Hynix, Micron, Kioxia and Western Digital.

Earlier in the sell-off, when the basket had fallen by roughly a quarter of its peak, Goldman cautioned that the heavily crowded positions in memory stocks made the sector vulnerable to additional liquidation. It pointed out at that time that long positions held by asset managers were still close to record levels. The further fall implies that much of that unwinding has already occurred.

The bank's overall stance on AI is also a cautious one. A Goldman strategist has pointed out that the credit spreads of the biggest AI cloud providers are at their widest point this year, and contends that share prices have not fully adjusted. This leaves the implication that either fixed-income markets are excessively gloomy or that equities still have room to drop. The same strategist believes that the peak of AI capex has already passed, moving investor attention from earnings upgrades to financing and returns on capital.

JPMorgan takes a more tempered position on the underlying situation. In a recent analysis of the memory sector, the bank found that order visibility was lengthening. It reports that the market looks to be anticipating a downcycle for memory driven by AI financing and power constraints as early as the first half of 2027. Still, it does not see any earnings cuts materialising over the next six months, citing long-term supply deals and customer prepayments that act as a cushion.

Together, the two banks set out the contrasting terms of the debate. Share prices are already factoring in a turn in the cycle, even as near-term earnings, in JPMorgan's assessment, remain defended. Whether the sell-off has overrun depends on how quickly AI investment cools and how it is funded.

The strain on memory comes against a wider backdrop of tension in the AI trade, from questions over the scale of off-balance-sheet commitments at large technology firms to the governance overhang at Nvidia. For investors, the key signposts ahead are hyperscaler spending plans, memory pricing and contract details, and whether credit spreads for AI-linked issuers keep widening.