Recent Bank of America Securities roadshows in Hong Kong indicate that most investors have turned pessimistic on memory chips, driven mainly by doubts over AI capital expenditure prospects and memory price trends, with sentiment deteriorating markedly compared with roadshows in August and September.
Jukan, an independent technology analyst at Citrini Research, said after visiting Shanghai that he had "lost conviction in his positions." A veteran investor pointed out to him that even if memory stocks trade at just 3 times 2027 earnings, they still cannot be called cheap. Historical data shows that once memory prices fall more than 60% from their peak, customers typically abandon prepayments and walk away from contracts, greatly reducing the value of the long-term agreements signed under the current super-cycle.
Hong Kong Roadshow: Most Investors Have Turned Bearish
In its Global Memory Technology weekly report published on November 6, Bank of America Securities disclosed that its research team recently traveled to Hong Kong to exchange views with multiple investors on the memory industry, and the discussions pointed to one signal: most people have turned pessimistic.
The report said: "We met with investors in Hong Kong to discuss memory, and most appeared to have turned bearish, mainly due to growing concerns about AI capital expenditure, memory price trends and competition from China."
A chart comparing BofA's memory indicator with memory stock price performance visually illustrates the sharp pullback in share prices from October to early November.
It is worth noting that some investors said they had already reduced positions in advance. At the same time, they pointed out that Samsung Electronics and SK Hynix offer relatively high shareholder returns and low valuation multiples. A small number asked about shorting opportunities, but most still expect memory manufacturers to maintain high profit margins and substantial free cash flow even if average selling prices decline in 2027 or 2028.
BofA concluded: "Overall, most investors said it is difficult to find new catalysts that could exceed expectations. Sentiment is clearly more bearish than during the August to September roadshows, when most people were still bullish, or at least neutral."
Shanghai Trip: Veteran Investor Says Even 3 Times Earnings Is Not Cheap
Meanwhile, Jukan, a semiconductor analyst at Citrini, recorded on social media his impressions after visiting a veteran investor in Shanghai. This veteran investor believes: "Even if memory stocks trade at only 3 times 2027 operating profit, they still cannot be called cheap."
What is the reason? He pointed to the fragility of long-term agreements (LTAs).
"Memory LTAs are basically meaningless. The best outcome for LTAs signed now is merely getting customers to share part of memory manufacturers' capital expenditure. Once prices fall, customers will happily give up their deposits (prepayments) and walk away."
He cited historical data to further support this: "Historically, whenever prices fall more than 60% from their peak, customers default and break contracts, even if LTAs have already been signed."
The valuation logic derived from this is: even if a 2027 shortage is certain and the stock trades at 3 times 2027 earnings, 2028 earnings are uncertain, making a higher valuation difficult to justify. He went even further: "Even if 2028 earnings are certain, we still have to consider 2029 earnings."
Jukan also raised a rebuttal regarding Taiwan Semiconductor Manufacturing (TSMC): Why does the market give TSMC such a high valuation? If the AI boom fades, would TSMC not also be hit hard?
The other party's answer pointed to the fundamental problem with memory stocks: "The issue is that memory prices have risen too violently. If memory manufacturers raised prices in a healthy way like TSMC, the situation might be different. But because they raise prices as violently as commodities, the market can only price memory stocks as commodity stocks."
The investor finally did some math: "Even if you buy at 3 times 2027 earnings, if the decline in memory prices in 2028 is enough to trigger LTA defaults, then that 3 times P/E translates to 7.8 to 10 times on 2028 earnings. So the current valuation is not cheap."
Jukan admitted at the end of his post: "If you ask me what I think about memory now... honestly, it is hard for me to give a clear answer. I hope memory stocks perform well, and I hope to see them hit new highs and keep climbing. But after this Shanghai trip... I no longer have that sense of conviction."
BofA and TrendForce: Divergence in DRAM Judgments
Although market sentiment has shifted, fundamental data have not yet shown obvious deterioration.
BofA's Memory Indicator read 175 in September, still within the super-cycle range, although it has retreated from a peak of 190 between March and April. South Korea's semiconductor exports in October were $56.6 billion, up about 259% year on year, marking 10 consecutive months of triple-digit growth.
On prices, the 16Gb DDR5 spot price hit a record high of $61 in late October before edging back to about $60.7 in early November; NAND spot prices remained under pressure, with the decline widening further from October to early November.
Daily chart of 16Gb DDR5 spot price
On the outlook, there is a clear divergence between BofA and TrendForce. TrendForce analyst Avril Wu is optimistic, expecting DRAM average prices to rise in all four quarters of 2027 and HBM prices to more than double in 2027.
BofA takes a more conservative view: it expects DRAM average prices to decline slightly in the second half of 2027 and fall about 8% in 2028 (rather than the 30% increase forecast by TrendForce); HBM prices have already risen 30%-50% year on year in 2026 and may rise only another 30%-50% in 2027, while TrendForce's target price of $4-5 per Gb "seems too aggressive." For NAND, BofA expects average prices to begin declining from the first quarter of 2027, but by a smaller margin.
A valuation comparison table for memory and semiconductor supply chain stocks visually presents the current 3-5 times valuation levels of Samsung, Hynix, Micron and others.
AI Capital Expenditure: The Core Variable Supporting Memory Demand
An important backdrop to investors turning bearish is doubt about the sustainability of AI data center capital expenditure.
BofA data show that combined capital expenditure by the five major hyperscale cloud providers, Amazon, Microsoft, Alphabet, Meta and Oracle, is expected to grow about 100% year on year in 2026 to $730 billion, and could exceed $1 trillion per year in 2027 to 2028.
However, BofA currently believes "it is still too early to expect cuts in US AI data center chip orders," citing reasons including hyperscale cloud providers actively pushing to sign long-term agreements (mainly five-year DRAM procurement), data centers under construction likely being completed as planned, and new AI data centers potentially being built more in Asia than in the United States.
Comments