Memory Chip 'Long-Short Battle': Insights from Goldman Sachs' Top TMT Expert

Memory chips are currently the biggest battlefield for long-short trading in the market, and Sean Johnstone, Goldman Sachs' European TMT expert, has provided his judgment.

In his latest report, Sean Johnstone pointed out that orders for HBM (High Bandwidth Memory) and DRAM are sold out through 2027. Long-term agreements (LTAs) have set price floors, gross margins remain in the mid-to-high 70% range, and structural AI demand will keep supply tight until 2028. This is the foundation of the bullish case.

However, bears are focusing on another dimension—the "second derivative" of price increases. Johnstone noted that the month-over-month price increase rate is decelerating and is expected to peak in the second or third quarter of 2027, followed by even a mild slight decline. Meanwhile, once gross margins touch the mid-80% range, customers begin redesigning solutions to reduce their reliance on memory—a natural market feedback mechanism.

NVIDIA's moves confirm this logic. Johnstone mentioned that NVIDIA is evaluating a solution using fewer layers of HBM stacking in its next-generation "Vera Rubin" superchip. The reason is straightforward: memory costs account for approximately 62% of Vera Rubin's overall bill of materials (BOM), with the cost of SOCAMM2 memory modules on the CPU side even higher than that of HBM4 on the GPU side. Such significant cost pressure is forcing NVIDIA to actively seek alternative paths.

For conservative investors, their action is to compress valuation multiples—downgrading them from 5x to 2-3x. Johnstone emphasized that this does not mean the cycle is ending, but rather that "the easiest scarcity beta has passed." He also posed a rhetorical question: Bears need to clarify whether they are shorting fundamentals or just shorting momentum.

Additionally, Johnstone noted that The Wall Street Journal reported further that Apple is testing products from Chinese memory chip manufacturer CXMT, covering multiple product lines such as iPhone and MacBook, aiming to alleviate memory supply pressures driven by AI demand.

Absolute Prices Are Still Rising, But Remaining Upside Depends on Two Variables

Johnstone stated that absolute memory prices will continue to rise until mid-2027. However, the space thereafter depends on two things:

1. The durability of LTA price floors—whether long-term agreements can truly support the price bottom;

2. Whether the HBM mix can hedge against the normalization pressure on traditional DRAM/NAND—that is, whether the premium on high-end products can compensate for the price decline in commodity memory.

Notably, Elon Musk recently stated publicly that AI demand growth is approximately 200%, while supply growth is only about 20%. This data reinforces the structural bullish logic. However, Johnstone pointed out that the market's current pricing focus has shifted—investors are beginning to price in "slowing price growth rates," while true capacity release will not occur until late 2027 to 2028, and will be concentrated mainly in HBM, not traditional memory.

Software and Other Tech Sectors: Goldman Sachs' Latest Judgment

Beyond memory, Johnstone also reviewed several other tech topics currently attracting market attention.

The software sector is showing clear divergence. Goldman Sachs believes that software is no longer traded solely on the single factor of being an "AI loser." Data infrastructure and developer tool platforms (such as NET, PLTR, TEAM, TWLO, etc.) have rebounded from their lows, with the market willing to trust that AI will bring incremental consumption and new workloads to these companies. Pure application software vendors (such as HUBS, etc.) remain under pressure, as investors need hard evidence that AI is expanding rather than replacing their core revenue.

Goldman Sachs software analyst Gabriela holds incrementally positive views on SNOW and PANW, but incrementally negative views on ADBE, INTU, and WDAY, citing potential pressure on top-of-funnel demand and core workflows for the latter.

Regarding AI model usage (open source vs. closed source), earnings calls from Pinterest and Duolingo both released clear signals: open-source models are accelerating the replacement of closed-source models. Pinterest management stated that the transaction cost per use of open-source models is less than 8% of comparable closed-source models; Duolingo CEO Luis von Ahn stated bluntly, "As long as the quality is roughly equivalent, we switch to open-source models because they are much cheaper," and expects the future model mix to tilt further toward open source.

In terms of AI financing needs, AI-related equity financing has accounted for approximately 40% of total U.S. secondary stock offerings this year. Goldman Sachs expects hyperscalers' capital expenditures to reach $1.1 trillion in 2027, exceeding operating cash flow by approximately $150 billion, with free cash flow turning positive only by 2028. Goldman Sachs credit strategists estimate that hyperscalers can cover approximately 35% of 2027 capital expenditures through debt financing, corresponding to a global bond issuance scale of approximately $400 billion. Meanwhile, the year-over-year growth rate of S&P 500 buybacks this year is approximately +11%, with newly authorized buyback scales approaching the historical record of nearly $1 trillion year-to-date. It is estimated that full-year buybacks of approximately $1.4 trillion will offset approximately $700 billion of primary market equity supply pressure.