Server DRAM prices set to surge up to 18% in Q3 — but long-term contracts are capping the upside

Market research firm TrendForce has forecast that contract prices for server DRAM will rise 13–18% quarter-on-quarter in the third quarter of 2026, rekindling investor interest in the memory semiconductor cycle. The projected increase, however, falls short of what many in the industry had hoped for — and the reason lies in a structural shift that has quietly reshaped how memory is bought and sold: the long-term supply agreement (LTA).

Long-term contracts put a ceiling on price gains

According to TrendForce, the major memory manufacturers had already front-loaded anticipated price increases into their second-quarter contract negotiations. More importantly, leading American cloud service providers (CSPs) have locked in multi-year LTAs, which effectively cap how much suppliers can charge those customers going forward. In practice, hyperscalers such as Amazon, Microsoft, and Google have traded away the possibility of paying spot-market prices — in exchange for guaranteed supply. Memory makers, in turn, have accepted a price ceiling in exchange for stable, predictable revenue streams.

The proliferation of LTAs is a lasting consequence of the semiconductor supply crunch of 2021–22. Having been badly exposed to supply disruptions during that period, major cloud operators began to favour multi-year commitments over one-off spot purchases. Samsung Electronics, SK Hynix, and Micron — the three dominant memory producers — reciprocated eagerly, seeing LTAs as a way to retain customers and spread the risks associated with large capital expenditure programmes.

As a result, the price gains forecast for the third quarter are expected to be concentrated among customers without existing LTAs, and in volumes supplied above and beyond what current contracts stipulate. TrendForce was explicit on this point: "Price increases from the third quarter onwards will be driven primarily by non-contracted customers and excess volumes beyond existing agreements."

A CPU bottleneck, and what happens when it clears

A notable detail in TrendForce's analysis is that CSP DRAM inventories actually rose more than expected in the second quarter. The culprit, the firm argues, was a shortage of server-grade central processing units. Since a server cannot be assembled without its CPU, memory chips that had already been purchased were left sitting in warehouses, unable to be put to use.

This bottleneck, however, is expected to be temporary. The rollout of Intel's next-generation server platform and the expanded supply of AMD's EPYC processors should gradually ease CPU constraints from the second half of this year into 2026. When that happens, pent-up server assembly capacity is likely to be released rapidly, drawing down accumulated DRAM stockpiles and triggering a sharp surge in effective demand.

A 2027 supply shortage looms — and the hoarding cycle may be starting

The more significant signal is how the market is positioning itself for 2027. TrendForce has warned of a potential supply shortage in server DRAM next year and suggests that even CSPs which have already secured volumes for the second half of this year may seek to build additional buffer stocks. This raises the spectre of "double ordering" — the tendency in the memory industry for buyers, fearing future shortages, to place excess orders that amplify the inventory cycle.

History offers a cautionary parallel. During the DRAM super-cycle of 2016–18, a confluence of surging data-centre demand and rising memory content in smartphones caused prices to more than triple over three years. The current environment — characterised by rapid expansion in AI server infrastructure and an accelerating shift towards high-bandwidth memory (HBM) that is tightening supply of conventional DRAM — bears a resemblance to that earlier episode that is difficult to ignore.

Prices will keep rising, but the pace will slow

TrendForce's overarching conclusion is clear: server DRAM prices will rise in each quarter from the second half of 2026 through to the second half of 2027, but the rate of increase will gradually moderate. On the supply side, the three major producers are likely to begin expanding capacity in earnest once profitability is sufficiently restored, which would bring more chips to market. On the demand side, the LTA-dominated contracting structure itself acts as a buffer, smoothing out the violent price swings that characterised previous cycles.

For South Korea's memory industry, this creates a double-edged dynamic. Samsung and SK Hynix stand to benefit from rising spot prices, yet their LTA commitments mean they cannot fully capture the upside — the agreed contract prices are structurally anchored below where the open market might otherwise clear. At the same time, heavy investment in the transition to HBM is constraining their capacity to expand conventional DRAM production. Paradoxically, if the supply tightness lasts longer than expected, that very constraint could open the door to even steeper price increases down the line.

Ultimately, three variables will determine how the memory market plays out in 2026–27: the speed at which CPU supply normalises; the durability of investment in AI infrastructure; and the resilience of LTA-set price ceilings when faced with genuine supply-demand imbalances. The memory cycle is alive and well — but it now operates on a far more complex contractual architecture than it did in previous downturns.