
TL;DR: Morgan Stanley expects about a fifth of the US data center capacity planned for 2028 to slip into 2029. Its memory analysts say that delay, weaker consumer electronics demand and new supply put DDR4 price increases near their peak. This article checks that call against current prices. The tracked DDR4 spot price fell about 5% through November 16, while the PC contract benchmark rose in October. Whether the delay cools DDR4 depends on whether the postponed projects were replacing old DDR4 servers or adding AI capacity, and the bank’s model puts DDR4 close to balance only in 2028.
Recent research from Morgan Stanley estimates that about 5 GW of US data center projects, roughly a fifth of the capacity planned for 2028, will shift into 2029. Other reporting points the same way, putting about 40% of pending US data center capacity in states that want to pause new projects. The bank’s memory analysts take the delay a step further and list it as one of three reasons DDR4 price increases are nearing their peak.
That raises the question this article tries to answer: will data center delays actually push DDR4 prices to a peak? New AI servers run on DDR5 and HBM rather than DDR4, so the link is not obvious. The article looks at it from several angles: what DDR4 prices are doing now, what the forecast assumes, the two ways a delay can reach DDR4, and what it means for DDR4 still in service.
What DDR4 prices are doing now
DDR4 got expensive because Samsung and SK hynix began winding down DDR4 production while demand from industrial and networking equipment and parts of data center infrastructure held up. That squeeze has not ended, though spot prices have started to ease.
The spot market is where traders buy chips without a contract. There, the average price of an 8Gb DDR4 chip, the building block of DDR4 modules, went from $46.86 on October 28 to $44.36 on November 16, according to the live spot board. That is a drop of about 5% by our arithmetic, in small daily steps, as buyers held off. The contract benchmark is still climbing. The October contract benchmark for an 8Gb DDR4 PC chip was a record $27.50, up 5.8% from September, and contract prices for conventional DRAM, which includes DDR4, were forecast in late September to rise 10% to 15% in the fourth quarter. Nanya, a Taiwanese memory maker that gets only about 10% of its revenue from DDR5, said in October that demand for DDR4 remains strong and expects its prices to rise further this quarter, by different amounts for different products.
So DDR4 spot prices have turned down while the latest contract benchmark was still rising, and the bank’s model puts DDR4 into a small surplus only in 2028. For anyone buying DDR4 spares, the spot dip is an early sign of easing, not yet a confirmed fall.
How a data center delay reaches DDR4
New AI servers run on DDR5, HBM and other newer memory, and published coverage of the note does not say which products carry the older chips. There are two connections. The second involves far more memory per machine, although no source measures either one in total.
The first is small amounts of DDR4 inside new equipment. NVIDIA’s SN5600 AI Ethernet switch ships with 32 GB of DDR4 for its control computer. The AST2600 server management chip, which lets administrators monitor and restart a server remotely, supports up to 2 GB of DDR4, although its successor moves to DDR5.
The second runs through servers that would normally have been retired by now. With DDR5 scarce and expensive, some operators have kept DDR4 machines in service. In April, Meta began extending some data center servers from six to seven years because the memory shortage left it short of non-AI servers. An internal memo expects that shortfall to last through 2027, and the company says the change affects “a small portion of servers.” The reports do not say what memory those servers use, but machines bought six or seven years ago predate the first DDR5 server platforms, so on our reading most would be DDR4. It is also reusing DDR4 pulled from decommissioned machines as add-on memory in its own server designs. A small group of hyperscalers already runs this in production, while most other companies are still at the pilot stage. Modules reused this way also never reach the used market. For on-premises operators, an Omdia analyst said in April, the shortage “just means they delay a refresh or upgrade”. A DDR4 server kept for an extra year can create demand for replacement or upgrade modules, depending on failures, upgrades and the spares already on hand.
| Connection | How much DDR4 | Direction |
|---|---|---|
| Parts in new switches and management chips | Up to 32 GB per device | Some newer designs moving to DDR5 |
| Older servers kept in service | Replacement and upgrade modules, depending on failures and spares on hand | Depends on when DDR5 becomes easier to buy |
Would the delay end the extra years for old DDR4 servers?
Fewer new US data centers than planned in 2028 means fewer new servers buying DDR5 that year. If DDR5 becomes easier to get, keeping a seven-year-old DDR4 server running makes less sense, and the extensions start to unwind.
The delay can also work the other way, and much depends on what the postponed capacity was for. A project meant to replace an existing DDR4 fleet would keep that fleet running for another year, which supports DDR4 demand. A project meant to add new AI capacity has little direct effect on older servers. No source says how the delayed 5 GW splits between the two. The timing does not line up either. The delay hits 2028 projects, while the shortfall in that memo runs through 2027, so a 2028 slip does nothing to make DDR5 easier to buy this quarter. None of the sources here measures which effect is larger.
New supply may matter more than either. A former Nanya president warned on November 11 that new capacity from China’s CXMT, due within two years, could cause a global memory glut, which would weigh on prices with or without a data center delay. The bank’s own price outlook has also swung this year: cautious on DDR4 in March, positive again from May, and cautious now.
What it means for DDR4 still in service
A DDR4 server doing useful work has no reason to come out early because of a 2028 forecast. The open question is what happens when the extensions end. If DDR5 supply loosens and the servers kept running through the shortage are retired over a few quarters, fewer buyers would need DDR4 spares just as more modules come out of those machines. That is our inference, not something any forecast has measured. Anyone holding modules that are already out of service can get a quote for used memory to see what they are worth now.
What would show the forecast is wrong
Reports of the note list what would keep prices rising: cloud customers ordering faster again, new chip capacity arriving later than planned, and data center construction moving faster. Three other signals show whether the 2028 surplus is on track: whether DDR4 output keeps shrinking as Samsung and SK hynix exit, whether operators that extended server lives start retiring those machines, and how many pulled modules they reuse rather than sell.