How do you source memory chips during industrial NAND and DRAM shortages?

Why embedded and industrial memory goes short first

Wafer capacity is a trade-off, and leading-edge lines chase the highest margin per wafer. When server and AI demand accelerates, that capacity moves, and the mature nodes industrial boards are actually built on — TSOP-48 parallel NOR, BGA-63 SPI NAND, low-density DDR3/DDR4 in motor drives, medical displays, smart meters and telecom gateways — get allocated away first.

For the manufacturer, redesigning around a different density or interface is expensive: a board re-spin, EMC re-qualification, firmware driver work. Sourcing the exact existing part number is usually the cheaper route, which is why demand concentrates so hard on a handful of memory part numbers during a crunch.

The two failure modes that are specific to memory

• Moisture taken on in storage (MSL non-compliance). Plastic BGA and TSOP memory packages absorb water from the air. JEDEC classifies most memory as MSL 3 (168 hours of floor life) or MSL 5 (48 hours); a lot that has been sitting unsealed in a humid market stall has usually exceeded that many times over. At reflow the trapped water vaporises and the package delaminates or cracks — the popcorn failure — and the damage is internal, so it passes a visual check and fails in the field.
• Desoldered salvage. Allocated memory is pulled from scrapped server boards and telecom line cards, re-balled, cleaned, and sold as new. The tell is on the solder: original factory spheres are uniform and mirror-smooth, while a re-balled part shows dull or oxidised balls, height variation, and flux residue around the ball bases. These parts pass continuity and die early under thermal cycling.

What we check before accepting a memory lot

1. 01 Moisture-barrier bag and humidity indicator card. The factory seal has to be intact, and the indicator card inside has to still read dry. A card that has gone pink past the 10% spot means the lot has been open long enough to need baking before it goes anywhere near a reflow oven.
2. 02 Die revision and speed-grade suffix. Memory part numbers carry bus timing and die revision in the trailing characters. A wrong suffix is not a cosmetic difference — it can fail timing closure, or need bootloader and driver changes the buyer did not budget for.
3. 03 Ball and lead condition under the macro camera. Factory BGA spheres are uniform in height and finish. Reworked parts show oxidised or dulled balls, height spread, and flux residue around the base — the signature of a re-ball.
4. 04 Rebake and reseal where the seal has been broken. A lot bought in cut-tape or from an opened bag is baked to the J-STD-033 profile and resealed in a fresh barrier bag with desiccant before it ships, so the buyer receives material that is ready for the oven rather than material that needs another week of preparation.

Parts that move first when memory tightens

The parts that go short first are the boring ones — the SPI NOR flash and small EEPROM on almost every industrial board, not the exotic densities. Winbond W25Q64JVSSIQ and W25Q128JVSIQ (SPI NOR), Macronix MX25L3206EM2I-12G, ISSI IS25LP128-JBLE, Microchip AT24C02C-SSHM-T and 24LC256-I/SN (I²C EEPROM) and STMicroelectronics M24C64-WMN6TP are the numbers we are asked for most often during an allocation cycle, along with DDR3 such as the Micron MT41K256M16TW-107:P.

When one of those is on allocation, the practical move is not to redesign. It is to find the exact part number in a stock position that has been held dry and can be proven to be factory material — which is a sourcing problem, not an engineering problem.