HBMs have become the product that returns memory manufacturers to profit after a roughly two-year period of financial losses. As for now an irreplaceable and integral element of powerful AI accelerators, for which there is unprecedented demand (e.g. AMD announced an increase in expected revenue in this segment from originally [v lednu] expected $3.5 billion to $4 billion) is reaching astronomical prices that are rising, and despite the usual decrease in demand as prices rise, demand is also increasing in this case. It is interesting to note that The Next Platform website estimates that the current HBM makes up about $110 per 1 GB of the end price of the products, and HBM3E will push the amount up to $170.
Hynix, which is the first ever manufacturer to historically offer HBM memory (the second is Samsung, the third is Micron, and the other places are contested by Chinese manufacturers who would like to enter this market in 2026), announced at the end of February that all HBM that will only produce in 2024. Now, less than 10 weeks later, he has announced that he has practically sold out of production capacity for 2025 as well.
HBM (Hynix)
The company also expects to start mass production of the twelve-layer HBM3E in the next (third) quarter. This means a capacity of up to 36 GB per case. Twelve-layer HBMs are very difficult to manufacture in terms of yield, as the individual layers cannot be tested before connection, and the chips are therefore made from all the silicon produced and are tested only after encapsulation. With respect to yield, only Hynix and Samsung can afford the combination of twelve layers. In the longer term, however, the number of layers is expected to increase up to sixteen, which Hynix could achieve in 2028 with HBM4.
HBM3E memories should appear on available products around summer to early fall, they will concern the Nvidia H200 accelerators (H100 with HBM3E) and AMD Instinct MI350 (Instinct MI300 with HBM3E and perhaps an improved manufacturing process).
Tags: Hynix practically sold HBM
