J.P. Morgan’s latest assessment regarding HBM is highly significant for the AI memory narrative. Ahead of Micron’s earnings release, JPMorgan projects that the supply-demand imbalance will persist through 2028, even after accounting for the risk of "de-specing" in HBM.
According to JPMorgan’s model, the HBM supply deficit stands at 20% for CY26, 19% for CY27, and 16% for CY28. Even more notable is the expectation that the cumulative deficit will reach 23 weeks by the end of CY28. In other words, supply is not expected to fully catch up with demand, even as new HBM capacity comes online.
The issue of "de-specing" is crucial here. Theoretically, reducing the quantity or specifications of HBM used per AI accelerator could lower total HBM demand—a topic that dominates much of the market discussion. However, JPMorgan’s calculations already factor in this risk, yet still forecast a significant supply deficit in the HBM market for the 2026–2028 period.
The primary reason for this goes beyond just the number of gigabytes of HBM used per GPU. The number of AI accelerators is rising rapidly, and HBM demand is no longer driven solely by NVIDIA GPUs; Google TPUs, AWS Trainium, and other custom ASICs also utilize HBM. Consequently, as AI compute capacity expands, HBM demand grows across a diverse range of customers and accelerator architectures.
JPMorgan’s dollar-based forecasts for the HBM market are also striking. The model projects the HBM TAM (Total Addressable Market) to reach approximately $160 billion in CY27 and roughly $282 billion in CY28. This growth is not driven solely by bit demand; HBM prices, product mix, and supply constraints are also key factors.
In my view, one of the most significant aspects of the HBM story is that its production is far more complex than that of standard DRAM. HBM involves stacking DRAM dies, utilizing processes such as TSVs (Through-Silicon Vias), base dies, and advanced packaging. As the number of stacked layers and product complexity increase, yield and packaging capacity become increasingly critical.
Moreover, HBM and conventional DRAM share a portion of the same wafer production capacity. According to a statement made by Samsung on September 29, HBM’s share of global DRAM wafer capacity is projected to rise to approximately 30% by 2027, up from the current level of around 20%. Samsung also points out that ramping up HBM production could constrain the supply of conventional DRAM.
Therefore, the supply tightness in the HBM sector should not be viewed solely through the lens of rising HBM prices. Allocating more capacity to HBM could also impact the supply of standard DRAM. In other words, strong HBM demand driven by AI could bolster pricing power across the broader DRAM market.
Meanwhile, JPMorgan estimates that Micron’s forward bit production coverage has surpassed 35% and could rise above 50% in the future.
From an investor's perspective, a key point here is that it may be premature to assume the memory investment thesis is weakening simply due to discussions regarding HBM "de-specing." Even after factoring in the impact of de-specing, JPMorgan still forecasts a supply deficit through 2028. At the same time, HBM’s share of DRAM capacity is growing, and AI customers are actively seeking to secure future memory supplies through long-term agreements. In short, the key question in the AI memory narrative is no longer just “how much HBM will be produced?” It is about how quickly manufacturers can meet the rising demand for HBM and what impact this will have on the supply of conventional DRAM. JPMorgan’s forecast of a supply deficit extending through 2028 indicates that this question remains highly significant.
Not investment advice.
