JPMorgan forecasts 63% CAGR for HBM bit demand 2026-2028, with persistent 15-22% supply deficit

Source: NetEase Tech

Overview

A recent research report from JPMorgan Chase indicates that even with HBM specification downgrades, bit demand compound annual growth rate will still reach 63% from 2026 to 2028. Total cumulative bit demand over three years remains unchanged at 163 billion Gb, with supply-demand gaps persisting at double-digit levels. AI is transforming memory from a cyclical commodity into a strategic asset: HBM wafer consumption is three times that of DDR5, ASICs are expected to surpass NVIDIA as the largest buyer by 2027, and the market is targeting the hundred-billion-dollar scale.

Specification Downgrade Concerns: What the Market Fears

The trigger for concern was a "spec reduction" by NVIDIA. According to TrendForce, due to DRAM supply shortages extending to 2027 and verification delays for HBM4e, NVIDIA has shifted its Rubin Ultra HBM configuration from the originally planned "HBM4e 12-layer" to parallel evaluation of alternatives including "HBM4e 8-layer." The Vera Rubin superchip's SOCAMM capacity has been halved. The market's immediate reaction: Has AI's appetite for HBM peaked?

JPMorgan's latest report offers a contrary view. The bank estimates that even if specification downgrades materialize, HBM bit demand CAGR from 2026 to 2028 remains 63%, with total cumulative bit demand locked at approximately 163 billion Gb. Specification changes alter product mix, not total demand. Breaking it down:

• The report raised demand forecasts for 2026 and 2027, with only a slight reduction for 2028
• Due to increased share of 8-layer economy products and delays of 16-layer to at least 2029, supply shortages persist across the next three years
• Supply-demand tightness improves from -20% to -16%, narrowing the gap but extending the shortage duration
• Even accounting for Rubin series downgrades and slower 12-layer/16-layer rollout, HBM supply-demand gaps remain approximately 15%, 14%, and 22% for 2026, 2027, and 2028 respectively

In short, "downgrading" is not cooling demand—it is a compromise forced by scarcity: when the ideal product is unavailable, buyers settle for what they can get.

First Impact: Memory Upgraded from Cyclical Commodity to "Strategic Asset"

Why can't HBM demand be reduced even with downgrades? AI poses a fundamental challenge to memory. Micron presented a striking divergence at Hot Chips 2026: computing power roughly triples every two years, while memory capacity grows less than 2x—the "memory wall" continues to rise. Autonomous driving illustrates this clearly:

• L2+ and L3 levels require over 200 GB/s memory bandwidth
• L5 level requires over 500 GB/s
• Stacking GPUs cannot solve cost and power constraints
• Only high-bandwidth memory like HBM can break through this wall

HBM's capability comes from advanced packaging: multi-layer DRAM vertical stacking, through-silicon vias (TSV) for interconnection, micro-bumps connecting DRAM to GPU, 1,024 I/Os per module (rising to 2,000 with HBM4), and bandwidth over 10 times that of traditional memory. The trade-offs are equally stark:

• Wafer consumption is approximately 3x that of DDR5
• Output per wafer is only one-third of standard DDR
• Yield rates hover between 50% and 60%
• From wafer start to packaging takes over two quarters
• In a typical GPU package, four 12-layer HBM modules occupy approximately 90% of the package silicon area

This "capacity-hungry, high-barrier" nature makes HBM a DRAM capacity black hole. JPMorgan estimates that from 2025 to 2028, 58% of new DRAM capacity will be allocated to HBM, with HBM's share of total DRAM capacity rising from 19% to 31%. Guosen Securities points in the same direction: HBM wafer starts as a share of total DRAM wafer starts will rise from 18% at end-2025 to 22% at end-2026, reaching 30% by end-2027.

Crowding-out effects follow:

• DDR4 experiences "exit-driven price increases": 16Gb DDR4 spot prices rose from $6.2 in June 2025 to nearly $77 in February 2026
• Smartphones are hit hardest: Counterpoint lowered its 2026 global smartphone shipment forecast by 14.3%
• In flagship phones, DRAM has overtaken SoC as the most expensive component

Memory's nature has changed: from a market-driven commodity to a pre-locked "strategic futures" contract. The three major memory manufacturers have sold out their 2026 HBM capacity in advance, and 2027 allocations are largely negotiated, with most customers receiving only 60% to 70% of their initial requests.

Second Impact: Industry Power Structure Reshuffled, Buyers Enter the "Capacity Grab"

Scarcity reshapes not only prices but also the industry's power structure.

First change: Demand-side shift. JPMorgan estimates that in 2026, NVIDIA will still account for 58% of total HBM demand. By 2027, cloud vendors' self-developed ASICs will overtake NVIDIA:

• ASIC share rises to 48%
• NVIDIA falls to 43%
• ASIC system shipments grow 102% year-over-year, far exceeding NVIDIA's 15%

The single-customer-dominated landscape is loosening, pushing HBM product definition toward customization.

Second change: Buyer intervention in supply. Multiple tech companies have reportedly offered to fund SK Hynix's purchase of EUV lithography equipment and invest in its Yongin new fab. Suppliers are requiring customers to prepay 30% to 40% of contract value in cash to lock capacity. This "production alliance" model is historically rare in the memory industry—traditionally, chipmakers produced based on forecasts and sold openly. Now, downstream players directly participate in capital expenditure. AI competition has evolved from model development to supply chain control.

Third change: Standards and competitive landscape shifting. JEDEC is reviewing a proposal to relax HBM package height standards from 775 micrometers to approximately 900 micrometers, clearing the way for 16-layer and 20-layer high-stack configurations. HBM4E can still use existing thermal compression bonding equipment, providing a buffer for yield and cost. On the competitive front, SK Hynix's dominant ~58% market share is being challenged. JPMorgan expects Samsung and Micron's combined HBM revenue share to reach 59% by 2027. Notably, the scramble for share amid scarcity has not triggered a price war: the bank expects HBM average selling prices to rise 54% year-over-year in 2027 and another 25% in 2028, with operating margins maintained at 60% to 70%.

Third Impact: Technology Roadmap Rewritten, Packaging Becomes the Second Battleground

The specification downgrade controversy also reveals another shift: HBM's competitive focus is moving from process technology to packaging and architecture.

Product stratification is moving toward "multi-SKU":

• 8-layer economy products: extended lifecycle, serving cost-performance demand
• 12-layer: high-performance focus
• 16-layer: delayed to at least 2029 due to yield and thermal constraints

Packaging roadmap: Thermal compression bonding (TCB) remains mainstream. Hybrid bonding, while viewed as a long-term direction, faces uncertain timing. Bernstein Research suggests that if HBM4E maintains 12-layer design, hybrid bonding adoption may be delayed to 2028. Equipment makers are responding: Morgan Stanley expects advanced packaging equipment demand to grow 59% in 2026 and 56% in 2027.

Customization is the key trend. Samsung has proposed upgrading the HBM base die from a communication layer to a customized AI platform and disclosed the ZHBM concept—vertically stacking HBM directly on XPU, targeting approximately 70% reduction in total DRAM power consumption compared to HBM5 and 2.3x bandwidth improvement. NVIDIA's NVHBM custom product is expected to launch by end-2028. Memory chips are evolving from temporary data storage to a critical link in the AI computing chain, reshaping pricing power, competitive dynamics, and technology direction.

Conclusion

Aggregating institutional forecasts, the visibility of this HBM cycle is remarkably high:

| Source | Forecast | |--------|----------| | JPMorgan | HBM market size reaches $160B–$282B from 2027–2028, accounting for 18%–24% of the three major memory makers' total DRAM revenue | | Gartner | 2026 HBM market ~$79.24B, up >130% YoY; exceeds $100B in 2027; reaches $152.77B in 2028 | | Global memory market | Revised upward to $1.82 trillion by 2028 | | Guosen Securities | Since 2000, six memory cycles have followed "fast demand, slow supply" logic; this cycle adds AI demand, supply cleanup, and technology upgrades; HBM and enterprise SSD tightness may extend to end-2027, gradually easing in 2028 |

For industry observers, two variables warrant close attention:

1. New capacity ramp-up after 2027: EUV equipment costs hundreds of millions of dollars with long lead times, creating a bottleneck for advanced chip expansion
2. Sustainability of AI capital expenditure: If cloud vendors slow investment, the "fast demand, slow supply" balance could tilt in the opposite direction

Specifications can be downgraded, but physics and capital expenditure cannot. As AI pushes computing power to unprecedented heights, the underlying scarcity of HBM is unlikely to fade anytime soon.