Primary sources: SK hynix's CES 2026 press release (2026-01-05), its HBM4 development-completion release (2025-09-12), and its TSMC Technology Symposium 2026 exhibit page (2026-04-23). This is a secondary analysis for procurement and system selection; calculated values and media relays are labeled.

From 36 GB to 48 GB: Capacity Scales Linearly. Difficulty Does Not.

SK hynix HBM4 48GB 16-hi stack model shown at the TSMC 2026 symposium: TSVs, logic die, interposer and package substrate (official photo)
SK hynix HBM4 48GB 16-hi stack model shown at the TSMC 2026 symposium: TSVs, logic die, interposer and package substrate (official photo)

At CES 2026, SK hynix publicly showed a 16-layer, 48 GB HBM4 stack for the first time. The same release accounted for its predecessor: the 12-layer, 36 GB HBM4 had set an industry speed record of 11.7 Gbps, while the 16-layer version was "in development on customer timelines." Three months later, at the TSMC Technology Symposium, SK hynix exhibited a 3D model and magnified internal-structure views of that same 16-layer 48 GB part. The symposium's theme said the quiet part out loud: overcoming technological limits requires the integration of memory and logic.

On paper, 36 GB to 48 GB is a 33% increase (calculated: 48/36 − 1), and the per-layer capacity has not changed — both generations use 3 GB layers, i.e., 24 Gb dies (calculated). But going from 12 layers to 16 does not scale linearly in difficulty. Package height limits, wafer thinning, through-silicon via (TSV) yield, warpage, and thermal gradients all get worse, and several of them get worse fast. To understand this part, look past the capacity number and into the packaging process.

MR-MUF: One Monolithic Pour, Not Layer-by-Layer Tiling

Official SK hynix HBM4 package product photo (September 2025 development-completion press release)
Official SK hynix HBM4 package product photo (September 2025 development-completion press release)

SK hynix's development-completion release gives the official definition. MR-MUF — Mass Reflow Molded Underfill — stacks the chips, injects a liquid protective material between them, and cures it into shape. Compared with schemes that lay down film-type material one layer at a time, it dissipates heat more efficiently. The advanced MR-MUF variant adds two things: better warpage control and less stress on the stacked chips.

In engineering terms: a single reflow pass completes the multi-layer interconnects, so throughput is far higher than layer-by-layer thermocompression. The liquid underfill fills every inter-layer gap and, once cured, serves as both structural support and thermal pathway. The price is extreme sensitivity to warpage — if a die deforms during reflow, voids can remain in the gaps, and voids become thermal and reliability liabilities.

Three process routes are worth comparing. TC-NCF uses layer-by-layer thermocompression with non-conductive film: precise, but slow takt time. MR-MUF molds the whole stack at once: efficient and thermally superior, but it lives or dies on warpage control. Hybrid bonding removes the microbumps entirely and joins copper to copper directly — the next-generation route for stacks beyond 16 layers. According to an April 2026 report by The Elec (media relay), SK hynix has validated a 12-layer hybrid-bonded stack with markedly better yield than two years ago, but economics remain the barrier, and 16-layer products stay on MR-MUF. That explains why the "more advanced" bonding technology has not immediately displaced it.

Four Hurdles on the Way to 16 Layers

SK hynix CES 2026 showcase zone featuring HBM4 and SOCAMM2 AI memory (official booth render)
SK hynix CES 2026 showcase zone featuring HBM4 and SOCAMM2 AI memory (official booth render)

The first is the height limit. According to a July 2026 DigitalToday report (media relay), JEDEC relaxed the HBM package thickness ceiling from 720 µm to roughly 775 µm — and it is those extra 55 µm that make "16 layers with the bumps kept" feasible again. Packaging VP Jaesik Lee's Hot Chips 2026 talk, as relayed by AET, put numbers on the result: an HBM4 package about 775 µm tall, 12.8 × 11 mm in footprint, with up to 16,148 base microbumps and more than 20,000 TSVs.

The second is thinning and pitch. The same talk relay reports each die ground down to about 0.9× its original thickness, inter-layer gaps cut in half, and microbump pitch shrunk to about 0.9×. Dies that thin warp at the slightest provocation, and the molding compound must fill narrower gaps without trapping a single bubble.

The third is TSV yield. Stack yield is approximately per-layer yield raised to the number of layers (a principled calculation, for illustration only: at 99% per layer, 16 layers give a theoretical ~85% — not actual yield data). Every added layer tightens the tolerance for per-layer defects, which is exactly why the official messaging stresses that advanced MR-MUF "reduces mass-production risk."

The fourth is the thermal gradient. Heat has to cross the whole stack to reach the substrate, so top and bottom dies run at different temperatures — with consequences for refresh behavior and long-term reliability (principled analysis). The official claim that MR-MUF beats film-type approaches on heat dissipation is, in essence, a claim about lowering the resistance of that vertical thermal path. For system builders, a higher stack thermal resistance means heatsink and cold-plate designs must reserve a more deliberate thermal path for the HBM (principled analysis). It is also why 16-layer products and liquid cooling keep getting tied together.

The Bandwidth and Capacity Math at 11.7 Gbps

The official numbers from the 2025-09-12 release: HBM4 uses 2,048 I/O — double the previous generation — for double the bandwidth; power efficiency improves by more than 40%; performance in AI services rises by up to 69%; and operating speed exceeds 10 Gbps, well beyond the JEDEC standard's 8 Gbps. The 11.7 Gbps figure is the industry record demonstrated by the 12-layer 36 GB version (CES release).

Run the bandwidth math (calculated): 2,048 I/O × 11.7 Gbps ÷ 8 ≈ 2,995 GB/s, or about 2.93 TB/s per stack. At the JEDEC 8 Gbps baseline, the same interface yields 2 TB/s. Compare SK hynix's official HBM3E product page — 9.2 Gbps+, 1.18 TB/s, 36 GB — and the generational jump in HBM4 comes mainly from doubling the I/O count, not from per-pin speed. The capacity math (calculated, illustrative): 48 GB = 16 × 3 GB; an accelerator carrying eight HBM4 stacks holds 384 GB of memory — enough for larger models and bigger KV caches.

For inference economics, the three variables each do a distinct job: capacity decides whether the model fits, bandwidth decides token throughput, and efficiency decides output per watt in a power-capped room. That is why SK hynix put "40% better efficiency" in the headline.

What This Signals for the AI Supply Chain — and for Buyers

Mind the official wording: the 16-layer 48 GB part is "in development on customer timelines," not shelf stock. The Hot Chips relay likewise says 12-layer is in mass production while 16-layer is still in qualification. At the same CES booth, the 12-layer 36 GB HBM3E was described as the product "driving the market this year," alongside SOCAMM2, LPDDR6, and 321-layer 2 Tb QLC — a deliberate two-generation portfolio. The other storyline is cHBM (custom HBM): moving some compute and control functions that used to sit on the GPU or ASIC into the HBM base die (the official definition). The TSMC symposium theme and Korean media reporting both point to the HBM4 base die being built on TSMC's logic process (media relay).

For integrators and end customers in Guangzhou and the Pearl River Delta, the advice is concrete. Do not write 16-layer 48 GB into near-term delivery contracts as an off-the-shelf specification. When selecting systems, verify the actual HBM configuration, speed grade, and firmware support with the GPU or accelerator vendor — a demo record is not the shipping spec for every customer. Size the stack count from your model's capacity and bandwidth requirements before you negotiate price. The risks are equally clear: HBM supply is highly concentrated, specifications drift with customer customization, and prices swing hard at generation transitions. Leave change clauses in place before you lock memory specs into a long-term contract.

The efficiency numbers matter on the facility side too (principled analysis). An AI accelerator's power budget is dominated by the compute dies and the HBM; a memory-efficiency gain of more than 40% means either a looser system power budget at the same inference throughput, or more throughput under the same power cap. For local customers selecting inference appliances, write tokens per watt — not peak compute — into the acceptance criteria, and require written confirmation of HBM generation, stack count, and speed. That is how you avoid the gap between "demo-record speed in the marketing" and "standard baseline in the shipment."

Sources

Note: the bandwidth/capacity conversions and the yield illustration are calculations (labeled above). The 775 µm height, 12.8 × 11 mm footprint, 16,148 microbumps, 20,000+ TSVs, and 0.9× thinning figures come from media relays of Hot Chips/JEDEC material, not from first-hand slides.