AI accelerators are often described as a memory story: more high-bandwidth memory, or HBM, means more compute.

But more memory only becomes more compute when the memory can be joined to the logic on an advanced package.

That join now shapes how fast usable AI compute can actually be added. Extra memory does not, by itself, mean more accelerators in service.

HBM is built by stacking multiple DRAM dies for AI accelerators. That memory is then integrated with the accelerator logic on an advanced package such as TSMC’s Chip-on-Wafer-on-Substrate, or CoWoS.

What happened

On 29 July 2026, SK hynix said HBM4 mass shipments had begun in the second quarter and would ramp in the second half. Customer demand, the company said, exceeds its supply capabilities. It had finalised long-term agreements with around 10 customers, including key strategic partners. Second-quarter revenue was KRW 79.3 trillion, part of a first half that crossed KRW 100 trillion for the first time.

The Korea-Taiwan relationship is already inside that stack. At TSMC’s 2026 technology symposium, SK hynix CDO Ahn Hyun said HBM4 uses TSMC’s advanced logic processes for the base die. The join begins upstream of the final package.

Taiwan then has to bring that memory and logic together at scale. TSMC is adding packaging capacity rapidly, but capacity is still tight enough for Chief Executive C.C. Wei to say it is limiting customers’ growth.

On 16 July 2026, Wei also said CoWoS remains the majority of TSMC’s advanced packaging, and that backend capacity is still in shortage. On 14 May 2026, TSMC operations vice president B.Z. Tien said CoWoS and related advanced-packaging capacity was expected to grow at more than 80% a year through 2027, according to Focus Taiwan. The expansion and the shortage are therefore not contradictory: capacity can be growing rapidly while still falling short of customer demand.

Those packages still have to yield. Tokyo Electron (TEL) told investors on 30 April 2026 that it holds a high share in HBM interconnect etch and a high share in HBM bonders. Advanced-packaging revenue in the year to March 2026 was about JPY 200 billion, with fiscal 2027 advanced packaging guided at more than 60% growth. Those tools sit inside the equipment layer needed to produce HBM and advanced packages at the required yield. Yield is produced by a whole process of equipment, integration, materials, design and know-how. TEL is one critical input into that layer.

That creates the critical gate in the chain. An HBM stack can exist, but it does not become part of a finished AI accelerator until the package can bring memory and logic together.

What it means

At the critical HBM-to-accelerator layer, an AI accelerator becomes a three-country manufacturing story. Accelerator capacity at this layer depends on coordination across all three.

SK hynix’s record quarter sits beside allocation, not abundance.

TSMC’s packaging ramp sits beside a chief executive still describing a growth constraint. Wei also indicated that TSMC is open to customers using external advanced-packaging capacity, including Intel’s EMIB-T.

SK hynix is also moving further downstream. It has said memory competitiveness is expanding into system architecture and packaging, and its mid-to-long-term plans include the P&T7 advanced-packaging facility, to be phased according to customer demand. Memory and packaging are converging, which makes that coupling clearer. It does not remove the current packaging constraint.

An AI-accelerator order is not filled when the memory is allocated. The stack still has to meet the package, and the manufacturing chain has to yield enough finished accelerators to turn that memory into compute.

The next test is whether Wei’s packaging gap narrows, or whether the HBM4 ramp keeps running into the same constraint.

What to watch

• Packaging-gap language. Whether TSMC still describes packaging capacity as limiting customers’ growth. If that language persists, the packaging constraint is still shaping how quickly those customers can expand.

• HBM4 ramp versus the company’s own description. Whether “demand exceeds supply” remains SK hynix’s description as HBM4 shipments ramp. The test is whether allocation remains the operating condition as shipments rise.

• Japan’s tool path. Whether TEL’s advanced-packaging revenue tracks its FY2027 guidance of more than 60% growth from about JPY 200 billion. That would show whether equipment demand is scaling alongside the HBM and advanced-packaging ramp.