HBM Boom Shifts AI Chip Spending to the Machines That Build Memory

Japanese equipment maker TOWA booked ¥32.97 billion in orders for the July–September 2026 quarter, up 125.8% year over year, in the clearest signal yet that the AI chip cycle has produced a second investment front: the machinery that manufactures high-bandwidth memory.

The order figure, reported alongside quarterly sales of ¥17.53 billion, gives TOWA an order-to-sales ratio of roughly 1.88. New commitments are running at nearly twice recognized revenue. South Korea's Wonik IPS points the same direction from the other end of the fab. Its semiconductor equipment revenue is projected to grow from ₩707.3 billion in 2025 to ₩1.408 trillion in 2027 — cumulative growth of 99.1%.

The stakes are straightforward. The AI investment story has so far been measured through the revenue of Nvidia, Micron, SK hynix, and other suppliers of processors and memory for AI data centers. The equipment makers behind those chips now show the cycle extending from wafer fabrication into advanced packaging, across U.S., Japanese, South Korean, Dutch, and Israeli suppliers.

Why does HBM need more machines than DRAM?

HBM differs from conventional DRAM because fabrication is only the first step. Multiple DRAM dies must be thinned, connected through through-silicon vias (TSVs), stacked, bonded, and inspected before the finished package can attach to an AI accelerator.

Each added operation adds equipment demand, processing cost, and potential yield loss. That produces two distinct sources of equipment spending, and they are running simultaneously.

• Capacity expansion. SK hynix, Samsung, and Micron are adding wafer processing and advanced packaging capacity to meet AI demand.
• Manufacturing complexity. Each HBM generation requires more sophisticated deposition, etching, bonding, and process-control tools.

The question investors face is whether this is another cyclical capex bump or a sustained expansion. Recent results from Japanese and South Korean equipment makers show the cycle now extends well beyond the largest U.S. wafer fabrication equipment firms.

What are the U.S.-listed plays?

Four names dominate the U.S.-listed opportunity, each at a different stage of the HBM process.

Lam Research (LRCX) sells the etch, deposition, and wafer cleaning tools used to build TSVs, which connect stacked DRAM dies. TSV fabrication involves deep silicon etching and thin-film deposition. Lam reported September-quarter 2026 revenue of $8.22 billion, up 54.4% year over year and 22% sequentially. GAAP operating margin reached 40.0%, against 37.4% in the prior quarter and 34.4% a year earlier. The company attributed sequential growth to foundry and memory demand together, not HBM alone.

KLA (KLAC) sells inspection, metrology, and process control. As HBM makers raise stack heights and shrink interconnects, a single defect can scrap an entire memory stack, raising the value of catching failures early. KLA's September-quarter revenue was $4.14 billion, up 28.9% year over year. The financial distinction matters: Lam benefits directly from etch and deposition spending, while KLA benefits from the measurement steps that complex devices require. The two demand streams can grow at different rates even when serving the same customers.

Applied Materials (AMAT) participates in deposition, chemical mechanical planarization, and advanced interconnect steps relevant to TSV fabrication and wafer surface preparation. It reported July-quarter 2026 revenue of $9.12 billion, up 25.0%. On October 26, 2026, it announced an agreement to raise its stake in BE Semiconductor Industries (BESI), the Dutch maker of advanced assembly equipment, to roughly 20% of outstanding shares. The deal expands a strategic relationship begun in 2020 and remains subject to regulatory approval, so the figure is a target, not a completed transaction.

Camtek (CAMT), headquartered in Migdal HaEmek, Israel, offers the most concentrated exposure to advanced packaging inspection and metrology. September-quarter revenue reached about $161 million, up 28% year over year. Management projects revenue above $800 million in 2027, versus roughly $595 million expected in 2026 — growth of about 35%, well above the broader equipment market. The smaller revenue base cuts both ways: higher percentage upside, greater sensitivity to customer spending decisions and qualification schedules.

Company	Quarter	Revenue ($B)	YoY growth
Lam Research	Sep. 2026	8.22	54.4%
KLA	Sep. 2026	4.14	28.9%
Applied Materials	Jul. 2026	9.12	25.0%
Camtek	Sep. 2026	0.161	28.0%

What does TOWA's order surge reveal?

TOWA's divergence between orders and sales is the useful signal. Orders grew 125.8%; sales grew 14.1%. Customers are committing capital faster than equipment can ship.

The Kyoto-based company, which supplies molding and encapsulation equipment, also raised guidance. Its forecast for the fiscal year ending March 2027 increased from ¥64 billion to ¥71 billion in revenue, a 10.9% revision, and from ¥10.24 billion to ¥13.15 billion in operating profit, a 28.4% revision.

The arithmetic shows operating leverage. The additional ¥7 billion of revenue carries ¥2.91 billion of additional operating profit — an implied incremental margin of 41.6%, though product mix and expenses also play a role. TOWA's disclosures identify expanding AI and data center investment, including HBM, as demand drivers, and the company serves other packaging applications, so the growth is not exclusively HBM.

Where does Wonik IPS fit?

Wonik IPS (240810.KQ), based in Pyeongtaek, works the front end: PECVD, ALD, and thermal processing systems for DRAM, NAND, and other applications.

It reported preliminary third-quarter 2026 revenue of about ₩337.8 billion, up 25.8% year over year, with operating profit of roughly ₩66.7 billion, up 142.0% from the year-ago quarter. Operating margin was about 19.7%. Semiconductor equipment sales jumped from ₩189.6 billion in the second quarter to ₩306.6 billion in the third — sequential growth of 61.7%.

Metric (₩B)	2025	2026E	2027E
Semiconductor equipment	707.3	1,049.2	1,408.0
Display equipment	202.6	123.3	178.1
Total revenue	909.8	1,172.5	1,586.1
Semiconductor share	77.7%	89.5%	88.8%

The projections cover all semiconductor applications, not HBM exclusively. But the segment shift is stark: display equipment revenue declines in 2026 while semiconductor equipment grows 48.3%, then another 34.2% in 2027. Together, TOWA and Wonik IPS indicate that memory investment is expanding across multiple stages of manufacturing — packaging at TOWA, wafer processing at Wonik.

How global is the supply chain?

The supplier base extends far beyond U.S. manufacturers, and the distribution explains why HBM growth does not lift every vendor equally.

• Lam Research — Fremont, California: TSV etch, deposition, cleaning
• KLA — Milpitas, California: inspection, metrology, process control
• Applied Materials — Santa Clara, California: deposition, CMP, bonding; planned ~20% BESI stake
• Camtek — Migdal HaEmek, Israel: packaging inspection and metrology
• Wonik IPS — Pyeongtaek, South Korea: PECVD, ALD, thermal processing
• TOWA — Kyoto, Japan: molding, encapsulation
• Hanmi Semiconductor (042700) — Incheon, South Korea: thermocompression bonding for HBM die stacking
• DISCO (6146) — Tokyo, Japan: grinding, thinning, dicing
• BESI — Duiven, Netherlands: hybrid bonding, advanced assembly

Higher stack counts raise the value of die thinning and bonding. Smaller interconnects add inspection, metrology, and surface preparation requirements. Front-end DRAM expansion benefits etch, deposition, and thermal toolmakers. The opportunity tracks the processes being upgraded, not the count of HBM devices produced.

Could hybrid bonding reshape the market?

Current HBM production relies on established bonding processes, but higher stacks and tighter interconnect pitches strain package thickness, electrical performance, and yield. Hybrid bonding offers finer-pitch copper-to-copper connections through direct bonding of prepared surfaces, reducing dependence on conventional microbumps.

The technology carries its own demands: exceptionally flat, clean surfaces, precise alignment, and tight defect control — all of which push spending toward surface preparation, metrology, inspection, and specialized bonding equipment. Applied's planned BESI stake connects it directly to that transition. Timing, however, remains uncertain, and the two bonding approaches may coexist for years if different HBM configurations require different solutions.

Can equipment demand outlast the capacity build?

Complexity spending differs from capacity spending. A memory maker can add wafer capacity to raise output, but building more advanced HBM may require new tools even where wafer capacity is sufficient.

There are limits. Throughput gains, process integration, and yield improvements can reduce the number of systems needed. Manufacturers can delay purchases if memory prices weaken or AI infrastructure investment slows. New facilities will eventually add supply, and if HBM capacity outruns demand, price pressure could curb capex — the traditional relationship between memory supply, pricing, and capital investment still applies.

The investor takeaway from the reported numbers: the companies best positioned are not those with the greatest exposure to HBM capacity expansion, but those whose equipment stays essential as memory manufacturing advances through successive technology generations.