NextFin News - The artificial-intelligence buildout has found its next choke point, and it is not the chip. Eric Lee, chief executive of Taiwan's Scientech Corporation, discussed demand and the business outlook in an interview at SEMICON Taiwan this week, and the message from the advanced-packaging equipment supplier is consistent with what the company has been signaling for months: order visibility now extends into 2028, delivery lead times have roughly doubled, and the semiconductor industry's binding constraint is migrating from wafer fabrication to the packaging lines that assemble AI accelerators.
The comments land as Scientech, listed in Taipei under ticker 3583, closes a record order cycle. The company said earlier this year it had accumulated as much as NT$13 billion (US$415 million) in contract liabilities - advance payments from customers that sit on the balance sheet as a proxy for future revenue - and that delivery lead times for its own-brand equipment have stretched to eight to twelve months, roughly double the four-to-six month norm for comparable semiconductor tools. Scientech's shares closed at NT$762 on September 2, up from NT$724 at the end of August, in a 52-week range of NT$298 to NT$1,010.
The Numbers Behind the Buildout
Scientech's financials show the packaging boom has already moved from narrative to earnings. First-half 2026 revenue reached NT$6.12 billion, up 6.8% from NT$5.73 billion a year earlier, but the profit-quality shift is the real story: earnings per share of NT$9.16 in the first half alone accounts for about 66% of the full-year 2025 print of NT$13.82. Gross margin expanded to 37%, up three percentage points from 34% for full-year 2025, while operating and pre-tax margins both reached 15% and 16% respectively, up from 14%.
The mix is what is changing. Revenue from Scientech's own-brand equipment business hit NT$2.56 billion in the first half of 2026, already 69% of the full-year 2025 total of NT$3.72 billion. Management expects own-brand equipment and reclaimed-wafer sales to account for more than half of total revenue this year, up from around 40% in prior years. Full-year 2025 revenue grew 17.37% to NT$11.37 billion, the first time the company cleared the NT$10 billion threshold.
To keep pace, Scientech is spending at a record rate. Capital expenditure for 2026 is projected at NT$1.75 billion, the highest in company history and well above the NT$1.25 billion deployed in 2025. The money is going into two plants: Hukou Plant 2, which held its topping-out ceremony on March 23 and is due to begin operations before year-end, and a larger expansion in Tainan spanning roughly 21,800 square meters, scheduled for completion by the end of 2027. By 2028, the company says it will have doubled its own-brand equipment capacity.
"Our own-brand capacity has been fully booked this year. We are preparing for next year and the year after," Scientech Vice Chairman M.T. Hsu said at a media gathering earlier this year. "Customers' demand for advanced packaging equipment used in AI and HPC-related applications is very strong. We cannot catch up with their pace."
Why Packaging Became the Bottleneck
For most of the semiconductor era, advanced packaging was a back-end cost center - a necessary step, but not a strategic one. The AI accelerator changed that calculus. Modern AI chips such as Nvidia's data-center GPUs do not ship as bare dies; they arrive as multi-chip modules in which the logic die is bonded alongside stacks of high-bandwidth memory inside a single package. That process, known in TSMC's lexicon as CoWoS (Chip-on-Wafer-on-Substrate), determines how many accelerators can actually reach customers.
The capacity math explains the squeeze. TSMC is targeting roughly 130,000 CoWoS wafers per month by late 2026, nearly quadruple its late-2024 output. Yet industry estimates put 2026 demand for advanced-packaging capacity near one million wafers for the year, with lead times on TSMC's CoWoS lines running 52 to 78 weeks. Third-party research estimates suggest the top three customers account for more than 85% of available capacity, with Nvidia alone believed to hold roughly 60%.
Every CoWoS-class package requires wet-process steps - cleaning, stripping, bumping, electroplating - and that is Scientech's core competency. The company's proprietary equipment supports 2.5D and 3D packaging flows, and its wafer-reclaim business supplies the monitor wafers fabs use to calibrate tools without consuming production-grade silicon. As Scientech put it in its own release on the Hukou expansion, wet process equipment "plays an essential role in enabling high-performance advanced packaging." When packaging becomes the binding constraint, the equipment that enables it gains pricing power and visibility.
Scientech is not alone in reading the signal. ASE Technology Holding, the world's largest independent packaging and test provider, said in April that revenue from its leading-edge advanced-packaging business should rise about 10% to more than US$3.5 billion in 2026. In the same month, ASE broke ground on a new chip-testing campus in Kaohsiung with an investment exceeding NT$108.3 billion, with the first phase due to start operations in April 2027 and the second in October 2027. Across the Taiwan Strait, TSMC has confirmed plans for an Arizona advanced-packaging facility to be operational before 2029.
The Transmission Chain: From GPU Orders to Equipment Revenue
The link between a hyperscaler ordering AI accelerators and Scientech recognizing equipment revenue runs through a chain that most investors do not trace in full. It begins when a cloud provider places an order for AI systems. That order flows to the GPU designer, which books capacity at the foundry. The foundry prints the logic die and, critically, reserves CoWoS packaging slots - the step where the die is married to high-bandwidth memory. Only after the package is assembled does the accelerator ship, and only then does the hyperscaler's capital expenditure convert into deployed compute.
Each link in that chain has its own lead time, and the packaging link is currently the longest. A wafer fab can ramp in a matter of quarters once its tools are installed; a CoWoS line requires not just TSMC's own bonding equipment but an ecosystem of wet-process, plating, and inspection tools from second-tier suppliers. That is why the equipment order book leads the OSAT revenue line by roughly two years: the tools must be ordered, built, shipped, qualified, and ramped before the packaging house can bill for output.
This sequencing has a direct implication for how to read the sector. When Scientech reports NT$13 billion in contract liabilities, it is not booking future revenue - it is recording cash already received for tools not yet delivered. That cash advance is the strongest possible signal that customers expect to be running packaging lines for years, because no rational buyer pre-pays for capacity it does not intend to use. The advance payment converts a demand forecast into a balance-sheet commitment.
The chain also explains why packaging equipment is more resilient to a GPU cycle than it first appears. Even if accelerator order growth moderates, the packaging capacity already committed must still be built out - the tools are already paid for, and the lead times mean the revenue is effectively locked in. What would break the chain is not a soft quarter of GPU shipments but a cancellation of the packaging-line build itself, and there is no evidence of that in the order book.
Cyclical Wave or Structural Shift?
The central question for investors is whether this is a cyclical inventory upswing that will mean-revert, or a structural re-rating of packaging's role in the semiconductor value chain. The evidence points to structural.
Three distinctions matter. First, the duration of visibility: Scientech's order book extends to 2028, and the company is committing capital on a two-to-three-year payback horizon - behavior inconsistent with a short-cycle inventory restock. Second, the nature of the demand: AI accelerators are not a single product cycle but a multi-year infrastructure buildout, with investment forecasts for AI data-center capital expenditure reaching roughly US$1 trillion globally in 2026 and a path toward about US$1.2 trillion by 2027. Third, the technological lock-in: as packages grow larger - TSMC is expected to begin production of a 14x reticle CoWoS platform supporting 20 high-bandwidth memory stacks in 2028 - the wet-process and bonding steps become more numerous, not fewer, raising the equipment content per package.
History offers a useful contrast. The last time the semiconductor equipment sector enjoyed multi-year visibility was during the extreme-ultraviolet lithography transition, when each new node required a fundamentally new class of tool and customers committed years ahead. The advanced-packaging buildout resembles that pattern more than it resembles a memory-cycle inventory restock, where lead times expand and contract within a single year.
A cyclical reading would require evidence of a demand cliff - a pause in accelerator deployment, a shift away from large-model training, or a sudden normalization of lead times. None is visible. The more plausible cyclical component sits in the timing of equipment deliveries and the sequencing of capacity additions, not in the direction of demand.
The Second-Order Trade the Market Is Underweight
Most coverage of the AI buildout focuses on the hyperscalers buying GPUs or the foundries printing dies. The second-order implication is that revenue recognition cascades upstream before it appears in OSAT earnings. Equipment makers like Scientech book orders and advance payments years ahead of the revenue they enable, which means their earnings inflect ahead of the packaging houses that install the tools. The NT$13 billion in contract liabilities is a leading indicator for an entire equipment supply chain, not just one company.
There is also a geographic second order. Scientech has said it plans to open a sales and services office in the United States in the first half of 2027, aligned with a major customer's new advanced-packaging fab under construction in Arizona. That is the CHIPS Act buildout becoming real at the equipment level: packaging capacity is following logic fabrication onshore, and the tool vendors that can service customers in both Taiwan and the U.S. gain a moat that pure domestic players lack.
There is a third-order expectation gap worth naming. The market has largely priced the AI buildout as a foundry and GPU story. What remains underpriced is the re-rating of the back end of the line. In the historical semiconductor value chain, packaging captured a small share of total value and traded at commodity-like multiples. If advanced packaging becomes the binding constraint on AI accelerator supply for a sustained period, the value share and the multiple both re-rate - and the first beneficiaries are not the OSATs but the equipment vendors whose order books prove the constraint exists.
Looked at another way, the AI buildout is turning packaging from a commodity service into a differentiated capability. When a single advanced package can account for a substantial fraction of an accelerator's bill of materials, the companies that control yield, throughput, and process integration in packaging capture a larger share of the value pool than the historical back-end model would suggest.
The Counter-Thesis
The strongest case against the structural read is capacity overshoot. TSMC, ASE, and a cohort of smaller equipment suppliers are all adding packaging capacity at once. If AI accelerator demand normalizes faster than expected - if inference workloads prove less packaging-intensive than training clusters, or if hyperscaler capital expenditure rolls over - the industry could wake up in 2028 with too much CoWoS-class capacity and a collapsed equipment order cycle. Equipment stocks, which trade on order momentum rather than trailing earnings, would be the first to reprice.
There is also a concentration risk: with the top three customers estimated to account for more than 85% of advanced-packaging capacity, a design-win loss or a customer's decision to bring equipment development in-house could materially alter a supplier's trajectory. TSMC has a history of co-developing process equipment with vendors and then internalizing know-how. A single customer shifting its equipment roadmap could erase years of a supplier's R&D investment.
These risks are real but do not overturn the structural call on the current evidence. The falsifying signal is specific and observable: if Scientech's contract liabilities decline for two consecutive quarters, or if equipment lead times fall back below six months while capacity additions continue, the multi-year visibility thesis is broken and the cyclical-overshoot view takes over. Until then, the order book, the margin expansion, and the peer-capex commitments all point in the same direction.
What to Watch: Scenarios by Time Horizon
In the short term, the key reads are Scientech's monthly revenue prints and the trajectory of gross margin as Hukou Plant 2 ramps. A margin dip would signal execution strain or unfavorable mix as new capacity comes online; a margin rise would confirm operating leverage and pricing power. Investors should also watch whether the NT$1.75 billion capital-expenditure plan stays on track through the second half of 2026.
Medium term, three data points dominate. First, TSMC's CoWoS capacity updates - specifically whether the 2027 target of more than 200,000 wafers per month holds. Second, ASE's ability to bring its NT$108.3 billion Kaohsiung campus online on schedule in 2027. Third, whether the U.S. advanced-packaging fab in Arizona draws the expected cohort of equipment suppliers, which would validate the onshoring second-order thesis.
Long term, the structural thesis stands or falls on one question: does advanced packaging's share of the accelerator bill of materials keep rising as packages grow more complex? If heterogeneous integration continues to add steps - more bonding layers, more HBM stacks, co-packaged optics - the equipment content per package rises and the structural call holds. If packaging complexity plateaus, the re-rating reverses.
Three scenarios frame the path. The base case is continued tightness through 2027, with lead times staying above six months and Scientech's own-brand revenue share climbing past 50% of the total. The upside case is a faster-than-expected ramp of HBM4 and 20-HBM packages, which would lift equipment content per package and pull 2028 orders forward. The downside case is a hyperscaler capex pause that pushes contract-liability conversions into 2029 and forces equipment suppliers to compete on price for a smaller pool of new orders.
The AI buildout was supposed to be a story about transistors. It is becoming a story about how those transistors get put together - and the companies that sell the tools for that job are discovering that the bottleneck is where the pricing power lives.
Explore more exclusive insights at nextfin.ai.
