Managing Partner
The race for AI compute is no longer won at the wafer fab alone; it is won across the industrial infrastructure that turns silicon into functioning, deployable systems. This article explains why hardware buyers must secure packaging, memory, substrates, cooling, and resilient sites before those constraints strand capital in idle clusters.
Securing the Physical Integration Stack
Hardware buyers and capital allocators have long assumed that securing wafer fabrication capacity, the nanometer-scale printing of logic dies, was the primary boundary in the race for AI compute dominance. That assumption is now costing companies billions in idle cluster capacity. Advanced logic wafer supply is stable. The actual limits have migrated downstream to a physical integration stack: advanced packaging, organic substrates, high-bandwidth memory, on-package liquid cooling, and machine tooling. Securing AI compute is no longer a semiconductor procurement problem; it has become an industrial infrastructure problem. Winners will secure scarce integration capacity, contract around multiplicative yield risk, and site clusters where cooling, power, and water permissions are durable.
Manufacturing Bottlenecks: The Advanced Integration Stack
Silicon wafer supply is no longer the binding constraint for AI accelerators. Instead, individual logic dies must be stitched together with high-bandwidth memory on silicon interposers. Advanced packaging determines the physical compute limits of next-generation hardware.
The structural transition from monolithic fabrication constraints to physical integration constraints is illustrated below:
Research from Epoch AI demonstrates that leading AI chip designers accounted for approximately 90% of global advanced packaging (CoWoS) and High-Bandwidth Memory (HBM) supply in 2025 while using only 12% of advanced logic die production. This highlights a fundamental architectural reality. To train models with trillions of parameters, accelerators like Nvidia’s Blackwell or custom cloud ASICs cannot rely on a single monolithic die. Instead, they require heterogeneous integration, where a logic die is surrounded by multiple high-bandwidth memory stacks on an ultra-thin silicon interposer.
Even with advanced packaging capacity secured, the next bottleneck is the organic integrated circuit substrate. Heterogeneous integration requires mounting the logic-and-memory package on a high-layer-count organic substrate that routes electrical signals to the system board. The industry standard is Ajinomoto Build-up Film (ABF) substrates. The material scaling requirements of advanced AI accelerators relative to legacy client processors are shown below:
An advanced AI accelerator uses 10 times as much ABF film as a standard PC processor. Because these large packages require 8 to 16 layers of micro-circuitry to connect the logic dies to the HBM stacks, high-layer organic substrates can account for 30% to 40% of the total package bill of materials (BOM) cost in complex configurations, according to semiconductor manufacturing market research (see Intel Market Research). This cost can exceed the manufacturing cost of the individual logic dies themselves, representing a major upstream material chokepoint.
The substrate supply chain is highly consolidated. Three companies, Unimicron, Ibiden, and Shinko Electric, control 75% of the global market share. Expanding substrate factory lines takes 18 to 24 months due to long equipment lead times. Procurement strategies must secure upstream substrate allocations directly rather than relying on chip designers or packaging houses. To mitigate these physical integration bottlenecks, advanced packaging and memory suppliers are executing massive capital expenditure expansions. However, a significant temporal lag remains between immediate hyperscaler cluster deployment schedules and the physical online dates of new packaging facilities. The projected growth and concentration of advanced packaging capacity are detailed below:
TSMC targets an in-house Chip-on-Wafer-on-Substrate (CoWoS) capacity of 127,000 wafers per month by the end of 2026. Independent assembly and test providers like ASE Technology plan to offer an additional 40,000 wafers per month of spillover capacity. NVIDIA has reportedly pre-booked more than half of TSMC’s projected CoWoS capacity for the 2026-2027 cycle, reserving an estimated 800,000 to 850,000 wafers to support its Blackwell Ultra and Rubin rollouts, according to supply chain disclosures (see Wccftech). This leaves custom ASIC developers and smaller cloud providers to compete for the remaining residual capacity.This capacity squeeze is further complicated by sovereign equipment delivery schedules. While ASML is ramping up shipments of extreme ultraviolet (EUV) tools, TSMC has deferred High-NA EUV integration for its A16 and A13 nodes. TSMC senior vice president Kevin Zhang noted that the $400 million per-unit cost makes High-NA tools commercially unviable for near-term nodes. Instead, TSMC is extending its existing Low-NA EUV multi-patterning lines. This strategy stabilizes wafer costs while keeping the bottleneck downstream at the packaging step.
C-suite leaders must move beyond standard commercial agreements with Tier-1 chip designers. Procurement teams must directly co-invest in upstream capacity by pre-funding capital expenditures and cleanroom tooling at organic substrate makers (such as Ibiden or Shinko) and advanced packaging OSATs. Securing the physical tooling is the only way to guarantee system delivery.
The physical integration bottleneck does not end at the factory gate. Once advanced accelerators are successfully packaged and shipped, they must be deployed in clusters of tens or hundreds of thousands of units. At this system level, the compute bottleneck shifts from manufacturing capacity to thermodynamic limits and municipal resource permitting. As individual accelerators surge past the 1,200W threshold, air-cooling systems become physically incapable of dissipating heat without catastrophic thermal throttling. Direct-to-chip liquid cooling, utilizing localized cold plates and Coolant Distribution Units (CDUs) to circulate specialized fluids directly over the silicon package, is now a mandatory design standard. The actual bottleneck remains the regulatory permitting of the water required to reject heat from the facility. The water-use profiles of different data center cooling technologies are compared below:
Large data centers can consume up to 5 million gallons of municipal water per day, according to the Environmental and Energy Study Institute. This massive resource draw has triggered severe public and regulatory backlash in primary data center hubs. In May 2026, North Carolina advanced a state-level bill (SB 730) that proposes to ban water-intensive open-loop cooling, eliminate local financial incentives, and remove sales tax exemptions on electricity for non-compliant facilities. Similarly, Virginia lawmakers passed 15 data center bills in March 2026, establishing a mandatory Department of Environmental Quality permit process for any facility utilizing 100MW or more.
For capital allocators, this regulatory shift imposes heavy CAPEX penalties. To secure permitting in water-scarce or highly regulated regions, developers are forced to adopt dry-cooling, closed-loop glycol systems. While these closed-loop dry cooler systems virtually eliminate evaporative water draw, they introduce significant capital cost premiums-often ranging from $2,500 to $4,500 per kW of IT load for advanced fluid heat-rejection equipment, according to facility economic analysis (see Energy-Solutions.co). These systems also require a substantially larger physical footprint for massive air-cooled dry-cooler arrays, directly impacting data center site selection, building structural load ratings, and cluster density economics. Prioritize regions that support waterless closed-loop dry-cooling. Legacy hubs like Northern Virginia now carry up to a 40% CAPEX penalty for compliance. Clients should prioritize regions with lower permit friction and long-term power agreements over legacy convenience, accepting minor latency penalties in exchange for multi-year permitting safety.
To navigate this highly constrained physical and municipal landscape, leading technology companies are executing a structural pivot in how they contract for advanced hardware. The legacy model of placing standard purchase orders with chip designers is insufficient. Instead, procurement teams are restructuring their supply chains through structured Capacity Reservation Fees (CRFs) and pass-through multi-tier visibility agreements. The structural contrast between traditional spot purchasing and the advanced CRF procurement model is illustrated below:
Because advanced packaging and HBM are highly consolidated, non-spot markets, leading buyers secure capacity by paying upfront, non-refundable Capacity Reservation Fees (CRFs) in exchange for strict take-or-pay volume commitments. Crucially, these contracts are no longer restricted to Tier-1 chip vendors. Sophisticated procurement teams now write pass-through visibility clauses directly into their contracts, allowing them to verify capacity and tool allocations with Tier-2 and Tier-3 suppliers, such as organic ABF substrate makers (Ibiden) and liquid-cooling CDU manufacturers. The shift to heterogeneous integration also fundamentally changes the economics of hardware assembly. In monolithic chip design, a manufacturing defect simply ruins a single die. In heterogeneous packaging, where multiple high-value chiplets (GPU logic, HBM stacks, and silicon interposers) are stitched together, the final yield is multiplicative.
Because these components are physically sealed together, a 1% defect in a single HBM stack or a minor hairline fracture in the substrate results in scrap of the entire finished system. On a $40,000 Blackwell-class system, an early-stage packaging line operating at a 60% yield results in an effective Cost-per-Functional-Unit (CFU) of ~$66,600. To mitigate this, procurement contracts must explicitly shift scrap risk to the component level, ensuring the OSAT bears the cost of low manufacturing yields rather than passing it to the system builder.To manage this yield trap, procurement heads must negotiate contracts based on Cost-per-Functional-Unit (CFU) scrap sharing rather than raw packaging costs, while simultaneously reserving dedicated testing and validation capacity. Known Good Die (KGD) pre-testing and multi-day thermal burn-in validation at the OSAT represent the ultimate operational gates. Without dedicated testing capacity, finished goods will pile on the warehouse floor.
To help hardware buyers and capital allocators navigate these overlapping physical and geopolitical constraints, Santiago & Company has built a proprietary diagnostic index: the Santiago Integration Resilience Score (SIRS). The index evaluates active and planned advanced packaging facilities globally across two core metrics:
The visual distribution of these resilience scores across global integration hubs is illustrated below:
The SIRS index reveals a counterintuitive reality for procurement heads. Samsung Cheonan scores the highest (77.0) despite regional tensions because its physical supply chain is highly consolidated. In South Korea, Samsung’s logic foundry, SEMCO’s ABF substrate plants, and Samsung’s HBM lines sit within a tight 50-kilometer radius, virtually eliminating transport lags and customs bottlenecks. In contrast, Intel’s domestic New Mexico facility (Fab 9) scores 44.4, despite operating in the United States, a low geopolitical risk environment. Its score is dragged down by extreme supply chain fragmentation, with logic wafer starts originating 8,000 kilometers away in Ireland and substrates traveling 10,000 kilometers from Asia. For capital allocators, this proves that geographic proximity to sub-component suppliers is just as critical to supply assurance as national sovereignty.
Supply chain priorities are highly dynamic. To understand how changes in strategic assumptions alter facility rankings, we conducted a multi-scenario sensitivity analysis across three distinct corporate profiles:
The facility scores across these scenarios are detailed in the index table below:
Under the Geopolitical-Heavy Profile, Intel’s New Mexico facility (Fab 9) jumps from 44.4 to 53.3 (+8.9), overtaking ASE Malaysia and TSMC Taiwan, driven by its ultra-secure domestic footprint, demonstrating how sovereign risk modeling shifts capital allocation choices. Conversely, under the HBM-Heavy Profile, Amkor Peoria improves to 66.3 (+1.7) because its physical distance to Micron’s US memory facilities (1,500 km) is closer than its distance to Ibiden’s ABF substrate lines in Asia (10,000 km), illustrating the dynamic nature of physical component mapping. C-suite leaders must demand pass-through visibility clauses in procurement contracts, establishing quarterly capacity audits down to substrate makers and coolant CDU manufacturers to provide early warning of supply delays. Contracts should also mandate CFU scrap sharing rather than raw packaging costs, ensuring the OSAT bears the financial penalty of low manufacturing yields rather than passing it to the system builder.
This analysis is based on primary corporate disclosures, earnings transcripts, and legislative records from the first half of 2026. Core data and capacity curves are derived Santiago proprietary data and from the following primary sources:
Methodology Note:
The Santiago Integration Resilience Score (SIRS) is a quantitative framework for measuring the structural risk of advanced packaging supply chains. Rather than relying on subjective assessments, the index models physical spatial logistics and standardized geopolitical risk ratings across seven global advanced packaging facilities:
Aerospace and defense companies will convert record demand into revenue only by managing qualified throughput. The scarce asset is qualified throughput, and it has to be managed node by node. The approved combination of site, process, workforce, and part at the individual production-node level.
Santiago & Company announced the public release of the AI Workforce Stability Architecture (AWSA) at the United Nations WSIS Forum 2026, introducing an open framework that helps governments monitor, prepare for, and respond to the economic and fiscal impacts of AI-driven labor displacement in service-export economies.
AI is moving on a 15-month clock; governments are moving on a five-year one. For service-export economies, the winners will not be those that forecast labor disruption most precisely, but those that pre-wire fiscal and workforce responses before the evidence becomes undeniable
SpaceX does not need to disrupt the wireless industry to reshape it. By turning satellite coverage into negotiating leverage, it is forcing carriers to rethink who owns the customer relationship when coverage comes from space.
