The evidence describes an AI-infrastructure build-out whose principal constraints now extend well beyond GPUs. Memory, advanced packaging, power equipment, optical connectivity, storage, and data-center electricity are becoming coordinated capacity constraints. NVIDIA remains positioned at the center of accelerated-computing demand, but the pace at which that demand converts into shipped systems will depend on procurement discipline across the entire infrastructure stack.
The most material signals are long-duration power commitments, persistent memory scarcity, extended transformer and equipment lead times, advanced packaging investment, and the development of alternative architectures including high-bandwidth flash (HBF), glass substrates, hybrid bonding, solid-state transformers (SSTs), and heat-assisted magnetic recording (HAMR). A systematic analysis reveals a common operating principle: customers are accepting longer contractual commitments, higher inventory costs, and more complex supplier coordination in exchange for greater certainty of supply.
The strongest evidence concerns hyperscaler power procurement. Microsoft has entered into a 20-year off-take agreement with Constellation Energy to restart the 835-megawatt Three Mile Island Unit 1 reactor, a claim supported by four sources 67,78. Existing nuclear operators control scarce operating capacity and can contract directly with hyperscalers 52, while hyperscalers are already signing 20-year power agreements 52. Data-center leases commonly run for 15–20 years 89, colocation contracts generally last five to 15 years 10, and traditional hyperscale or fixed-capacity leases commonly span 10–25 years 85. AI infrastructure is therefore being contracted and financed as durable, utility-like capacity. This supports demand visibility for NVIDIA’s customers, but it also makes power availability, deployment timing, and total system economics decisive variables.
The Procurement System Is Becoming the Constraint
Power availability and equipment lead times
Let us examine the data dispassionately. Transformer lead times reportedly expanded from approximately 50 weeks to nearly two years 69. Other reports place current lead times at two to four years 10, compared with approximately 24–30 months before 2020 10, while some high-capacity units are quoted at as much as five years 10. A data center cannot be energized merely because its GPUs, servers, or permits are available. If transformers and grid equipment are not delivered, the planned compute capacity remains theoretical.
SSTs are presented as a potential long-term response. They require specialized power semiconductors 84 and could eventually displace low-voltage transformer and centralized UPS manufacturers 84. The timing, however, remains unsettled. SSTs are still developmental 38 and one source expects meaningful deployment to remain years away 38. Against this, Sungrow launched a ready-to-use SST product in China in July 2026 84, with the launch described as evidence of commercial progress 84. The correct conclusion is not that legacy transformer suppliers face immediate displacement. It is that power-electronics content may become more important in the architecture of AI data centers, and suppliers must commercialize SSTs rapidly to participate in the prospective 800VDC transition 84.
Battery accreditation introduces another procurement variable. Qualification depends on duration, state of charge, and the ability to sustain output during correlated scarcity periods 62. Deterministic solar forecasts can cause service-level-agreement violations 28, which illustrates the limitation of intermittent generation for facilities operating under strict uptime commitments. Announced capacity must also be distinguished from contracted and energized capacity. The six-party AI-infrastructure arrangement remains a set of memorandums of understanding subject to final agreements 81, whereas one binding master services agreement provides for an initial 10-megawatt delivery in Dowagiac, Michigan 63. These are different levels of evidence and should not be treated as equivalent.
Long-term agreements and allocation priority
Memory procurement demonstrates how scale changes the operating environment. Large memory buyers obtain superior allocation through scale and long-term agreements, while smaller customers face residual-market pricing 31. Large buyers generally operate under long-term agreements (LTAs) 31; newer agreements reportedly extend to five years 31, SK hynix customer contracts generally last around five years 41, and advance memory agreements can run for as long as seven years 3. Samsung expects LTAs to make its memory business more stable and predictable 4, reports average agreement duration of five years 15, and intends to prioritize customers able to provide firm, long-term demand commitments 4.
The description of 2027 DRAM and HBM capacity as “sold out” may therefore refer to binding demand, strategic reservations, or allocation agreements rather than fully paid orders 43. The distinction matters. A reservation can improve supply visibility, but it does not by itself guarantee that demand will convert into delivered systems or that the buyer will avoid changes in price, specification, or timing.
Current tightness is visible outside NVIDIA’s immediate supply chain. A reported DRAM shortage is delaying Apple processor packaging 6 and contributing to longer wait times for affected Apple products 24. At the same time, Apple and its manufacturing partners remain optimistic about fulfilling early iPhone 18 Pro orders 25, A20 Pro wafer production is reportedly proceeding smoothly with satisfactory yields 25, and Apple does not appear to be accumulating large wafer inventories in advance 24. These claims are compatible. Leading-edge wafer yields can remain healthy while DRAM or packaging capacity becomes the bottleneck. Successful launch execution can coexist with tight supply when allocation and prioritization are effective.
The procurement conclusion is direct: NVIDIA’s access to HBM and related memory must be assessed not solely through aggregate industry capacity, but through the duration, enforceability, and priority of its own supply arrangements. Large customers with firm commitments will be served differently from customers dependent on residual capacity.
Storage Technology Is Becoming Heterogeneous
NAND and solid-state storage
NAND is nonvolatile storage used in smartphones and other electronics 44, enterprise SSDs, memory cards, and large storage arrays 90. It retains data for years without power 90 by storing charge in floating-gate or charge-trap structures 90. Demand for NAND and SSDs is rising partly as an alternative to hard drives 31, although increased DRAM supply does not eliminate NAND demand 90. The NAND recovery has nevertheless been slower than expected 45.
Supply growth is increasingly dependent on process productivity rather than simply adding wafer capacity. Sandisk expects to grow supply primarily through node transitions and productivity improvements 57. Samsung’s V10 bonded NAND is scheduled for mass production in August 2, is described as having more than 400 layers 48, and is designed partly around energy efficiency 49. Samsung’s V10 BV-NAND is its tenth-generation technology 47, as is SanDisk’s BiCS10 90. These developments indicate that storage economics will be influenced by layer count, bonding, energy consumption, yield, and controller integration—not only by the number of wafers processed.
Endurance remains a technical constraint. NAND cells support only a limited number of program/erase cycles 19, with QLC and PLC endurance ranging from a few thousand to tens of thousands of cycles 19. For AI infrastructure, this matters because model checkpoints, datasets, inference caches, and retrieval workloads place different demands on performance, latency, capacity, and write durability. A single storage medium cannot optimize all four variables simultaneously. The resulting architecture will be heterogeneous, with high-performance flash used where latency and throughput justify its cost and lower-cost media used for less demanding or more persistent workloads.
High-Bandwidth Flash
HBF illustrates the movement of storage technology toward the compute package. Sandisk’s first-generation HBF provides up to 512GB per module 19, with simulations indicating 512GB per 16-die stack 51. Because HBF stores state in nonvolatile NAND, it requires no standby refresh power 68. That feature may improve capacity-per-watt economics, but it does not eliminate package-level engineering constraints.
Adjacent logic heat can accelerate charge leakage through the Si3N4 trap layer and SiO2 79. Package inductance and resistance create L·dI/dt and IR voltage drops 79, which can produce power-supply noise and ground bounce 79. Proposed deep-trench capacitors above 300 nF/mm² would act as local charge reservoirs in the base die and interposer 79. The engineering requirement is therefore system-level: memory density must be combined with thermal control, signal integrity, local power delivery, and suitable interconnect design.
For NVIDIA, HBF is relevant not because it replaces every other memory technology, but because it expands the design space between conventional memory and storage. The commercial outcome will depend on whether the resulting capacity, latency, endurance, and power characteristics justify the additional packaging complexity.
HAMR and the continuing role of hard drives
HDDs remain economically relevant. HDD is still the lower-cost solution for bulk storage 37, and HAMR is intended to preserve hard disks’ competitiveness against substantially more expensive flash in hyperscale storage 60. Seagate and Western Digital benefit from stronger nearline demand and higher areal density 60 and compete in nearline hard drives 60.
Seagate is targeting a 50TB HAMR HDD 13, expects customer validation to begin in 2027 13, and has proposed 5TB per platter with ten platters by late 2027 12. Seagate is accelerating that 50TB timeline 14. Western Digital’s 44TB HAMR drive remains on track for the first half of 2027 23, and the company has begun shipping 40TB ePMR/UltraSMR drives 23. The substitution of HDD by flash is therefore delayed rather than eliminated 5.
This supports a diversified storage ecosystem around NVIDIA data centers. Flash is appropriate for performance-sensitive data, active model operations, and latency-critical applications. HDD remains appropriate for bulk capacity where cost per terabyte dominates. The relevant investment question is not which technology wins universally. It is how efficiently infrastructure operators assign each technology to the workload it can serve at the lowest total cost.
Advanced Packaging Is a Bottleneck in Its Own Right
Packaging has become one of the clearest cross-sector beneficiaries of AI demand. Kulicke & Soffa is investing in thermo-compression bonding (TCB) 59, conducting hybrid-bonding research and development 59, and developing both TCB and hybrid-bonding technologies 59. The current TCB cycle is strategically important to K&S 55. TCB and hybrid bonding are nevertheless likely to coexist for an extended period rather than produce an abrupt replacement cycle 59, and hybrid-bonding adoption is occurring gradually 59.
This favors suppliers with process breadth, equipment reliability, and customer-qualification relationships. It does not support an immediate winner-take-all conclusion. Over time, the TCB market could broaden from concentration toward multiple vendors as K&S expands its participation 59. The constraint is not merely equipment availability; it is the time required to qualify an entire process window across substrates, materials, thermal conditions, and customer designs.
Horng Terng Automation, a Taiwan-founded semiconductor-equipment company established in 1994 66, jointly develops equipment with customers for new package designs and thermal materials 66. Because package designs and thermal materials continue to evolve, the company requires ongoing customer co-development and engineering 66. V Technology’s LIBRA tester performs terminal-to-terminal electrical inspection of package substrates, interposers, and redistribution layers 83, while the sector is shifting from flat-panel displays toward semiconductors 83. Kinsus is supporting the transition of an advanced-packaging architecture from research and development into mass production 30, and advanced organic substrates enter the manufacturing flow late in production 56.
These facts are material to NVIDIA. Advanced GPU packages require coordinated substrate, interposer, bonding, inspection, and thermal capabilities. A delay at any one of these stations can prevent a completed system from shipping even when accelerator demand and logic yields are strong. Qualification cycles can therefore delay monetization without reducing end-market demand.
Glass substrates and demanding component specifications
Glass is a longer-duration packaging theme. Samsung Electro-Mechanics identifies glass substrates as a longer-term commercialization opportunity 39 and is preparing to commercialize them as a medium- to long-term new business 39. Huawei is developing glass as an alternative interposer or substrate material to organic and silicon options 36, with mass production targeted for 2027 36. Samsung Electro-Mechanics can supply MLCCs, silicon capacitors, and package substrates on a turnkey basis 39. Yet only a small number of suppliers can meet the performance and reliability requirements for MLCCs of at least 100 microfarads at 125°C 39.
The opportunity and the constraint are therefore the same: more capable packaging and power-delivery components may improve AI-system performance, but the supplier base remains narrow in the most demanding applications. Glass substrates should be treated as a developing architectural option, not as a near-term substitute already capable of relieving every packaging constraint.
Equipment Demand Is Strong, but Qualification Determines Conversion
KLA management provided constructive guidance 21. Customers are already discussing deliveries for the second half of 2027 35, and KLA could have shipped more equipment if it had been able to build more 35. Supply-chain lead times affecting KLA are expected to normalize to approximately seven to nine months 34. These claims indicate robust semiconductor capital-equipment demand and constrained near-term capacity.
Tokyo Electron is exposed to deposition, etch, cleaning, and materials-engineering demand 46, while the sector is shifting from traditional oil-sealed vacuum systems toward dry, oil-free systems 75. A Tokyo Stock Exchange Prime Market company with security code 7717 is working to shorten equipment lead times 83. Cohu, by contrast, faces uncertainty over converting customer qualifications into production orders 40, although it planned to ship a production configuration at the end of August 2026 40.
A new memory customer selected PDF Solutions’ DirectScan machine for a “very unique capability” 65, but qualification of the memory opportunity is expected to take close to one year 65. The distinction is essential. Backlog, customer discussions, and equipment selection provide evidence of demand; qualification-to-order conversion determines when that demand becomes revenue.
Renesas’ product commercialization likewise depends on successful customer qualification 20. Its Gen 3 MRDIMM includes a Device Equalization Self-Train Mode that allows users to tune memory timing and receiver-equalization training 20. SiTime’s acquired Renesas timing-products business adds a substantially expanded global workforce 87. These examples reinforce the broader conclusion that AI systems require increasingly complex timing, signal-integrity, memory, and interconnect solutions. Specialized suppliers may benefit, but validation cycles lengthen as system requirements become more demanding.
Procurement Resilience: Contracts, Buffers, and Visibility
The cluster provides extensive evidence that customers are accepting inventory and contracting costs to reduce disruption risk. Toyota identified approximately 1,500 critical parts requiring alternative sourcing or stockpiling 88, mapped suppliers several tiers deep after the 2011 earthquake and tsunami 88, and allocated resources to alternative sourcing and inventory despite its just-in-time model 88. It required two to six months of inventory for critical components 88 and continuity provisions in supplier contracts 88, deliberately modifying lean manufacturing where quantified disruption risk justified the cost 88. The automotive ecosystem spans OEMs, hundreds of Tier 1 and Tier 2 suppliers, and logistics providers across 20–30 countries 17.
The lesson for NVIDIA is direct. Multi-sourcing, buffer inventory, and end-to-end supplier visibility are no longer signs of inefficient procurement when the cost of a line stoppage exceeds the cost of carrying stock. A supply chain with several single points of failure is not lean; it is merely under-instrumented.
Blanket orders
Blanket orders provide one mechanism for securing supply while managing cash and inventory. They can satisfy supplier minimum order quantities while distributing deliveries over time 26, align cash outlays and inventory with production 26, and secure availability without immediate receipt of the full quantity 26. They are particularly useful for displays, touch panels, customized modules, and stable long-term production 26. They can make deliveries more predictable and reduce last-minute purchasing 26, lower the risk that component unavailability stops production 26, and preserve a component version when alternatives would require redesign or further testing 26. They also provide configuration continuity for electronics components 26 and reduce production-downtime risk 26.
The mechanism is not risk-free. Under the applicable GTCS, a blanket order must specify total quantity, delivery schedule, and final delivery date 26. Its effectiveness depends on forecast quality, specification stability, sample validation, a realistic schedule, safety stock, supplier communication, and awareness of binding final quantities and dates 26. Delivery schedules should be reconciled regularly with the current production plan 26, and quantities aligned with standard pack sizes can simplify logistics 26. Safety stock should be calibrated to lead time, demand variability, and customer risk tolerance 26.
A blanket order secures an option on supply; it does not eliminate counterparty, working-capital, or demand-forecast risk. ASBIS customer credit terms generally range from seven to 90 days 29, export-sale credit periods vary and are normally supported by letters of credit 11, and one company’s consignment model reduces inventory exposure and may improve cash-flow stability 61. Historical telecom precedent is cautionary: Lucent Technologies and Nortel Networks reportedly lent billions to customers, making orders initially appear stronger than underlying demand 33.
Inventory discipline and the difference between announced and firm capacity
The procurement system must distinguish among four states: announced capacity, reserved capacity, qualified capacity, and delivered capacity. The six-party AI arrangement remains subject to final agreements 81; a binding master services agreement provides for an initial 10-megawatt delivery 63; and new memory opportunities may require close to one year of qualification 65. Treating all three as equivalent would convert an analytical process into an accounting error.
The same discipline applies to announced memory, storage, and power projects. SK hynix’s Cheongju M17 facility is expected to become operational in late 2028 9. That may support future memory supply, but it cannot relieve current HBM constraints. Announced battery and semiconductor capacity does not guarantee actual supply, as illustrated by Northvolt’s collapse 76. NVIDIA’s procurement quality should therefore be assessed through firm delivery schedules, supplier qualification, energized capacity, and inventory positioning—not through headline capacity alone.
Regionalization and the Competitive Environment
South Korea remains a critical manufacturing and market node. Belinker connects U.S. enterprises to South Korean manufacturing 1, with South Korea serving as the manufacturing-infrastructure location tied to its U.S. enterprise customers and Chicago as its North American base 1. South Korea is also accelerating semiconductor infrastructure: President Lee Jae Myung instructed officials to improve regulations and sharply reduce approval timelines for the Gwangju semiconductor complex 32. Labor costs remain a variable, with non-semiconductor industrial unions demanding higher pay 18. Reliance on Samsung as a dominant employer exposes vocational schools and graduates to concentration risk 16, although specialized semiconductor schools reportedly enable some students to enter Samsung jobs as early as age 17 16.
Huawei plans to launch Ascend 950 processors and Atlas 950 SuperPod systems in South Korea in Q4 2026 70, using master distributors Hansol PNS and SK Shieldus 70. These claims are single-source and should be treated as an emerging channel-development signal rather than confirmed market share. They nevertheless indicate that NVIDIA’s competitive environment is shaped increasingly by local distribution, government policy, and sovereign-computing requirements, not solely by chip benchmarks.
China’s revised IC layout-design policy is intended to support VLSI innovation 42 and forms part of broader institutional support during the 2026–30 15th Five-Year Plan 73. Designers, integrated device manufacturers, foundries, technology licensors, and other participants may need to update IP-registration, originality-certification, compliance, and litigation procedures 7. The PLA has often obtained advanced semiconductor capabilities through civilian procurement 22, increasing the strategic significance of technology controls. Controlled-goods enforcement also requires screening arbitral awards, parties, and goods against the Strategic Goods Control Order 2025 and RIEA Regulations 77. These conditions raise compliance costs and may accelerate parallel supply chains, favoring vendors with diversified manufacturing and robust export-control processes.
Adjacent Components and Materials
Optical connectivity is an enabling layer for distributed AI systems. Silicon germanium can expand optical-market demand through both higher link volumes and greater analog content per lane 54. Demand for high-speed, low-loss advanced PCB materials benefits Elite Material, ITEQ, Nan Ya Plastics, Resonac, and Mitsubishi Gas Chemical 58. The useful commercial lives of 800G optical products are extended under a strict regulatory scenario 86, while validation from multiple customer groups reduces the likelihood that optical-transceiver demand is solely the result of double ordering by a narrow supplier group 54. These signals support continued investment in high-speed networking around NVIDIA clusters, although they do not establish a specific NVIDIA supplier relationship.
Component lead times create additional system-level risk. Vishay’s backlog is growing 53, customers are placing orders more than 52 weeks in advance 53, and competitor lead times for polymer tantalum products are very long 53. Samsung Electro-Mechanics is accelerating construction of its third MLCC plant in Calamba, Philippines 50. Sanyo Chemical’s electrolytes support stable operation of aluminum electrolytic capacitors and are positioned for high-reliability applications 82, while its permanent antistatic agents serve electronic components and packaging 82. These constraints can affect power-delivery and system-assembly schedules even when GPU supply itself is adequate.
Polysilicon economics have an infrastructure dimension. The larger solar-polysilicon market can provide the industrial foundation for the smaller semiconductor-grade market by subsidizing fixed costs 64. A viable solar-grade market is increasingly necessary because producers depend on high-volume, lower-purity output for acceptable overall unit economics 72, and polysilicon producers require solar-grade volumes to sustain production across categories 72. Canadian Solar depends on imported cells and wafers 64, while U.S. policy contains a severability provision that keeps remaining provisions effective if one provision is invalidated 72. These are indirect signals, but they remain relevant to the cost and resilience of the electricity and semiconductor-manufacturing infrastructure supporting AI data centers.
Sanyo’s portfolio-transformation example illustrates the supplier capabilities required in complex industrial value chains. Its “Accelerating Transformation” plan seeks faster profit growth through responsiveness and value creation 82. Its Medium-Term Management Plan 2030 emphasizes portfolio transformation, capital efficiency, faster research and development and digital transformation, human-capital reform, and customer co-creation 82. Sanyo cites more than 1,000 patents as a competitive foundation 82, targets lubricant additives, specialty-fiber chemicals, and permanent antistatic agents in Core Materials 82, and identifies Wellness advantages in healthcare-material creation, recombinant technology, and clinical-site insight 82. Its portfolio includes Silk-Elastin 82, including wound-healing and meniscus-regeneration applications 82.
Sanyo also has overseas production in South Korea, Thailand, and the United States 82, sales offices across China, South Korea, Taiwan, the United States, and Thailand 82, and an Asian production base established through Sanyo Kasei Thailand in 1997 82 and a Shanghai trading company established in 2007 82. Its ability to translate durability requirements into heat resistance and processability 82 and combine customer needs with proprietary technology through needs-and-seeds research and development 82 is representative of the supplier differentiation increasingly required by advanced AI systems. Its plans to create employee growth opportunities and identify skill gaps 82 further emphasize the importance of technical labor.
Implications for NVIDIA
The evidence supports a constructive but qualified view of NVIDIA’s position. Demand for accelerated computing is supported by long-duration hyperscaler leases, power contracts, memory LTAs, and networking upgrades. Yet these same commitments show that the industry is becoming more capital-intensive and infrastructure-constrained. NVIDIA can capture value through its integrated platform of accelerators, networking, software, and systems, but revenue conversion will increasingly depend on customers securing electricity, transformers, advanced packaging, HBM, optical links, and qualified manufacturing capacity.
NVIDIA’s moat should therefore be assessed at the system level. A GPU performance advantage is diluted operationally if memory allocation is insufficient, package yields are delayed, power infrastructure cannot be energized, or high-speed interconnects are unavailable. Conversely, the ability to coordinate suppliers, support reference architectures, and optimize the full data-center stack could strengthen NVIDIA’s position as customers move from component purchasing toward long-duration infrastructure planning.
The development of HBF, HAMR, 800G optics, glass substrates, TCB, hybrid bonding, and 800VDC power systems indicates that the addressable opportunity is expanding across the architecture. It also shows that execution risk is spreading across more specialized vendors. The principal near-term risks are memory and packaging. DRAM and HBM scarcity can increase system costs and extend customer wait times even when leading-edge logic yields are healthy. LTAs improve supply visibility for large buyers but may disadvantage smaller customers and reduce spot-market flexibility 31. Qualification periods near one year 65 mean that a technology can be strategically important without contributing materially to near-term revenue.
The principal medium- and long-term risks are architectural and competitive. SSTs could eventually pressure legacy power-equipment suppliers while creating demand for specialized semiconductors. Huawei’s glass-substrate and AI-system initiatives could support alternative ecosystems. China’s policy and procurement framework may accelerate domestic substitution. The timing remains uncertain because SSTs and hybrid bonding are still developing, and several claims are single-source or forward-looking. NVIDIA’s valuation should therefore incorporate execution probability rather than multiply projected AI demand by accelerator share.
Several peripheral examples provide useful context. Costco deliberately differentiates Kirkland products while avoiding incumbent patent infringement 74, demonstrating that ecosystem participants can reduce dependence on established IP owners. Sony controls its console ecosystem and collects transaction fees 80. Its music business continues to grow streaming revenue 71, faces demand for next-generation mobile image sensors 71, and maintained higher PS5 production and pricing levels for FY2026 3. These claims are not NVIDIA catalysts, but they reinforce the value of platform control, recurring revenue, and ecosystem monetization—attributes commonly ascribed to NVIDIA’s CUDA and networking franchises.
Evidence quality also requires explicit weighting. KLA’s constructive guidance and customer discussions have only one source each, whereas Microsoft’s nuclear off-take has four sources 67,78 and SK hynix’s M17 facility timing has two sources 9. SanDisk’s debt-free position has one source 90. KLA customers discussing second-half 2027 deliveries 35 and the expectation that lead times normalize to seven-to-nine months 34 are encouraging, but they remain less robust than the multi-source nuclear and memory-contract evidence.
The dates require similar discipline. Most evidence falls between July 28 and August 11, 2026, making the cluster current as of the stated period. However, the deterministic-solar SLA claim and the skilled-labor TSF-AHP weighting are dated December 11, 2026 27,28, beyond the current date of August 11, 2026. They should be treated as out-of-period or possibly misdated inputs rather than contemporaneous market evidence. The unverified report of an LG Innotek–TDK robot-sensor collaboration 8 and the single-source Huawei channel claims 70 should likewise not drive the core investment conclusion.
Conclusion
The strategic issue is no longer whether AI demand exists. The issue is whether the supply chain can convert that demand into energized, qualified, and fully configured capacity. The bottlenecks are distributed across HBM, DRAM, NAND, advanced packaging, substrates, transformers, power semiconductors, capacitors, optical links, and storage media.
NVIDIA’s procurement strategy should therefore be judged by the firmness of its supply commitments, the resilience of its multi-tier supplier network, the quality of its inventory buffers, and its ability to coordinate the system around the GPU. Investors should monitor HBM availability, advanced-packaging access, firm delivery schedules, qualification-to-order conversion, energized data-center capacity, and the timing of alternative architectures such as SSTs and glass substrates 40,65,81,84. The evidence supports continued confidence in the underlying demand, but it does not support treating announced capacity as delivered throughput. Capital and attention should follow the actual constraint: relieve the bottleneck before expanding the unconstrained station.