$GLD $SLV Precious Metals & GAI: Mapping Gold and Silver Exposure Across the Hardware and Power Stack
Gold and silver enter the generative AI hardware and infrastructure stack primarily through 3 mechanisms: 1) corrosion-resistant, low-resistance contact surfaces and interconnects (gold-dominant for low-current signal integrity; silver-dominant for high-current power contacts and select RF surfaces), 2) metallurgical joining and attachment materials (silver-bearing solders and silver-sinter die attach; gold-tin solders in optoelectronics and hermetic/high-reliability packaging), and 3) electricity supply chain expansion associated with AI-driven data center load growth (silver-intensive solar PV deployment and silver-bearing grid hardware). The strongest fundamental linkage between generative AI build-outs and precious metals consumption is typically silver via the electricity ecosystem, while gold is more directly linked to electronics volumes and connector density but remains small versus the total gold market, where non-industrial demand dominates. A direct linkage to AI hardware demand is explicitly visible in 2024 gold technology demand: World Gold Council data shows technology demand rising 7% in 2024 to 326.1 t, with electronics at 270.6 t, and commentary attributing a meaningful portion of the electronics recovery to demand for high-end AI infrastructure. (World Gold Council)
The power-side scaling is material. The IEA projects global electricity generation supplying data centres increasing from 460 TWh in 2024 to over 1,000 TWh in 2030 and 1,300 TWh in 2035, implying an increase in average continuous load of ≈61.6 GW by 2030 and ≈95.9 GW by 2035, before considering the split between IT load and facility overhead. The IEA also expects renewables to meet nearly 50% of incremental data centre electricity demand growth to 2030, with renewables (primarily wind, solar PV, hydro) currently supplying about 27% of data centre electricity globally, and with PPAs and co-located renewables cited as financing channels for additional capacity. (IEA)
Silver market structure is more exposed to industrial incremental demand than gold. Silver Institute data shows total silver demand down 3% to 1.16 Boz in 2024, while industrial demand remained at a record level, and narrative attribution highlights record electronics and electrical demand, structural gains from PV and grid infrastructure, and a boost from AI-related applications. Supply-side metrics show 2024 mine production at 819.7 Moz, recycling up to 193.9 Moz, and a 2024 market deficit cited at 148.9 Moz, with cumulative deficits of 678 Moz over 2021-2024 reported in the Silver Institute’s World Silver Survey commentary. (The Silver Institute)
A practical investment conclusion follows from the structure of metal usage: direct embedded gold and silver in AI servers, networking, and power electronics is real but typically measured in grams per server and kilograms per MW, making it directionally supportive but unlikely to be a dominant driver of commodity prices at the total market level. By contrast, incremental electricity supply and grid build-out associated with AI data centre load growth has a clearer pathway to industrial-scale silver consumption through solar PV (and, secondarily, switchgear, relays, and high-current distribution hardware), though PV “thrifting” (reduced silver loadings per watt) is an active counterforce that weakens linearity between solar capacity additions and silver demand. (The Silver Institute)
PHYSICAL AND ENGINEERING RATIONALE FOR GOLD VS SILVER IN AI HARDWARE
Gold and silver are used in the AI ecosystem because the required operating regimes compress electrical, thermal, and reliability constraints into small volumes at high duty cycles. Silver has the highest electrical conductivity of common metals and very high thermal conductivity, making it valuable where ohmic losses and heat generation are limiting, especially in high-current contact interfaces and select power electronics attachment schemes. Gold is less conductive than silver but is chemically noble and highly resistant to corrosion and surface film formation, which stabilizes contact resistance in low-current signal interfaces, fine-pitch connectors, and certain wire-bond and metallization contexts where oxide or sulfide films would create intermittent failure modes. The engineering choice is therefore usually not a simple conductivity optimization; it is a combined optimization across contact resistance stability, fretting corrosion, galvanic compatibility, manufacturability, and lifecycle reliability under vibration and thermal cycling.
A core implication for AI data centres is that precious metals are used “thinly but widely.” Many applications use micron-scale coatings (gold or silver plating) across large surface areas and high connector counts, or small cross-sectional interconnects repeated at massive scale. This produces a pattern where consumption grows with: 1) unit volumes (servers, accelerators, switches, optical modules, power shelves), 2) connector and port density (especially as cluster fabrics scale), and 3) electrical distribution complexity (as rack power densities rise), but with strong incentives for thrifting because precious metals are cost-sensitive inputs.
DIRECT USES IN GENERATIVE AI COMPUTE HARDWARE
SEMICONDUCTOR PACKAGING AND MODULE ASSEMBLY
Gold in mainstream silicon device front-end processing is generally limited because gold is a deep-level impurity in silicon and is typically avoided in CMOS fabs; gold’s role is concentrated in back-end packaging and assembly. Historically, gold wire bonding (Au wire) dominated ball bonding for many IC packages; however, sustained cost pressure and process improvements have driven industry transitions toward copper wire bonding in many packages. A Kulicke & Soffa release from 2010 explicitly frames a customer transition “from gold to copper wire bonding” driven by the cost of gold wire and the maturation of copper wire bonding processes, consistent with a long-run structural substitution trend. (Kulicke & Soffa Industries, Inc.)
For generative AI accelerators specifically, the leading-edge GPU/TPU packages are typically flip-chip and advanced-package heavy, which reduces reliance on traditional gold wire bonding for the primary logic die. Nonetheless, gold remains relevant in 3 semiconductor-adjacent domains that scale with AI build-outs:
A) Supporting silicon content: AI servers require large quantities of ancillary ICs beyond the main accelerator (power management ICs, clocking, retimers, PCIe switches, NIC ASICs, microcontrollers, sensors). Many of these components still use wire bonding (gold or copper depending on node, package, and reliability requirements) and use gold plating/metallization on bond pads and leadframes.
B) High-reliability and specialty packaging: Certain modules in AI clusters (especially in optical interconnect and RF contexts) use gold-based solders and gold metallizations to achieve fluxless joining, hermeticity, and high-temperature stability.
C) Contact surfaces and testability: Even when the die interconnect is flip-chip, the package-to-board interface, module connectors, and test sockets frequently use gold plating for stable contact resistance over repeated insertion cycles.
Silver is increasingly relevant in power semiconductor packaging, especially in die attach. Silver sintering (Ag pressure sintering) is described as a high-reliability interconnection method in power electronics with high thermal conductivity (cited ranges of 130-250 W/(m·K) for silver-sintered paste) and high-temperature stability, with applications that include UPS transformation and storage, inverters, and related power infrastructure categories aligned with data centre power trains. (Power Electronics News) In practice, higher data centre power densities and the shift toward higher-efficiency power conversion (including SiC adoption in some designs) can increase the addressable footprint for silver-sinter processes, though cost trade-offs and alternative materials (including copper sintering) are explicit substitution vectors. (Power Electronics News)
PCB SURFACE FINISHES, CONTACT PADS, AND CONNECTOR PLATING
Gold usage in AI servers is materially driven by plated contact interfaces and PCB surface finishes. ENIG (electroless nickel immersion gold) is widely used as a PCB surface finish for solderability and as a contact surface. IPC-4552 specifies ENIG as a nickel layer capped with immersion gold, and sets a default minimum immersion gold deposit thickness of 0.05 µm (with an exception thickness of 0.04 µm under specific procurement conditions), measured at -4 sigma from the mean on defined pad areas. (https://t.co/sRg1uEcyy2) This gold thickness is extremely thin, which implies that PCB surface finish gold mass is typically measured in mg per board rather than g, but scales with board area, layer count, and the number of boards per system (motherboards, accelerator carrier boards, NICs, switch line cards, power distribution boards).
Connector contacts and edge fingers often use thicker “hard gold” plating than ENIG because these surfaces undergo mechanical wear and fretting. Industry references commonly cite hard gold thicknesses in the range of 30 µin (≈0.76 µm) to 50 µin (≈1.27 µm) depending on durability requirements and connector class, implying that connector gold mass can dominate PCB surface finish gold for systems with high connector counts and frequent insertion cycles. (indiumcorporation) AI servers and network equipment tend to be connector-dense: multiple high-speed board-to-board connectors, DIMM and PCIe card edge interfaces, mezzanine connectors, backplane connectors in modular systems, and high-port-count switch ASIC boards.
Silver appears in PCBs and assemblies through: 1) silver-bearing solders, 2) silver-filled conductive adhesives and epoxies in certain component attach contexts, and 3) silver plating in selected RF/high-frequency interfaces. The dominant solder family in lead-free electronics, SAC305, is specified as 96.5% tin, 3% silver, and 0.5% copper, which embeds silver into essentially every solder joint in assemblies using that alloy. (AIM Solder) The silver mass per server attributable to solder is therefore a function of the total solder mass used across motherboards, accelerator boards, power distribution boards, and power supply assemblies. Even if silver is only 3% by solder mass, high-component-count boards and large board populations across a hyperscale AI build can translate into meaningful aggregate silver demand.
MEMORY (DRAM, HBM MODULES, DIMMS) AND STORAGE (NAND/SSD)
Gold enters memory systems predominantly through connector plating and package interconnects. Server DIMMs (and some storage module form factors) use gold-plated edge contacts to maintain reliable low-resistance interfaces under repeated insertion and long service life. The use of gold in these contacts is consistent with the broader principle that gold is preferred for low-current signal interfaces where oxide formation would destabilize contact resistance. Memory IC packages may use gold or copper wire bonding depending on package type and cost/reliability constraints, contributing additional gold consumption at the semiconductor packaging level, though this is increasingly subject to substitution by copper wire in mainstream packages. (Kulicke & Soffa Industries, Inc.)
Silver enters memory and storage through solder, conductive adhesives, and some contact systems. High-speed memory and storage subsystems require signal integrity at multi-GHz rates; while precious metals are not directly used for bulk conductors (copper dominates), the contact and termination ecosystems still employ precious metal coatings to stabilize interfaces.
A key uncertainty for metal intensity in AI systems is the balance between HBM-on-package and off-package DRAM. Training-class AI nodes typically shift a portion of memory bandwidth demand onto HBM integrated with accelerators; however, CPU-side DRAM remains substantial for host processing and staging, and storage/network disaggregation strategies can shift where connector density appears (within servers vs within fabric-attached storage). Net precious metal demand therefore depends more on total connector count and board population than on the nominal “HBM vs DIMM” split.
NETWORKING (SWITCHES, NICS, OPTICS, AND CABLE PLANTS)
Generative AI clusters are unusually network-intensive relative to general-purpose compute. Scaling laws for distributed training and inference drive large, high-bisection fabrics (often multi-tier fat trees) with high port counts, large optical module populations, and dense cable plants. This tends to increase both gold and silver usage per unit of IT load through 3 pathways:
A) Gold in high-speed connectors and switch/line-card interfaces: High-speed electrical connectors rely on gold plating for stable contact resistance and wear performance. Switch line cards and NICs also use gold in PCB finishes and in edge connectors.
B) Silver in select high-frequency conductors and plated interfaces: At high frequencies, conductor surface properties matter due to skin effect. Many high-frequency coax and twinax constructions use silver-plated copper conductors to reduce RF losses and improve performance, implying that an expansion in high-speed direct-attach copper (DAC) and certain internal cabling can embed incremental silver. Product and materials references for coax/twinax commonly specify silver-plated copper conductors as a construction choice for high-frequency performance. (Netgate)
C) Gold in optical transceivers and photonics packaging: Optical modules contain lasers, photodiodes, driver/TIA ICs, and packaging that often uses gold wire bonds, gold metallization layers, and gold-tin solders for die attach and feedthroughs. Eutectic AuSn is explicitly described as a preferred option for high-reliability solder applications in optoelectronic packaging and laser diode die attach, with reference to uses including RF/DC feedthrough attach and optoelectronic packaging, and with a substantially higher melting point (280 °C) than Sn-Ag solders (≈221 °C). (Inseto UK) As AI fabrics migrate to higher speeds (400G to 800G and beyond), optical module counts can increase rapidly in large clusters, making optoelectronic gold usage a potentially material subcomponent of AI-linked gold demand within the broader electronics category.
DATA CENTRE ELECTRICAL INFRASTRUCTURE: WHERE SILVER DOMINATES
The portion of a data centre build that is not “IT hardware” increasingly becomes a primary driver of silver demand because AI data centres are power-dense and require extensive high-current distribution and protection. Silver plating and silver contact materials are widely used in electrical apparatus because they offer superior conductivity and can maintain performance over time under appropriate environmental control. A technical paper on degradation of power contacts in industrial atmospheres states that silver plating is widely used on contacts and other conductive parts in electrical apparatus such as switchgear and motor control centers, and explicitly enumerates silver present on bus, circuit breakers, protective relays, auxiliary relays, control switches, and test switches. (https://t.co/T0slRGmfMY) While that paper emphasizes corrosion risks in sulfur-rich environments, its system-level implication is that large electrical installations that prioritize reliability and low contact resistance frequently deploy silver-plated interfaces.
Within an AI data centre, the silver-bearing electrical infrastructure footprint spans:
Utility interconnect and substation gear (on-site or near-site): high-voltage switchgear, breakers, protective relays, and buswork.
Medium-voltage to low-voltage transformation and distribution: transformers, switchboards, busways, and panelboards; silver appears primarily in contact interfaces, relays, and protective devices.
UPS and power conditioning: rectifiers/inverters, static transfer switches, and internal contactors; silver appears in contacts, in some power module attachment methods (silver sinter), and in solder.
Rack-level power distribution: PDUs, busbars, and connectorized power shelves for GPU servers; higher rack power densities increase current levels and the stress on contact interfaces, pushing designs toward lower-resistance, higher-reliability contact systems where silver plating can be used to reduce resistive heating at joints.
A notable nuance is that silver usage in high-current gear is not only a function of current rating but also of reliability philosophy and maintenance intervals. Data centres typically operate in controlled atmospheres with filtration and humidity controls, which can reduce corrosion risks relative to industrial plants, supporting the use of silver-plated contacts without the same corrosion-driven failure incidence described for sulfur-rich environments. (https://t.co/T0slRGmfMY)
THERMAL MANAGEMENT, HVAC, AND COOLING INFRASTRUCTURE: SILVER VIA BRAZING AND JOINING
AI data centres increasingly require advanced cooling architectures (rear-door heat exchangers, direct-to-chip liquid cooling, immersion in some deployments), which expands the amount of heat-exchange equipment, copper tubing, manifolds, fittings, and refrigeration components per MW. Silver demand can arise in this subsystem through brazing filler metals used to join copper tubes and fittings. Copper Development Association guidance distinguishes brazing filler metals suitable for copper tube as including BCuP series alloys (phosphorus-containing) and BAg series alloys containing high silver content, indicating that silver-bearing brazes are a recognized and often-used approach in copper joining. (https://t.co/O6w8pFNMo4) The magnitude of silver embedded in cooling infrastructure is highly design-dependent (air-cooled vs liquid-cooled; on-site chilled water vs direct evaporative), and the silver content of brazing alloys varies widely by alloy family and performance requirements, creating a potentially non-trivial but uncertain silver lever in large-scale build-outs.
POWER GENERATION AND GRID EXPANSION: SILVER THROUGH SOLAR PV AND ELECTRIFICATION
The single largest structural linkage between AI-driven data centre growth and silver consumption is typically the incremental electricity supply build-out, with solar PV a major silver-intensive technology. The IEA projects that renewables meet nearly 50% of the additional electricity demand from data centres between 2024 and 2030 in its Base Case, with renewables generation for data centres increasing at an annual average rate of 22% between 2024 and 2030, driven primarily by wind and solar PV growth and partially financed through PPAs with technology companies. (IEA) This establishes a clear causal pathway: AI load growth supports additional solar PV deployment, and solar PV manufacturing consumes silver.
The Silver Institute’s supply and demand commentary reinforces the PV linkage, stating that record electronics and electrical demand reflected structural gains in the green economy flowing through from PV and automotive sectors and grid infrastructure development, while also noting that “notable advancements within the PV segment led to a sharp reduction in silver loadings.” (The Silver Institute) This combination is critical: AI can drive more PV deployment, but the silver-per-watt embedded in that PV can fall due to thrifting and substitution, weakening the linearity of the relationship.
A 2024 industry estimate cites photovoltaic silver demand at 6,577 tons in 2024, with solar described as accounting for about 19% of worldwide silver metal demand in that year. (The Silver Institute) In parallel, Silver Institute/Metals Focus commentary reports that China alone added 278 GW of solar capacity in 2024, illustrating the scale of the PV build cycle into which AI-driven incremental demand would be layered. (The Silver Institute) The implication is that AI-driven incremental PV demand must be evaluated against a backdrop where global PV additions are already measured in the 100s of GW per year; AI may be a marginal driver rather than the primary driver at the global level, but it can be locally decisive in constrained grids or in regions where data centre clusters are the dominant incremental load.
The grid build-out pathway is additive to PV. Even when electricity is not met by new PV, incremental load growth requires upgrades to transmission, distribution, protection, and switching infrastructure, all of which can embed silver through plated contacts, relays, and switchgear components. Silver Institute commentary explicitly links silver demand strength to “grid infrastructure development” alongside PV and AI-related applications, consistent with silver’s role in electrification hardware. (The Silver Institute)
A DATA-DRIVEN FRAMEWORK FOR TRANSLATING AI BUILD-OUTS INTO METAL CONSUMPTION
STARTING POINT: ELECTRICITY GROWTH AS A PHYSICAL PROXY
IEA projections imply global data centre electricity consumption rising from 460 TWh in 2024 to over 1,000 TWh in 2030 and 1,300 TWh in 2035. Converting to average continuous load yields ≈52.5 GW in 2024, ≈114.2 GW in 2030, and ≈148.4 GW in 2035, implying incremental average load of ≈61.6 GW by 2030 and ≈95.9 GW by 2035. (IEA) The IT hardware portion is lower than total electricity due to PUE and facility overhead; therefore, metal intensity per MW should be referenced either to total facility load (for power infrastructure and energy supply) or to IT load (for servers, accelerators, and networking).
DIRECT HARDWARE METAL INTENSITY: WHY ESTIMATES ARE WIDE
Precious metal intensity per server varies across orders of magnitude depending on system architecture, board types, and connector density. E-waste and PCB studies demonstrate wide dispersion in gold and silver content per kg of printed circuit boards. A 2022 Waste Management study reports gold content in waste printed circuit boards ranging from 179.86 mg/kg to 3,694.51 mg/kg and silver from 809 mg/kg to 12,320.51 mg/kg, indicating that board grade and component mix can swing precious metal concentration by >10x. (ScienceDirect) A Chemical & Engineering News article provides a lower-bound illustrative point, stating that 1 t of PCBs contains at least 0.4 kg of silver and 0.09 kg of gold, underscoring that “average PCB” mixes can dilute precious metal concentration depending on what is included in the category. (Chemical & Engineering News)
AI servers and network equipment tend to skew toward higher-grade boards and higher connector density than consumer electronics, which tends to bias toward the upper half of concentration ranges. However, multiple counterforces push toward lower metal intensity: thinner gold plating, substitution to palladium/nickel or tin finishes in non-critical contacts, and copper wire bonding substitution for gold in many IC packages. (https://t.co/sRg1uEcyy2)
BOTTOM-UP MECHANISMS THAT SPECIFICALLY SCALE WITH GENERATIVE AI
A) Accelerator density and advanced packaging: AI nodes pack large numbers of accelerators per chassis, increasing the count of high-end boards, connectors, and power stages per server. Even if flip-chip reduces gold wire bond usage on the main die, the total ecosystem of auxiliary ICs and interconnects scales up.
B) Fabric scaling: Large training clusters scale port counts superlinearly with node counts in many network topologies, increasing the number of optical modules, connectors, and cables. This can raise gold usage (connector plating, optics packaging) and silver usage (silver-plated conductors in some cable types, solder, and contact materials).
C) Power density: Rising rack power density increases the amount and rating of switchgear, busway, breakers, and UPS infrastructure per unit floor area, embedding more silver-bearing contacts and potentially more silver-sinter usage in power modules. (Power Electronics News)
