AI Is Reshaping Electronics Gold Demand, Not Expanding It

Electronics gold demand in 2025 landed at 270.4 tonnes. In 2024, it was 270.8 tonnes. The number barely moved.

And yet, inside that flat aggregate, something structural is happening that the World Gold Council and major capital markets institutions say could define this demand category for the next decade.

The composition of electronics gold use is rotating away from consumer smartphones, which face a record-setting volume decline in 2026, and toward AI server infrastructure, which is absorbing hundreds of billions of dollars in annual capital expenditure. That shift is visible in the supply chain geography and the hardware mix, even if it does not yet register as a meaningful change in total tonnage.

What this analysis maps is where the gold is actually going, how thrifting and substitution are capping any volume uplift from AI, and what the compositional shift genuinely means for your framework as an investor, as opposed to what the “AI gold demand” narrative tends to imply. This is a story about structure, not surge.

Flat on the surface, shifting underneath

Start with the surface fact that any credible read of AI gold demand has to work through. Electronics gold consumption came in at 270.8 tonnes in 2024 and 270.4 tonnes in 2025, per the World Gold Council’s Gold Demand Trends: Full Year 2025, published 29 January 2026. Two years, essentially one number.

The WGC technology demand data covering 2024 and 2025 shows electronics consumption holding at 270.8 and 270.4 tonnes respectively, with AI-related server applications identified as the primary offset against declining consumer electronics volumes.

The most recent quarterly data extends the pattern. Electronics gold demand rose 3% year-on-year in Q1 2026 to 69 tonnes, up from 67.1 tonnes in Q1 2025, according to the WGC’s Q1 2026 technology report published 29 April 2026. Modest positive drift, not a break higher.

The wider technology grouping, which folds in industrial and dental applications, reached 322.8 tonnes in 2025, roughly 6% of total global gold demand exceeding 5,000 tonnes. The WGC’s own characterisation of the category sets the tone.

The broader gold demand picture in Q3 2025 reached a historic 1,313 tonnes across all categories, a figure that contextualises why a 0.4 tonne decline in full-year electronics consumption carries so little weight for price discovery relative to the movements in investment and central bank buying.

WGC framing: Technology-sector gold demand held “broadly steady” across 2024 and 2025.

That near-identical aggregate across two years tells you something specific: AI server demand and smartphone contraction are, at this stage, roughly cancelling each other out within electronics. Understanding that offsetting dynamic is the first step toward reading any future tonnage data correctly, because the interesting story is about what the number is made of, not the number itself.

Where smartphone decline enters the equation

The other side of the offset is the consumer handset. IDC, in a September 2026 revision, now expects worldwide smartphone shipments to fall 16.7% in 2026 to just over 1 billion units, a record contraction. That revision worsened an earlier May 2026 estimate of a 13.9% decline. Notably, total market value is still expected to grow 6.3% to $613 billion on higher average selling prices.

Counterpoint Research, cited by CNBC on 27 February 2026, projected a 12% shipment decline for 2026, the sharpest on record and the lowest volumes since 2013.

Falling smartphone unit volumes reduce one historically significant source of electronics gold demand. That contraction creates the space for AI server demand to grow its share of a fixed pool without moving the aggregate figure.

What AI hardware actually does with gold

If smartphones are ceding ground, the natural question is what the incoming demand is physically doing with the metal. The WGC identifies bonding wire and high-density interconnect boards installed in AI servers as the primary gold-use channel within the electronics segment, and its Q4 2025 report flagged demand rotating toward the server infrastructure that supports AI applications.

Gold enters AI server and accelerator hardware across four functional categories:

• Bonding wires and pads: high-performance CPUs, GPUs, and AI accelerators use gold bond wires or gold-plated pads in packaging families where fine pitches and high reliability are required.
• High-density edge contacts and interconnects: gold plating on edge connectors, sockets, and high-speed backplane contacts keeps resistance low and connections stable under dense compute loads.
• Advanced packaging layers: thin gold layers or gold-containing alloys appear in under-bump metallisation, micro-bumps, and fine redistribution layers in high-end accelerator and networking modules.
• Specialised optics and RF modules: certain optics, radio-frequency modules, and precision sensors use gold for thermal stability, corrosion resistance, and bondability.

AI servers and accelerator cards typically pack more chips and more high-density interconnects per rack unit than conventional general-purpose servers. That can lift the absolute gold content per server in some configurations, making AI-specific hardware somewhat more gold-intensive per unit than older designs.

The countervailing force is thrifting. Finer bond wires, thinner plating, and substitution in lower-criticality positions mean per-unit gold intensity does not scale linearly with chip count or compute density.

Gold’s physical properties, particularly its oxidation resistance, bondability, and performance at fine pitches, are not engineering preferences but hard constraints, which is precisely why the substitution frontier in high-reliability interconnects moves slowly even under sustained cost pressure.

The scale of the hardware being built is genuinely large. Goldman Sachs, in its 1 May 2026 analysis, put baseline AI capex at approximately $765 billion for 2026, covering compute, data centres, and power.

The capex backdrop: Goldman Sachs estimates roughly $765 billion in annual AI capex for 2026. That figure sits close to BloombergNEF’s estimate of near $750 billion for the 14 largest data centre operators and J.P. Morgan’s $697 billion for the five largest U.S. hyperscalers, up $173 billion from beginning-of-year projections.

Here is the signal worth holding onto. No published source, including the WGC, breaks out AI-server-specific gold tonnage as a distinct figure. That absence, set against hundreds of billions in hardware capex, tells you thrifting is already working hard enough to absorb much of the per-unit gold intensity that volume would otherwise deliver. Treating “AI uses gold” as a uniform or scalable proposition misreads the physics.

Why thrifting means AI volume does not equal AI gold demand

Thrifting is not a risk to the AI gold story. It is the mechanism that explains the flat aggregate you already encountered. Two levers do the work: applying thinner gold layers to hit the same functional outcome (thrifting proper), and replacing gold with palladium-coated copper, nickel, or advanced solder alloys where gold’s specific properties are not strictly required (substitution).

Four drivers are accelerating both, ordered here by weight of impact:

1. Persistent cost pressure at scale: across hundreds of millions of units, even small reductions in gold per connector or contact compound into significant savings.
2. Elevated and volatile gold prices: high prices give engineering teams a direct incentive to redesign interconnects and qualify cheaper alternatives.
3. Miniaturisation and advanced packaging: AI accelerators and high-end processors use fine-pitch interconnects and advanced substrates, enabling more precise deposition and the same electrical performance with less gold per connection.
4. Improving reliability of substitutes: copper and aluminium bond wires, once vulnerable to heat and corrosion, have improved enough to expand where gold is no longer strictly necessary.

The parallel from an adjacent sector is instructive. Solar panel manufacturers reduced silver content per cell through thrifting rather than eliminating the metal, showing how technology-critical metals behave under sustained cost pressure. The WGC’s stable technology figures across 2024 and 2025 are the empirical fingerprint of the same process at work in electronics.

Where gold remains effectively irreplaceable

Thrifting has hard limits, and they matter for how durable the floor is.

• Corrosion resistance: gold’s oxidation resistance keeps connections stable over long lifetimes in demanding sockets, edge connectors, and relay contacts, where substitutes show higher failure rates.
• Bondability and process robustness: gold’s wide process window tolerates minor manufacturing variation; alternatives often demand tighter control and specialised atmospheres.
• High-frequency, high-density performance: at extreme data rates and miniaturisation, gold’s mix of conductivity and fine-pitch compatibility is difficult to replace without hurting yield.
• Qualification cycles: once a gold-based interconnect is qualified in servers, networking gear, automotive, or aerospace, switching to alternatives requires lengthy reliability retesting, slowing substitution even where it is technically feasible.

The important point for positioning is that AI hardware concentrates its gold use precisely in high-reliability, high-density interconnects, the category where substitution risk is lowest. That makes the technology floor more durable than aggregate thrifting trends alone would suggest.

For you as an investor, the implication is direct. Even if AI server unit volumes double over the next three years, the realistic expectation for electronics gold demand is continued gradual drift rather than a step-change, because thrifting and qualification inertia are structural, not cyclical. Technology demand is a floor, not an amplifier.

How to read technology demand when you are a gold investor

The analytical layers now click into a single investor read, and it starts with proportion. At roughly 6% of total gold consumption, technology demand is dwarfed by the combined weight of investment flows, central bank purchases, and jewellery buying.

The proportional anchor: Technology represents approximately 6% of total gold demand, which exceeded 5,000 tonnes in 2025. A structural shift inside that 6% does not carry the price implications of a shift in central bank buying.

Electronics demand behaves as a non-monetary consumption floor. When prices rise, semiconductor manufacturers cannot simply remove the gold plating from their components; the most they can do is engineer thinner and thinner applications of it. That makes the category relatively inelastic, stable rather than price-amplifying.

The proportional weight of technology demand, at roughly 6% of total consumption, is why central bank reserve activity and real interest rates rank as the dominant gold price drivers, while electronics tonnage data informs demand structure without reliably signalling price direction.

The WGC’s geographic framing reinforces the point. East Asia’s electronics gold demand is supported by a strong AI supply chain and localisation strategies, while manufacturing weakness in Western markets drags on overall performance. AI is reshaping where demand originates without moving total tonnage, which is exactly why the aggregate held at 270.8 then 270.4 tonnes across the two years.

The scale of the buildout is worth seeing side by side, because the magnitude is what tempts the surge narrative.

The clearest tell sits inside that table. PwC models US$31.6 trillion of cumulative data centre capex through 2050, and none of the major institutional analyses from Goldman Sachs, J.P. Morgan, PwC, or BloombergNEF flags gold as a strategic bottleneck or central driver. Institutions modelling trillions do not treat gold as a supply constraint, which is itself the strongest evidence that electronics demand is a baseline input, not a catalyst.

The misinterpretation risk is confusing a sectoral compositional shift, AI hardware replacing smartphones, with total demand growth, and then positioning speculatively on tonnage data that does not support it. To avoid it, keep your monitoring priorities ranked correctly:

• Price-moving variables to watch: central bank reserve activity and real interest rate movements.
• Informative but not price-moving: AI capex headlines and electronics tonnage, useful for reading demand structure but not gold price direction.

Position technology demand as a structural floor rather than an AI-driven catalyst, and you sidestep the trap that “AI uses gold” implies a re-rating of gold’s price equilibrium.

What the compositional shift changes, and what it does not

The genuine insight the data builds toward is narrow and specific: AI server infrastructure is becoming the primary growth engine within electronics gold demand, but thrifting and the sheer weight of other demand categories make it a structural stabiliser, not a price catalyst. AI is redrawing the internal map of electronics gold use without expanding its territory.

The most recent data point makes the case in a single figure. Electronics demand rose 3% year-on-year to 69 tonnes in Q1 2026, per the WGC. A 3% quarterly rise, set against hundreds of billions in AI hardware capex, is the clearest quantitative statement that thrifting is doing its job, and that the current narrative is unlikely to translate into a demand-side price event.

There is genuine forward uncertainty worth naming. If AI hardware volumes scale far faster than thrifting can offset, a modest uplift in electronics tonnage is plausible, but it would need to be large relative to the 6% share to move price.

The illiquidity embedded in the tradeable gold supply structure means that even a modest genuine uplift in electronics consumption, if it ever materialised at scale from AI hardware, would interact with a market where available float is far more constrained than total above-ground stock implies.

The long-run backdrop: Goldman Sachs projects roughly $7.6 trillion in cumulative AI-related capex between 2026 and 2031. No institutional analysis in this body of research frames gold as a strategic constraint within that spend.

The one variable to track is whether thrifting rates in AI-specific hardware accelerate or plateau, since that determines whether the electronics floor creeps higher or holds flat across the next investment cycle. The absence of any published AI-server-specific gold tonnage figure remains a structural data gap that makes precise forecasting difficult and caution around AI gold demand narratives warranted.

This article is for informational purposes only and should not be considered financial advice. Investors should conduct their own research and consult with financial professionals before making investment decisions. Past performance does not guarantee future results, and financial projections are subject to market conditions and various risk factors.