The Silver Bottleneck

A bottom-up model of the global silver market, 2020–2035. Every figure is classed OBSERVED / DERIVED / ESTIMATED / SPECULATIVE and traces to a published evidence ledger and model code. Written August 22, 2026.

Prologue: ninety days that tested every theory

On October 14, 2025, a former JPMorgan metals trader told reporters there was “no free-floating silver left” in London. He was very nearly right. One-month lease rates — the cost of borrowing physical silver, normally 1% — went above 30% annualized; overnight rates touched roughly 200%. Metal was being flown, by commercial airliner, from London to Mumbai and from New York to London. Silver broke a 45-year-old nominal record at $54.48 on October 17. By January 29, 2026, it printed $121.67 intraday. Four sessions later it had fallen 30% — the second-steepest two-day drop in the metal’s recorded history, exceeded only by the collapse of the Hunt corner in 1980 — after CME raised margins twice in a week. By August 2026 the futures curve was back in full contango, lease rates were back at 2–3%, London vaults held more silver than at any time since 2021, and the price sat near $69: forty-nine percent below the spike, seventy-two percent above the prior year’s average.

Every claim anyone wants to make about silver — bull or bear — was stress-tested in those months. This paper is an attempt to build the machine that explains them, and to project it forward honestly to 2035. We built the model to falsify a specific thesis: that AI, electrification, solar and investment demand are colliding with an inelastic supply base to create persistent physical deficits much larger than conventional forecasts anticipate. The short answer, which the rest of this paper earns: the thesis fails as stated, and something more precise — and stranger — survives in its place.

Part I — Where silver comes from

The world mined 846.6 Moz of silver in 2025 (Metals Focus / Silver Institute, World Silver Survey 2026; all balance figures use its April 2026 vintage). That was a record, up 3%. But the composition matters more than the total: 73.9% of mine supply is a by-product of lead-zinc, copper and gold mines — decisions made about other metals. Primary silver mines are down to 26.1% of supply, a record low, and falling.

The subtlety usually missed: roughly half of that by-product stream is further encumbered by streaming contracts under which the operator receives $4–6/oz and 20–25% payability — the mine operator has no economic exposure to silver at all. The silver price signal reaches, at most, a quarter of the supply base. And that quarter is eroding: primary production fell for a third consecutive year in 2025, and primary-mine reserves declined in the best price year in history — a replacement ratio under 75%, flattered by cut-off-price arithmetic.

What did miners do with a 70% AISC margin? $10.8 billion of M&A (more than 2015–2024 combined), record buybacks, and the largest producer hedge book since 2013 — selling the price, not building mines. The greenfield queue nets out, after historical slippage, to roughly +30–40 Moz/yr by 2030 against 15–25 Moz of base decline. The historical calibration is unforgiving: after the 2011 peak, supply kept rising for five years while price fell 68% — the supply lag is four to five years in both directions. Supply is not collapsing, but it cannot accelerate. Whatever happens to silver will be decided on the demand side and in the vaults.

Recycling supplied 197.6 Moz in 2025 — a 12-year high, but up only 2% against a 42% price rise, because refinery throughput, not the scrap pool, was the binding constraint (Metals Focus’s own finding). Our recycling curve: ~211 Moz at $75, ~230 at $100, ~255 at $150 — with 6–12 month lags and a hard short-run growth cap.

Part II — Where silver goes

The AI number, built honestly

There is no published figure for silver per AI server. None. The Silver Institute’s own commissioned AI study says, verbatim: “Even in the absence of precise silver loading data, the link is clear.” Circulating claims of “1–2 oz per server” are unsourced. So we triangulated three ways — rack mass × e-waste assay, component build-up (SAC305 solder is 3.0% silver; signal connectors are gold-flashed; MLCCs de-silvered decades ago), and top-down allocation — and reconciled to 1.5–22 tonnes of silver per GW of data-center IT load, mid ~5.

AI and data centers consume roughly 0.5–7 Moz of silver per year — 0.1 to 1.1% of industrial demand. The 2030 high case is ~21 Moz, about 3%.

The AI electricity channel is bigger than the AI electronics channel — utility-grid equipment runs ~22 Moz/yr rising toward ~27 by 2030 on the IEA’s grid-capex path — but even AI-direct-plus-AI-induced-grid is under 20 Moz/yr of increment by 2030 in our bull case. Real, growing, and an order of magnitude too small to drive the balance. Robotics is smaller still: a humanoid carries perhaps 10–30 g of silver; at 25 g, the world would need 8.7 million humanoids per year to consume 1% of industrial demand. Actual 2025 production was about twelve thousand.

Solar: the engine that just shifted into reverse

Photovoltaics was the genuine demand revolution: 82.8 Moz in 2020 to 197.5 in 2024 — 29% of industrial demand. It has rolled over. When silver crossed $84 in December 2025 and $118 in January, silver paste hit 10–20% of cell cost, and within weeks Longi announced mass production of copper-metallized back-contact cells, high-copper TOPCon paste entered GW-scale production, and zero-busbar designs went mainstream. Blended intensity fell 20% in one year. The result: PV silver demand fell 6% in 2025 despite record installs, and is forecast down 19% in 2026 on installs down only 8% — the year volume growth lost the race to thrifting.

Our three intensity trajectories yield the finding that matters: at base-case thrifting, PV silver demand declines through 2030 at any install level below ~1,100 GW/yr — far above every deployment forecast. The burden of proof has moved to the bulls (specifically, to copper-reliability failure: Ag–Cu interdiffusion and TOPCon UV degradation are live research controversies).

Electrification, honestly bounded — and one wildcard

Automotive silver (~73 Moz in 2024 → ~94 by 2031) is solid but already inside the electronics numbers, not additive. Chargers add single-digit Moz. And one wildcard is not small: silver-carbon anodes in solid-state batteries — roughly 1 kg of silver per 100 kWh in the Samsung SDI architecture, 15–30× today’s per-vehicle content, mass-production target 2027. Two million such EVs would consume ~48 Moz/yr — the only candidate we found for a brand-new demand category of solar’s scale. We carry it at 30% probability; it is the tail that breaks the bear case.

Part III — The balance, assembled and stress-tested

One definitional trap governs the whole debate: the World Silver Survey “market balance” excludes ETF flows. The famous headline — five consecutive deficits, −716 Moz cumulative 2021–25 — understates the true call on above-ground stocks. Counting ETPs, the drawdown was −1,172 Moz over five years, including −318 Moz in 2025 alone.

Our model pins 2020–25 to the survey actuals (the identity check reproduces every reported balance to ±0.1 Moz) and builds 2026–35 bottom-up. Four scenarios:

At 2030: bear +208 Moz surplus, base +18 (balanced), bull −149, physical squeeze −288 with float exhaustion from 2028. The Monte Carlo (20,000 draws) puts the 2030 headline balance at median +5 Moz — an honest coin flip — but the balance including investment flows at median −46 Moz/yr with a 75% probability of continued float drain.

And the sensitivity tornado answers the question this project was built to answer:

The five variables that actually determine the silver deficit: investment demand (±115 Moz — nothing else is close), PV silver intensity (±51), mine base-decline (39), PV volume (35), and price-driven demand destruction and recycling (34/28). AI-direct ranks thirteenth of fifteen. The silver story is an investment-flows × thrifting × by-product-decay story. It is not an AI story.

Part IV — When does a deficit become a shortage?

A deficit is an accounting entry; a shortage is an event in a vault. What matters is the mobilizable float: London metal not allocated to ETFs, plus COMEX registered, plus what Shanghai and Singapore will release. In September 2025 that number was ~136 Moz — against ~450 Moz of daily London trading volume. The squeeze followed with mechanical inevitability.

Two lessons, both reproduced by our float model, both cutting against naive narratives. For the bears: headline stocks were never low; the shortage happened anyway, because fragmentation is real — bar sizes, purity tiers (99.9% London vs 99.99% industrial — the entire reason a four-nines Singapore contract launched in May 2026), tariff walls, India’s new 15% duty, China’s export licensing. Metal in the wrong vault, bar, purity or jurisdiction does not exist for delivery this month, and price is set at that margin. For the bulls: the float regenerates at price. Our model cannot reproduce the observed 2026 rebuild without assuming 150–250 Moz of episodic disgorgement from pools the squeeze narrative labels unavailable — and retail net-sold the January top. The “silver is gone” story and the “there’s plenty of silver” story are both true, at different prices, roughly one hundred dollars apart.

Part V — Price regimes, stated as conditions

Price follows the cumulative draw on mobilizable float with a 3–4 year lag — the biggest annual deficit (2022) coincided with the lowest price of the deficit era; the explosion came in year five.

• $50–75 (current regime): balanced-to-modest deficits, float 250–350 Moz, contango. Sustainable indefinitely absent an investment surge.
• $100: requires no shortage — only macro (weak dollar, ETP inflows >100 Moz/yr, gold >$5,000). January 2026 proved it.
• $150: requires renewed float stress on top of the macro — registered under ~80 Moz, lease rates >10%.
• $200+ ($212 is the CPI-adjusted 1980 peak): corner-equivalent conditions. History’s warning: no such price has ever held for more than weeks, because margin committees, governments and a hundred million ounces of disgorgement are all reaction functions with hair triggers.

Verdict

The popular thesis — AI eats the silver market — is wrong, and this paper documents exactly how wrong. The popular anti-thesis — deficits are fake, stocks are ample — mistakes accounting for availability and was disproven in real time in October 2025. What survives is a market with a new microstructure: a supply base that cannot accelerate, a demand base whose industrial core is self-stabilizing through engineering, and a shrunken, fragmented, policy-walled free float that converts ordinary investment-demand surges into extraordinary price events. Silver has become a metal in which the second derivative of investor behavior matters more than the first derivative of any factory’s consumption.

That is not a supercycle. It is a fragility regime — and fragility, unlike scarcity, is bidirectional.

Methods: model code with programmatically checked unit conversions; history validated to ±0.1 Moz against WSS 2026; N=20,000 Monte Carlo; float model calibrated to reproduce the October 2025 squeeze. Principal data gaps, disclosed: no public AI-rack teardown assay exists; grid-equipment contact loadings are engineering estimates; the WSS sub-application table is paywalled. The full evidence ledger, ten research briefs, and model download will be published on this page.