"The AI boom runs on tungsten — and the world is running out"

Ask most people what minerals power artificial intelligence and you'll hear the usual suspects: copper for data-centre wiring, lithium and cobalt for batteries, maybe rare earths for magnets. Tungsten almost never makes the list — which is strange, because it's quietly inside every one of the advanced chips doing the actual computing.

Tungsten is one of the more remarkable elements on the periodic table. It has the highest melting point of any metal — a blistering 3,422°C — which is why it spent decades as the filament inside incandescent light bulbs and still turns up in the electrodes of TIG welders and the heat shields of spacecraft. In chipmaking it does something subtler but just as essential: its ultra-fine, low-resistance deposits fill the microscopic "vias" that connect the stacked layers of a modern processor.

A recent report from Almonty Industries CEO Lewis Black ties those two worlds together with a warning. The AI boom, he argues, is consuming tungsten faster than the world can replace it, and global reserves are "running on empty." The specifics are stark: an American strategic-stockpile order that went unanswered, two factory shutdowns, and an uncomfortable truth about how few new tungsten projects will ever produce a single tonne.

The root of the problem is concentration. China controls roughly 80% of global tungsten production and has periodically tightened export controls on a metal it clearly regards as strategic. The US Geological Survey's tungsten statistics tell the same story — a market where the Western world's alternative sources, Vietnam, Russia, and a scattering of smaller producers, cannot scale quickly enough to offset the dominant supplier.

Here's the first insight the headline misses: this is less a geology problem than a lead-time problem. Tungsten mines are notoriously slow to permit and build, with the span from discovery to first production routinely running a decade or more. That means a demand spike that began when AI accelerators went mainstream a couple of years ago cannot be answered by new supply until well into the 2030s. No amount of capital shortens a mine's permitting timeline.

The second insight is about where AI's appetite actually bites. We narrate the boom through its glamorous inputs — GPUs, high-bandwidth memory, racks of servers — but the constraints that really bind tend to sit a layer down the stack. A couple of weeks ago I wrote about how AI data centres were draining the world's supply of multilayer ceramic capacitors, the "rice of electronics." Tungsten is the same story one level lower still: a metal almost nobody thinks about, caught in the same gravitational pull.

There's a specific reason AI hardware is unusually tungsten-hungry. The advanced logic chips inside AI accelerators carry more interconnect layers and finer, denser vias than older designs, and tungsten is the workhorse metal for filling those contacts. As leading-edge fabs push toward ever-smaller nodes, the volume of tungsten per wafer tends to climb even as individual features shrink — because there are simply more of them, each one needing its own tiny plug of metal.

The "stockpile order nobody can fill" is a revealing detail in its own right. Strategic reserves exist precisely for this scenario — a supply disruption you can buy your way out of. But a reserve only works if some seller, somewhere, has metal to spare. When a government issues a tender and no one responds, the buffer turns out to be theoretical. It's a quiet signal that the physical market is tighter than the official inventories imply.

The "awkward truth" about new projects is worth unpacking too. Tungsten is hard to mine profitably: grades are low, the chemistry is intensive, and the ore often sits in remote or politically sensitive places. Black's own company is among the few with a credible path — its Sangdong mine in South Korea and a development project in Spain — which is itself a measure of how thin the field of viable new sources really is.

A small aside for the curious: tungsten's chemical symbol is W, from wolfram, an old German name meaning roughly "wolf's froth." Tin miners coined it because the troublesome mineral seemed to devour tin ore the way a wolf devours sheep. It has been an annoyance and a prize in roughly equal measure ever since — a fitting origin story for a metal whose modern value is easy to overlook.

None of this is a reason for panic. High prices have a reliable way of summoning investment, and tungsten is a strong candidate for substitution and recycling — tungsten carbide is already among the most recycled industrial materials on earth. The likelier path is a grinding stretch of higher prices and supply anxiety, followed by a slow rebalancing as projects that looked marginal at the old prices suddenly start to pencil out.

What the tungsten squeeze really illustrates is that AI's material footprint reaches far beyond the data centre. Every watt, every chip, every circuit board reaches down into a web of obscure minerals that the industry — and the rest of us — have long taken for granted. As the boom matures, it may be the metals nobody names that quietly set the pace.