Open App →
Back to Blog
October 5, 2026 By Steve

Nobody Drills for Helium

Rare earths get summits and transformers get hearings. Helium has no lobby, no strategic reserve and no supply curve that answers to price. That's why the shortage hitting the AI buildout's memory rung is barely on the board.

The constraint nobody's watching

Most of the AI buildout's constraints have an audience. Rare-earth magnets have a truce deadline everyone's counting down to. Export controls get debated every week. Transformer lead times and grid queues have become their own beat.

Helium has almost none of that. There's no trade negotiation over it, no national champion building a mine-to-market chain for it, and since January 2024 no U.S. strategic reserve either. The Federal Helium Reserve was sold to Messer, and the USGS now lists the government stockpile as none.

That would matter less if helium were easy to substitute. It isn't. In a fab it cools wafers, provides an inert atmosphere and finds leaks. In a hard drive it fills the sealed enclosure that lets high-capacity drives exist at all. In a hospital it keeps MRI magnets superconducting. None of those uses has a drop-in replacement.

What actually happened

The shock came in two steps. Iranian drone strikes on Qatar's Ras Laffan complex on March 2 shut LNG production, and QatarEnergy declared force majeure two days later. Missile strikes on March 18 and 19 then did physical damage. In September, QatarEnergy's CEO said the two damaged liquefaction trains will take three years to repair.

Because Qatar's helium comes out of its LNG processing, helium stopped when the gas did. Qatar produced about 63 million cubic meters of helium in 2025, roughly a third of the world's 190 million.

The flattered number

"A third of the world's helium, gone" was the headline. It was true for a few weeks, and it's the wrong number to anchor on.

Production has partly restarted. The lasting loss is the damaged trains, and one consultant put it at about 30 percent of Qatar's volume in 2026, or roughly 11 percent of global supply. Other estimates of the structural cut land around 14 percent of Qatar's capacity. So the headline overstated the size of the hit.

It also understated the length. A one-third outage for a month is a scare. An 11 percent hole for three years is a regime. Markets handle the first with inventory. The second needs new supply, and new supply is the one thing helium can't produce on demand.

There's a second trap in the coverage. Hard drive prices rose about 46 percent between September 2025 and January 2026, and that gets cited as helium damage. It can't be. Those increases came before the first strike. They were AI storage demand. Helium is now stacking on top of a market that was already tight, which is a different story and a worse one.

Why price can't fix it

In most commodity squeezes, high prices pull in new supply. Helium mostly doesn't work that way.

Nobody drills for helium. It's recovered as a byproduct of natural gas processing, so supply follows gas economics. Helium is a rounding error in an LNG train's revenue. Doubling the helium price doesn't restart a liquefaction train, and it doesn't speed up a repair.

The alternatives are slow. Greenfield helium projects in Tanzania, Saskatchewan, Montana and Colorado run seven to ten years from exploration to commercial volume, by one industry estimate. Russia's Amur plant was designed for about 84.5 million cubic meters a year and has delivered a fraction of that after fires and sanctions.

One detail is worth watching. Fusion Worldwide reports that Qatar's repairs are paced partly by multi-year lead times on specialized gas turbines. If that holds, the helium outage is waiting in the same equipment queue that behind-the-meter data centers created. The AI buildout's power rung would be lengthening the shortage on its memory rung.

The buffers were sized in months

When the strikes hit, the major chipmakers said they were covered. TSMC reported more than two months of stock from multiple suppliers. Samsung and SK hynix were reported to hold about six months.

Those were sensible buffers for a disruption measured in weeks. They're the wrong size for one measured in years. Inventory buys time to find another source. It doesn't create one, and the other sources are already spoken for.

Where it lands first

Exposure isn't even, and the pattern matters more than the total.

South Korea sourced 64.7 percent of its helium from Qatar in 2025, according to TrendForce. Taiwan's mix is closer to 30 percent Qatar, 30 percent U.S., with the rest from other countries and domestic sources. That geography routes the shock straight to Korea's memory fabs, which make most of the world's HBM. The rung that was already the tightest in the AI supply chain is the one most exposed to helium.

From the fab side, a helium squeeze rarely stops a line. It erodes it. Thermal control gets less precise, defects creep up, yield slips and cost per good die rises. You won't see it as an outage. You'll see it in margins and in how many good parts come off the same wafer starts.

Then there's allocation. In a shortage, gas suppliers protect their largest contracts. That means the first buyers actually rationed are usually the smallest: university labs, research cryogenics and smaller hospitals running older MRI magnets. They use far less helium than a fab, but they have the least leverage. In past shortages they were rationed first, and there's no reason to expect this one to be different.

The least exposed are the operators who stopped buying helium as a consumable. Over 70 percent of Japanese and Taiwanese fabs run closed-loop recovery that captures 90 to 95 percent of their helium. TSMC is in that group.

The only new supply is the helium you already bought

That points to the one lever that works inside the shortage window: reclaim.

Samsung's Helium Reuse System, running on select lines since 2025, cuts about 4.7 tonnes a year, with roughly an 18.6 percent reduction projected if it's expanded company-wide. Retrofits reportedly run $300,000 to $1 million and take 12 to 18 months.

Those are small numbers against a fab's capex budget and a big one against the alternative. Reclaim doesn't add a molecule to world supply. It does take a buyer off the spot market, which is the next best thing. For the equipment side of the industry, it's also a real demand signal: reclaim and purification systems are about to become a line item nobody budgeted for.

You don't remove a bottleneck, you relocate it

Every constraint in this buildout has eventually moved somewhere less visible. Capital moved into structured credit. Silicon moved into packaging. Power moved into turbines, transformers and permits.

Helium is a constraint that skipped the move. It went straight to the rung that was already tightest, memory, and hit hardest in the country with the most concentrated exposure. It did it without a policy fight, a press conference or a deadline anyone circled.

That's what makes it under-covered. Visible constraints get managed. Quiet ones get absorbed until they show up in a yield number or a price sheet, and by then they're usually explained as something else.

What I'd watch

If you want to know whether helium is binding or just hurting, these are the dials:

The question

The buildout has spent two years learning that its limits aren't where it expected them. Most of them announced themselves.

This one is a byproduct of someone else's business, priced by someone else's market and repaired on someone else's schedule. So how many other inputs does the AI supply chain depend on that nobody drills for?