48 Days. That's How Long Before the Helium Runs Out for AI Chips.

Study Guide

Study Guide: Helium Supply Chain Crisis and AI Chips

Core Thesis

The missile strike on Qatar's Ras Laffan refinery complex has taken offline roughly 33% of the world's helium supply. Helium is an irreplaceable input in advanced semiconductor manufacturing, meaning this disruption threatens the entire AI hardware pipeline, from HBM memory chips to logic processors, with cost and availability impacts lasting years.

Key Concepts

Why Helium Matters for Chips

  • Plasma etching cooling: Helium is blown over wafer backs during etching to maintain temperature uniformity and prevent warping.
  • Vacuum chamber leak detection: Helium's tiny atomic diameter makes it the only practical element for testing seal integrity in fab clean rooms.
  • No substitute exists: No other element can replicate helium's thermal conductivity or particulate size at fabrication scale.
  • Purity requirement: Fabs require 6N purity (99.9999%), which only a handful of production sites worldwide can achieve.
  • Consumption is increasing: More advanced chips consume more helium per unit. A 300mm EUV fab may use 5,000 to 20,000 cubic meters per month.

Three Channels of Impact

  1. Helium supply to chip fabs: Direct physical input required for all advanced semiconductor manufacturing. South Korea imported two-thirds of its helium from Qatar.
  2. LNG prices and energy costs: Helium is a byproduct of LNG cryogenic distillation. LNG disruption raises energy costs for East Asian fabs. Taiwan (TSMC) imports 97% of its energy and holds only 11 days of gas reserves.
  3. Geopolitical restructuring: China is developing domestic helium production and pursuing the Power of Siberia 2 pipeline from Russia, potentially giving it a long-term cost advantage in chip fabrication over the Western-aligned semiconductor system.

Timeline and Scale

  • Qatar produced ~2.4 billion standard cubic feet of helium per year (~33% of global supply).
  • 14% of Qatar's helium capacity is permanently damaged with a 5-year reconstruction timeline.
  • A planned new helium plant (Helium 4) targeting 1.5 billion cubic feet/year is now delayed past 2027.
  • Liquid helium in ISO containers vaporizes after 35 to 48 days, so stranded shipments are time-limited.
  • The US Federal Helium Reserve ended sales in 2023, removing additional global supply.

Who Is Exposed

  • SK Hynix and Samsung: World's two largest memory chip manufacturers, heavily dependent on Qatari helium for HBM production.
  • TSMC: Makes logic chips in Taiwan, exposed on both helium and LNG/energy fronts.
  • US data centers: Physically insulated but chip-dependent on Taiwan and South Korea. The Arizona fab covers only a tiny fraction of demand.
  • European data centers: Vulnerable to LNG price increases with limited domestic energy alternatives.
  • Consumers: Memory costs (DRAM already up 70%), laptop and phone prices will reflect these supply constraints.

Practical Takeaways

  • Memory costs are projected to stay elevated at least through mid-2027.
  • Energy costs for Asian fabrication will be passed through to cost-per-chip and cost-per-flop.
  • If you are in IT procurement, buying compute sooner rather than later is advised, as structural costs will ratchet upward.
  • The fabs will not shut down entirely, but reduced capacity and slower production will constrain the scaling that AI inference demand requires.
  • Only ~0.1% of the world currently uses heavy agentic AI workflows, yet capacity is already maxed out. Projected 10x to 100x demand growth faces a supply wall.

Key Terms

  • EUV lithography: Extreme Ultraviolet lithography, the advanced chip printing process that requires helium.
  • HBM: High Bandwidth Memory, the specialized memory used in NVIDIA GPUs, AMD accelerators, and Google TPUs.
  • Ras Laffan: Qatar's massive industrial complex producing both LNG and helium.
  • 6N purity: 99.9999% purity standard required for semiconductor-grade helium.
  • Power of Siberia 2: Planned Russia-to-China natural gas pipeline that could shift energy economics in China's favor.
  • Cost per flop: The standard measure of inference computing cost, directly affected by chip and energy prices.
YouTube