The 10 tightest bottlenecks in the AI data center build-out right now
We score 50 parts of the AI supply chain from 0 to 100 on how hard they are to get more of: supplier concentration, lead time, capital needed, geopolitical risk and pricing power. These were the ten tightest at the time of writing (October 2026).
1Gas turbines (score 90, lead time 3–5 years)
Most new power for data centers in the next few years comes from gas. GE Vernova, Siemens Energy and Mitsubishi Heavy have order slots booked years out.
Gas turbines →2AI memory: HBM and DRAM (88, 12–18 months)
Every AI accelerator needs stacks of high-bandwidth memory. SK hynix, Samsung and Micron are effectively sold out into 2027.
AI memory →3Grid connection and transmission (88, 3–7 years)
Waiting to connect a large campus to the grid can take longer than building it. This is why so many projects build their own power.
Grid connection →4Transformers and switchgear (86, 2–4 years)
US producer prices for power transformers have roughly doubled since 2020, a sign that demand still outruns supply.
Transformers →5Rare earths and magnets (86, 3–6 years)
China controls most magnet production. Export licensing has shown how one input can halt motors, robots and drones.
Rare earths →6Leading-edge chipmaking (84, 2–3 years)
TSMC makes almost every advanced AI chip. New fabs take years and tens of billions of dollars.
Foundry →7Laser chips for optics: indium phosphide (82, 1–2 years)
The lasers inside 800G and 1.6T transceivers come from a handful of fabs, led by Lumentum and Coherent.
Laser chips →8Chip packaging: CoWoS (80, 12–18 months)
Advanced packaging that joins GPUs and HBM is a hard ceiling on how many AI chips ship each quarter.
Chip packaging →9Gallium, germanium and antimony (80, 2–4 years)
Minor metals used in chips and optics where China dominates supply and has restricted exports.
Minor metals →10Hyperscaler spending (80)
Not a shortage but the demand side: Microsoft, Alphabet, Amazon, Meta and Oracle's capex sets the pace for everything upstream.
Hyperscalers →What the ranking tells you
Notice how many of the tightest bottlenecks are physical and slow to fix: turbines, grid connections, transformers and rare-earth magnets all take years to expand. Chips and memory are tight too, but their supply can respond faster. That is why we think power equipment has one of the most durable shortages in the build-out.
How to use it
Open any bottleneck to see its lead time, the companies that control it, how their stocks have performed and what would weaken the case. The heat map then shows which tight bottlenecks have already been priced in by the market and which haven't.
The full ranking, with the stocks behind each bottleneck, is updated as conditions change.
Frequently asked questions
What is a tightness score?
Our 0–100 editorial score of how short supply is: concentration, lead time, capex needed, geopolitical risk and pricing power. It is judgement, not market data.
Which bottlenecks are easing?
At the time of writing: silicon wafers and chemicals, grid batteries, and lithium, graphite and cathodes.
How often do scores change?
When the evidence changes, for example new lead-time data, capacity announcements or export rules. Each bottleneck page explains what would prove the call wrong.
Get the daily brief
Everything that happened across the AI supply chain, in a 3-minute email every weekday. Free.
More guides
Figures quoted are as of the publish date; the linked pages show live data. Educational research by Singularity Research, not investment advice.