◆ INVEST DEEPER / Research by Amit Garg · October 2026 · Updated Oct 5
AI needs more than chips. It needs memory, packaging, power, cooling and cables, and some of those are in short supply. This page tracks the 12 possible bottlenecks that decide how fast AI gets built, and which stocks gain or lose while each one lasts.
This month in 30 seconds
- The picture: 2 of 12 bottlenecks are 🔴 Severe—storage and large gas turbines. Seven are 🟠 Tight; three remain ⚪ Not established. Strong demand alone does not confirm a shortage.
- What matters: pricing evidence is strongest in memory, storage and gas turbines, with product mix affecting some comparisons. Power also faces long waits: large-turbine agreements extend into 2031, including reservations rather than only firm orders.
- Next checkpoint: TSMC reports Oct 15, followed by memory, power and equipment results through November. Watch pricing, qualified capacity and delivery times.
The scoreboard
Click any row to open the full page. "When relief might come" is a rough forecast from company plans and industry sources, not a promise.
Bottlenecks
Bottleneck
Status
Trend
Confidence
Bottom line
When relief might come
Next key date
🔴 Severe
◆ First reading
High
Flash and hard drives each show all four warning signs on their own evidence. Supply is growing through denser products, but makers say demand is growing faster.
Into 2028. Supply is allocated, and makers are adding density rather than factories.
🔴 Severe
◆ First reading
High
Large gas turbines are being booked into 2030–2031, counting firm orders and reservations, and GE Vernova's prices per kW are rising, partly from a richer mix. Engines and fuel cells are alternatives with their own trade-offs.
2030+. Orders now being taken for 2031.
🟠 Tight
◆ First reading
Medium
Data centers draw on a skilled-trades workforce that is already short. Most contractors with openings struggle to fill them, pay is rising, and worker shortages remain the most common cause of project delays.
Unclear. Training is rising, but no data compares newly qualified workers with demand. Electricians take 4–5 years to train.
🟠 Tight
◆ First reading
Medium
Advanced chip substrates are short: makers report demand above capacity and firmer prices. Glass cloth supply is improving for some buyers, and a shortage of the insulating film isn't established.
2027+ for advanced substrates, as Ibiden's first expansion phase starts. T-glass capacity triples by March 2029. Makers say demand may still outrun it.
🟠 Tight
◆ First reading
High
New data centers wait years to get connected, and the biggest US grid has fallen short of its reliability target in two auctions in a row, with capacity prices stuck at the cap.
2028+. Needs new plants and power lines that take years.
🟠 Tight
◆ First reading
Medium
Many leading AI accelerators depend on HBM, and only three companies make it at scale. Micron has sold most of its 2027 HBM at significantly higher prices; how long the tightness lasts depends on qualified output and demand.
Tight through 2027 (TrendForce). New fabs start output from mid-2027, but meaningful, qualified HBM output lags by several quarters.
🟠 Tight
◆ First reading
Medium
The indium phosphide lasers inside AI optics are booked well ahead, and makers expect to trail demand even after expanding. Price rises for these lasers aren't confirmed yet.
2027–2028. New InP laser capacity and fabs come online, at different speeds by laser type.
🟠 Tight
◆ First reading
Low
Large power transformers remain constrained: 128-week waits and a ~30% US supply gap in 2025 data, with strong transformer orders in 2026. Prices are high but plateauing.
2028+. New US factories open 2027–2028. Smaller distribution units already improving.
🟠 Tight
◆ First reading
Medium
TSMC says packaging capacity is so tight it limits customers' growth. Capacity is growing fast, but TrendForce expects only slight relief by 2027.
2027 for slight relief (TrendForce). Alternatives matter more from 2028.
⚪ Not established
◆ First reading
Low
TSMC confirms advanced-node supply is short of demand, but public data can't yet measure how severe the shortage is or how long it lasts.
Unclear. TSMC says the demand-supply gap is "so big" but doesn't foresee bottlenecks to its expansion plans. Data on waits and prices is thin.
⚪ Not established
◆ First reading
Low
Demand for data center electrical gear is strong and some product lines have extended lead times, but a category-wide switchgear shortage isn't established.
Unclear. Makers are adding capacity now, but public data doesn't establish a shortage.
⚪ Not established
◆ First reading
Low
Growing fast and suppliers are expanding, but no maker has published lead times, a demand gap or cooling-specific pricing. Shortage not established.
Unclear. Suppliers are expanding fast, but no public data shows whether supply is keeping up. Watch small parts like fittings.
Bottleneck
Status
Who benefits
Who's held back
What changed this month
First issue. Every bottleneck is a first reading, so there's no change to report yet. From November this section leads with what moved and why.
Dates to watch
Dates marked * are expected dates from earnings calendars. The company hasn't announced them yet.
Oct 8: TSMC September sales, plus Samsung's preliminary Q3 results. First reads on chip and memory demand.
Oct 15: TSMC Q3 earnings call. Packaging capacity, 2nm ramp and pricing.
Oct 22*: Intel and Comfort Systems results.
Oct 26–30*: Amkor (26), SK hynix (27), Bloom (27), Hubbell (27), GE Vernova (28), Vertiv (28), Seagate (28), Samsung (29), Western Digital (29), ASE (29), EMCOR (29), nVent (30). Schneider Electric and Ibiden report Oct 29 per their calendars.
Nov 3–11*: Eaton and Lumentum (3), Caterpillar, Talen and TTM Technologies (4), NRG and Nittobo (5), Vistra (6), SanDisk (early Nov), Constellation (9), Coherent and Siemens Energy (11).
Nov 18*: NVIDIA results.
Dec 23*: Micron results.
How to read this page
Each bottleneck is checked against four warning signs every month:
- Long waits. Buyers wait longer, or the seller says it's sold out a year ahead.
- Orders above supply. The company or a trusted tracker says demand is bigger than what can be made.
- Prices going up. A confirmed price increase for this exact product, from a representative measure such as a market price or index. One isolated deal isn't enough.
- New capacity not enough. Even after planned expansion, supply still falls short.
Each sign is 🔴 red, 🟢 green, ⚖️ mixed, or ⚪ unknown. Mixed means we have recent data but it points in different directions. Unknown means we don't have recent data that settles the sign. Neither counts either way.
- ⚪ Not established: fewer than 2 signs confirmed red or green. We can't call it either way yet.
- 🔴 Severe: all 4 signs confirmed red.
- 🟠 Tight: 2 or more red, and more reds than greens.
- 🟡 Balanced: reds and greens are equal.
- 🟢 Easing: more greens than reds.
Confidence shows how much of the scorecard rests on recent data. We count signs that are unknown or rest only on data more than 12 months old (marked "dated"): none = High, 1–2 = Medium, 3 or more = Low. A mixed sign has recent data, so it doesn't lower confidence. Every sign on a bottleneck page shows its source date, so you can see how fresh it is.
Trend is what we actually measured against last month. Forecasts go in the page's outlook line, not here. A shortage moving from 🟠 to 🟡 often matters more to a stock than the color itself, because that's when pricing power starts to slip.
Who benefits are sellers that can charge more while supply is short. Who's held back are buyers that have orders but can't get enough parts, or pay more for them.
A tight bottleneck isn't a buy signal. A stock can already price in the shortage.
Archive
- Oct 2026: first issue. 2 Severe (storage, gas turbines), 7 Tight (HBM, transformers, grid access & power availability, optical lasers, CoWoS, chip substrates, skilled construction labor), 3 Not established (chip wafers, switchgear, liquid cooling).
Invest Deeper · Educational research, not investment advice.
Made with Bullet