Material Insights · Understanding the Material World

Data centres and Semiconductors

Data centres now bring in about nine-tenths of the leading AI chip designer's revenue and over half of the leading foundry's, and each new generation of AI accelerator packs more power into a rack, so chip roadmaps set data centre power and cooling needs while chip designers themselves rely on large compute farms.

Data centres (energy) and Semiconductors (industry): 4 relationships.

How they are linked

1. Industry input

Physical flow · Strong · Data centres → semiconductors

Chip design depends on simulation and verification run on large compute farms, so the site rates data centre capacity 5 of 5 for EDA software and IP-core vendors and 4 of 5 for fabless designers.

  • 2 enterprise types in semiconductors that depend on data centres
  • 5 of 5 criticality for EDA and IP-core vendors (simulation and verification)

This is the smaller of the two flows: the chip industry buys compute from data centres, mostly cloud capacity, for design and AI-model testing. As designs grow in size, design compute grows with them.

Enterprise types that depend on it:

  • EDA software and IP-core vendors: simulation and verification run on large compute farms (criticality 5 of 5)
  • Fabless chip designers: design and AI-model testing compute (criticality 4 of 5)

Sources: Material Insights

2. Largest chip customer

Transition · Strong · Data centre build-out → chip demand

AI training and inference run on accelerators and high-bandwidth memory made at leading-edge fabs, and data centre operators buy them in very large volumes.

  • $193.7 billion NVIDIA data centre revenue, fiscal year to January 2026, up 68%
  • $215.9 billion NVIDIA total revenue in the same year
  • 55% high-performance computing share of TSMC revenue, Q4 2025

About nine-tenths of the largest chip designer's sales now come from data centres, and high-performance computing is more than half of the leading foundry's revenue. This ties chip earnings to hyperscaler capital spending, so a pause in data centre building would reach chipmakers quickly.

Sources: NVIDIA via US SEC; TelecomLead

3. Chips set power density

Physical flow · Rising · Accelerator chips → data centre power and cooling

Each accelerator generation draws more power and is packed more tightly into racks, so the chips chosen decide how much power a hall needs and whether it must be liquid-cooled.

  • 120 to 140 kW power per rack for NVIDIA's rack-scale systems, 2025
  • 7 to 20 kW power of a typical conventional rack
  • 17% growth in data centre electricity use worldwide, 2025

Racks drawing several times the power of conventional designs cannot be cooled by air alone, so older halls need refitting. The IEA expects data centre electricity use to double by 2030, and the chips installed decide how that load is packed into sites.

Sources: TechSpot; IEA

4. Tight supply chains

Shared risk · Medium · Chip and power equipment supply → data centre build-out

A data centre needs both advanced chips and power equipment, and supply of each is concentrated, so a shortfall in either delays capacity.

  • up to about US$56 billion TSMC capital spending signalled for 2026, as much as 37% above 2025

The IEA notes that supply chains for advanced chips and IT components tightened in 2025, as did those for gas turbines and transformers. Foundry capacity expansion takes years, so chip supply and power connections both pace the build-out.

Sources: TelecomLead; IEA

For investors

Key learnings

  1. Data centres made up about 90% of NVIDIA's revenue in fiscal 2026 and high-performance computing 55% of TSMC's in Q4 2025, so chipmaker earnings track hyperscaler capital spending closely.
  2. Each accelerator generation raises rack power, with current systems at 120 to 140 kW per rack, which shifts value to liquid cooling and power equipment.
  3. Chips and power equipment are both supply-constrained, so the slower of the two sets how fast data centre capacity grows.

What to track

  • Hyperscaler capital expenditure guidance: Demand for accelerators and memory over the next year (Quarterly · company results)
  • TSMC revenue by platform: Strength of data centre chip demand at the foundry level (Monthly sales, quarterly breakdown · TSMC)
  • Data centre electricity use: How chip deployment is turning into power demand (Yearly · IEA)

For policymakers

Key learnings

  1. Power planning for data centres needs chip roadmaps, because rack density sets load per site and cooling needs.
  2. Chip supply for data centres depends on a few leading-edge fabs, which makes it a strategic dependency for countries building AI capacity.
  3. For India, data centres are a domestic market for chips, while chip design firms need large local or cloud compute, so data centre and semiconductor policies support each other.

What to track

  • Leading-edge foundry capacity and capex: Future availability of accelerators (Quarterly · foundry results)
  • Average rack density in new builds: Power and cooling standards needed for new data centre parks (Yearly · industry surveys)

Sources

  1. Semiconductors: materials, Material Insights (2026)
  2. NVIDIA announces financial results for fourth quarter and fiscal 2026 (Form 8-K exhibit), NVIDIA via US SEC (2026)
  3. TSMC delivers record 2025 revenue and profit as AI chip demand fuels growth, TelecomLead (2026)
  4. Closed-loop liquid cooling powers Nvidia push for greener AI, TechSpot (2025)
  5. Data centre electricity use surged in 2025, even with tightening bottlenecks driving a scramble for solutions, IEA (2026)