Powering AI: The Semiconductor Ecosystem at the Foundation of Data Centers
📄 Read the full report ↗SIA–Deloitte · Jun 2026 · 45 pp
Summary
SIA and Deloitte conduct a “virtual teardown” of a state-of-the-art AI data server rack, mapping every semiconductor component from AI accelerators down to voltage regulators. A single rack contains over 4,500 packaged chips (~20,000 individual dies), with semiconductors accounting for 95% of the rack’s content value and over 50% of total data-centre capex. The report projects cumulative data-centre investment of over $4 trillion through 2028, with up to $2.8 trillion spent on semiconductors. Annual AI data-centre semiconductor revenue could reach $1.2 trillion by 2028 — a nearly tenfold increase over four years.
Key Insights
AI accelerators capture 74% of a server rack’s semiconductor value. Logic chips overall account for 65% of content. Memory and storage make up over 20%. Yet 52% of chips in a server cost less than $10 each — the long tail of power management, sensors, and analog ICs is numerically dominant but value-subordinate. (pp. 33, 36)
83% of power-tray dies use foundational/mature nodes (>28nm). AI infrastructure does not run on leading-edge alone. Compute trays use 49% advanced (<10nm), but power trays are 83% foundational and CDU (cooling) trays are 100% foundational. The entire mature-node segment is load-bearing for AI. (p. 35)
Revenue trajectory: $27bn (2022) → $320bn (2025) → $1.2tn (2028). The 88.8% CAGR from 2022–2028 is driven by AI training initially, but inference is forecast to grow from 20% of demand (2024) to 80% by 2032, at 122% CAGR through 2028. (pp. 31–32)
A single AI server rack costs $1.5M–$4M. A leading data centre housing 10,000 racks would contain over 45 million chips. Future racks may individually require 1 MW of power, up from 100–120 kW today. (pp. 5, 9, 19)
Supply chain is concentrated in five geographies. Design in the US and Korea; fabrication in Taiwan, Korea, US; advanced packaging in Taiwan, Korea, Malaysia, US; module assembly in China, Taiwan, Vietnam. India is absent from every stage. (p. 38)
What This Means for India
India’s OSAT/ATMP investments are invisible in the AI supply chain. Micron Sanand (ATMP) and CG Power’s OSAT facility do not register in SIA’s supply-chain mapping because they are not yet operational at scale and do not serve the AI data-centre segment. For India to appear on this map even in the “traditional packaging” row, it needs operational capacity serving global customers, not just approved projects.
The mature-node opportunity is real but narrow. 83% of power-tray dies and 100% of cooling-tray dies use foundational nodes — exactly the segment Tata/PSMC’s Dholera fab targets. If Indian fabs can produce power-management ICs and analog chips at competitive cost, they could serve a segment of AI infrastructure that is growing at the same rate as the headline AI chip market. But the window to enter is shrinking as China expands mature-node capacity aggressively.
AI data-centre buildout is the single largest demand driver for the next decade. $2.8 trillion in semiconductor spending through 2028 is not abstract — it translates into specific chip types, packaging technologies, and supply-chain positions. India’s semiconductor strategy needs to articulate which slice of this $2.8 trillion it is targeting and what it would take to capture even 1% of it.
Data Extracted
AI data-centre semiconductor revenue trajectory (Figure 15, p. 31)
| Year | Revenue (est.) |
|---|---|
| 2022 | $27 billion |
| 2023 | $45 billion |
| 2024 | $123 billion |
| 2025 | $320 billion |
| 2026 (f) | $670 billion |
| 2027 (f) | $905 billion |
| 2028 (f) | $1.2 trillion |
Process node distribution by tray type (Figure 17, p. 35)
| Tray | Advanced (<10nm) | Current (10–28nm) | Foundational (>28nm) |
|---|---|---|---|
| Compute | 49% | 13% | 38% |
| Interconnect | 6% | 94% | — |
| IPMI | 1% | 80% | 19% |
| Power | — | 17% | 83% |
| CDU (Cooling) | — | — | 100% |
Supply chain geography (Figure 20, p. 38)
| Stage | Key locations |
|---|---|
| Product design | US, South Korea |
| Wafer fabrication | Taiwan, South Korea, US |
| Testing & die cutout | Taiwan, South Korea, Malaysia, Vietnam, Philippines, US |
| Advanced packaging | Taiwan, South Korea, Malaysia, US |
| Traditional packaging | Taiwan, China, South Korea, Malaysia, Vietnam, Philippines |
| Module/board assembly | China, Taiwan, Vietnam |
India: absent from all stages.
Source
SIA & Deloitte (2026), Powering AI: The Semiconductor Ecosystem at the Foundation of Data Centers, Semiconductor Industry Association, June 2026, https://www.semiconductors.org/wp-content/uploads/2026/06/SIA-AI-Data-Center-Report_June-2026.pdf.