The economics inside an AI rack

The commercial story inside an AI server rack is not the same as the story told by its component count. SIA and Deloitte’s June 2026 report estimates where value sits in a representative configuration and shows a sharp concentration in accelerator modules. That is useful context for understanding the hardware stack, but it is not a market-share table or a forecast of any supplier’s profit. The analysis below uses the report as a dated model of a system, rather than evidence that every deployed rack has the same economics.

Value is concentrated

Figure 16 assigns 74% of server-rack component content value to the AI accelerator category, compared with 8% for CPUs and 5% each for volatile memory and networking/interconnect components. These are approximate shares in one modeled rack, not revenues reported by chip companies. Accelerator modules themselves contain multiple types of silicon and packaging; treating the accelerator share as pure logic-chip value would therefore misread the chart. The source’s denominator is server-rack content value, not the cost of a complete data center. (SIA/Deloitte, June 2026, p. 33, figure 16.)

Count tells a different story

In figure 19, chips priced below US$10 represent 52% of the modeled rack’s chip count, while the category above US$1,000 represents 3%. The two measures answer different questions: count identifies the breadth of the component base; content value identifies where expenditure is concentrated. Neither establishes bargaining power, supplier margins or the cost of replacing a failed component. A low unit price is not evidence that a component is optional. (SIA/Deloitte, June 2026, p. 36, figure 19.)

Why a single configuration is not a market model

The report explicitly describes variation with accelerator selection. Its configuration assumptions are therefore central to interpreting the value split. A rack aimed at a different workload or built around different hardware may have a different mix. High-bandwidth memory means memory designed to move large amounts of data close to the accelerator; it is part of the wider system, not an interchangeable label for all memory. For business analysis, the useful question is how a company’s actual products fit that architecture, not whether its name can be attached to the largest bar. (SIA/Deloitte, pp. 13 and 34.)

What to watch

Company filings are still needed to evaluate product exposure, customer concentration, margins and cash generation. The report supports a distinction between value concentration and system dependence; it does not turn that distinction into a return forecast. This article deliberately omits the report’s aggregate infrastructure-spending projection because its executive summary and the discussion on page 33 use different time horizons. That inconsistency must be resolved before the figure is reused as a headline.