The demand model starts from two cash totals and walks each one down to the same eight components. Hyperscaler spend is what the named buyers lay out for property and equipment. OEM server spend is what server vendors invoice. The same machines can appear in both, so the combined total counts each server once.
From buyer capex to a bill of materials
The buyer file begins with cash capex. In 2025 the calendarised total for the seventeen companies is $468 billion. The fiscal-year sum of the same companies is $435 billion. The bill of materials uses the calendarised figure.
An IT share takes capex to IT capex, $391 billion in 2025. A server share of IT then takes that to server spend, $248 billion. Buildings and the rest of the facility stay above the server line. Server spend is split across three kinds of machine — GPU racks, custom-chip racks, and general-purpose servers — and then across eight components. The eight lines add back to server spend.
From OEM revenue to the same bill
The OEM file begins with worldwide server vendor revenue, $444.1 billion in 2025. A haircut removes services and the non-server lines inside that revenue and leaves $429 billion of server hardware. The same three machine types and the same eight components follow. On a branded server, assembly and margin also holds the vendor’s gross margin, so the eight lines add back to hardware.
One server, one demand
The seventeen buyers’ server bill is taken from capex. Machines those buyers purchase from a branded OEM, or through the unbranded channel, are already inside that bill, so they come out of the OEM side. What remains is everyone else: enterprise, other clouds, and the rest of the market. In 2025 that residual is $210 billion. Add it to the buyers’ $248 billion and server demand is $458 billion.
Above the server line, data-center IT is $685 billion in 2025. It includes the buyers’ IT capex plus the non-server IT that sits on top of everyone else’s servers. The forecast carries the same identity forward. In 2032, data-center IT is $2,616 billion, server demand is $1,750 billion, the seventeen buyers are $1,110 billion of that, and accelerators are $878 billion.
Component demand
The table is the eight components after the overlap is removed. Figures are billions of dollars. 2025 is the latest closed year. 2032 is the forecast. Component lines are rounded to the nearest tenth, so a column can differ from the total row by a tenth. The total row is the model’s server demand.
| Component | Hyperscaler Capex, 2025 | Other buyers, 2025 | Total demand, 2025 | Total demand, 2032 |
|---|---|---|---|---|
| GPU and custom accelerators, including high-bandwidth memory | 107.7 | 84.0 | 191.7 | 878.3 |
| Host CPUs | 35.2 | 24.0 | 59.2 | 180.2 |
| DRAM | 30.8 | 20.6 | 51.4 | 145.2 |
| Storage, SSD and hard disk | 16.8 | 11.4 | 28.2 | 83.6 |
| Network silicon | 17.2 | 12.7 | 29.9 | 123.0 |
| Other chips | 2.9 | 2.3 | 5.2 | 16.0 |
| Rack, power, cooling, and chassis | 23.5 | 16.4 | 39.9 | 152.5 |
| Assembly, test, and margin | 13.7 | 38.4 | 52.1 | 171.2 |
| Server demand | 247.9 | 209.8 | 457.7 | 1,749.9 |
Accelerators are the largest line: $192 billion of the $458 billion in 2025, and $878 billion of the $1,750 billion in 2032. Host CPUs are $59 billion, then $180 billion. High-bandwidth memory sits inside the accelerator line. The DRAM row is host memory.