Subscribe

Methodology

JP Data’s AI infrastructure model is built from the server bill of materials, not from a single top-down TAM. Demand starts with OEM, ODM, and hyperscaler-direct channels, then splits accelerated versus general-purpose systems. That spend is rolled into CPU (x86 and Arm), accelerators (GPU versus ASIC / TPU), server DRAM (HBM versus DDR), and enterprise SSD (a slice of Others). The result is a consistent set of dollar series from 2020 that shows where AI infrastructure money actually lands in the stack.

Supply is estimated the same way the market fills those dollars: Intel and AMD x86 plus Arm server CPU vendors; NVIDIA GPU, AMD Instinct, and captive or merchant ASIC (TPU, Trainium, Maia, MTIA, Gaudi); SK hynix, Samsung, and Micron server DRAM; and enterprise SSD from Samsung, SK hynix, Kioxia, Sandisk, Micron, and Solidigm. Coverage is supply divided by demand. Lead time is modeled channel weeks for CPU, GPU, and HBM, checked against research survey bands. Hyperscaler CapEx (MAG, MANGOS, neoclouds, and company rows) is the funding signal behind the demand.

The live model runs through 2027. The Forecast view extends server, CPU, GPU, and ASIC series to 2032, with history through 2025, model years in 2026–2027, and JP Data extensions thereafter. Numbers are not a one-time snapshot. The data is updated frequently as earnings, channel checks, CapEx guidance, and component availability change, so the views stay aligned with current market conditions.

Value proposition

• One connected view of demand, supply, tightness, and spend — not isolated chip or server estimates.
• AI versus general-purpose, GPU versus ASIC, and HBM versus DDR on the same dollar basis from 2020.
• Coverage and lead time to see when the stack is tightening or easing, alongside the CapEx that funds it.
• A forward path through 2032, with the nearer years tied to the live model rather than a disconnected long-range story.
• Frequent refreshes so positioning, capacity, and TAM work reflect the market as it stands, not last quarter’s print.

How clients use the data

• Hedge funds use the series to test vendor and hyperscaler earnings, see whether coverage and lead times support pricing power, and judge whether CapEx and accelerator mix are running ahead of or behind consensus.
• Hyperscalers use demand, coverage, and lead time to plan CPU, GPU/ASIC, and HBM availability, compare merchant GPU versus captive ASIC, and size the server BOM against CapEx plans.
• Semiconductor companies use the TAM split to see where dollars are moving across CPU, GPU, ASIC, and HBM, where supply is short, and how hyperscaler mix and forecast growth affect product and capacity decisions.

What a subscription includes

Subscribers get ongoing access to the Demand, Supply, Signals, and Forecast views on this site — the dollar series, coverage and lead-time indicators, hyperscaler spend, and the 2032 extension — as JP Data refreshes the model. Updates flow into the same charts and tables, so you are not reconciling a stale deck against a new market. Briefings and custom cuts of the model are available as part of the engagement when you need them for a name, a region, or a product line.