Goldman sees AI capex doubling from $765B this year to $1.6T by 2031 — $7.6T cumulative spend across compute, data centers, and power.
Compute scales from $494B to $1.1T because chip supply eventually catches up if you throw enough fabs and time at it.
Power is the real constraint. Only $73B by 2031 in the model, but grid interconnection takes 4–10 years vs 2 years to build a data center. US electricity demand from AI jumps from 10% to 25% by 2030. That's the choke.
Advanced packaging is sold out. TSMC's CoWoS is fully allocated. Amkor is the main independent alternative, building capacity in Vietnam and Arizona to give hyperscalers supply outside Taiwan.
Onto Innovation supplies inspection tools that decide if HBM stacks pass QC — locked in a $240M purchase agreement through 2027.
Kulicke & Soffa makes bonding equipment. Amtech Systems already seeing AI packaging hit 40%+ of revenue this quarter.
Interconnects inside the rack are their own layer. Astera Labs makes retimers that keep signals clean across dense AI server architectures. Credo makes the purple AEC cables linking GPUs inside clusters — revenue up 272% YoY. Fabrinet assembles optical transceivers and just reported a 252% jump in orders as rack power density blew past prior limits.
Data center REITs are the real estate play. Equinix and Digital Realty dominate interconnection and wholesale leasing. Iron Mountain pivoted from filing cabinets to data centers and outperformed $NVDA over the past year.
Semiconductor equipment sits underneath the entire fab buildout. Applied Materials, Lam Research, KLA touch nearly every wafer that becomes an AI chip or HBM stack. Teradyne tests finished chips before they ship — roughly 70% of revenue now tied to AI.
Robotics and automation round out the less obvious corner. Rockwell Automation, Nordson, Fortive benefiting as AI-driven automation spreads into factories and industrial equipment.
The AI infrastructure buildout is a multi-year, multi-layer trade. Packaging, interconnects, power, data centers, and the picks-and-shovels names underneath it all.
Compute scales from $494B to $1.1T because chip supply eventually catches up if you throw enough fabs and time at it.
Power is the real constraint. Only $73B by 2031 in the model, but grid interconnection takes 4–10 years vs 2 years to build a data center. US electricity demand from AI jumps from 10% to 25% by 2030. That's the choke.
Advanced packaging is sold out. TSMC's CoWoS is fully allocated. Amkor is the main independent alternative, building capacity in Vietnam and Arizona to give hyperscalers supply outside Taiwan.
Onto Innovation supplies inspection tools that decide if HBM stacks pass QC — locked in a $240M purchase agreement through 2027.
Kulicke & Soffa makes bonding equipment. Amtech Systems already seeing AI packaging hit 40%+ of revenue this quarter.
Interconnects inside the rack are their own layer. Astera Labs makes retimers that keep signals clean across dense AI server architectures. Credo makes the purple AEC cables linking GPUs inside clusters — revenue up 272% YoY. Fabrinet assembles optical transceivers and just reported a 252% jump in orders as rack power density blew past prior limits.
Data center REITs are the real estate play. Equinix and Digital Realty dominate interconnection and wholesale leasing. Iron Mountain pivoted from filing cabinets to data centers and outperformed $NVDA over the past year.
Semiconductor equipment sits underneath the entire fab buildout. Applied Materials, Lam Research, KLA touch nearly every wafer that becomes an AI chip or HBM stack. Teradyne tests finished chips before they ship — roughly 70% of revenue now tied to AI.
Robotics and automation round out the less obvious corner. Rockwell Automation, Nordson, Fortive benefiting as AI-driven automation spreads into factories and industrial equipment.
The AI infrastructure buildout is a multi-year, multi-layer trade. Packaging, interconnects, power, data centers, and the picks-and-shovels names underneath it all.