Sector Initiation Preview: May 2026

$10 trillion of capex. 190 GW of compute. One bottoms-up map.

Rolf Bulk

Head of Semiconductors & Infrastructure · Futurum Equities

A fully integrated view of the AI supply chain — from datacenter capex through GPUs, CPUs, memory, foundry, and semicap. Built bottom-up from unit economics, not top-down extrapolation.

The thesis, in five numbers

Demand is inflecting. Supply is the binding constraint.

Cumulative datacenter capex through 2030. Annual capex grows from $1.3T this year to ~$2.7T by 2030.
$ 0 T
Installed AI datacenter capacity. Base case: 190 GW by 2030 (range: 145–215 GW).
GW
GPU/ASIC spending in 2030, up close to 5x vs. 2025 ($195B).
$ 0 B
Wafer starts/month required in 2030 across Logic, DRAM, NAND, and CoWoS to enable the AI build-out.
0 M
Top picks across compute, memory, foundry, and semicap.
0

Coverage Universe

Fourteen names. The full AI compute stack.

NVDA

NVIDIA

AMD

AMD

AVGO

Broadcom

TSM

TSMC

ASML

ASML

AMAT

Applied Materials

LRCX

LAM Research

KLAC

KLA

MU

Micron

000660.KS

SK HYNIX

005930.KS

Samsung

INTC

Intel

ARM

Arm Holdings

8035.T

Tokyo Elctron

Exhibit 32 · One chart from the report

XPU spending is set to increase to more than $0.9T in 2030, with Nvidia retaining a leading position.

Nvidia
Broadcom
AMD
Others

What's inside · 12 chapters

The argument, end to end.

I
Demand inflection
Token consumption up 2.5x in six months. Hyperscaler RPO exited 1Q26 at $2.2T.
II
The AI Infrastructure build-out
Installed GW capacity grows 5–8x by 2030. Still early innings.
III
The $10 trillion capex ticket
Annual DC capex hits $1.3T this year (+85%), $2.7T by 2030.
IV
Money & Power
Financing mix is shifting; power eases as the binding constraint as gas turbines and flexible load come online.
V
Accelerators — a $1T opportunity
XPU revenue grows from $360B to $936B by 2030. Nvidia retains share leadership; AMD ramps; AVGO scales as the dominant ASIC vendor.
VI
CPUs — back in the limelight
Revenue $29B → $135B (37% CAGR). CPU-to-XPU ratios approach parity.
VII
Memory — upcycle longevity underappreciated
Memory could be a $1T category by 2030. The longevity of this cycle is structurally underestimated.
VIII
Logic — almost doubling in the BoM
XPUs increasingly designed on leading-edge nodes; wafer-content per accelerator scales as complexity rises.
IX
Advanced Packaging
CoWoS, the 2024–25 bottleneck, eases. Other constraints take over.
X
Semicap — $380B of WFE
EUV is the binding back-half-of-decade constraint. ~410 incremental tools required against ASML's ~90/yr ceiling.
XI
Reconciling demand & supply
Where the model says the bottlenecks bite, in what order, and through what year.
XII
Owning the build-out
Valuation framework + the 5 named top picks across compute, memory, foundry, and semicap.

Read the full initiation report.

Apply for access to Semiconductors & Infrastructure Research and Coverage. Approved applicants receive the full report, access to supporting models, and ongoing research on our semiconductor & infrastructure coverage.

Rate your experience