🌊THE FORCES

Structural Trends

Nine forces reshaping the AI infrastructure value chain. Each trend is written twice — once in technical detail, once in plain English — and maps to the specific companies that benefit most.

Current snapshot

9 trends tracked · 4 constraints · 5 structural shifts

Reviewed monthly · Last updated April 2026

The Nine Trends

1

The Great Packaging Bottleneck

CONSTRAINT

Every AI chip needs to be glued to its memory on a precision silicon base — and only TSMC can do it at scale.

Status: Current and getting worse through 20275 primary beneficiaries · 10 total
2

The Memory Supercycle

CONSTRAINT

A GPU without memory is useless. Three companies make all the memory, and they are sold out through 2026.

Status: Current — SK Hynix CFO confirmed 2026 HBM supply fully sold out3 primary beneficiaries · 9 total
3

The Optical Rewiring of AI

STRUCTURAL SHIFT

At AI cluster speeds, copper wires hit a physics ceiling. Light through glass is the only way forward.

Status: Current (1.6T pluggables shipping) → Accelerating (CPO in 2026) → Structural (all-optical racks by 2028–2030)3 primary beneficiaries · 9 total
4

The Power Grid Reckoning

CONSTRAINT

You cannot plug in the AI future on a grid built for the 20th century. The equipment isn't there, the electricity isn't there, and the lead times are measured in years.

Status: Current crisis → Multi-year structural constraint → Long-tail nuclear build4 primary beneficiaries · 14 total
5

The Liquid Cooling Mandate

CONSTRAINT

Chips are generating so much heat that air literally can't carry it away. Every future AI rack needs plumbing.

Status: Tipping point now — air cooling is already obsolete for frontier AI racks3 primary beneficiaries · 9 total
6

The Custom Silicon Surge

STRUCTURAL SHIFT

Every major hyperscaler is now designing its own AI chip. NVIDIA's monopoly on AI compute is ending.

Status: Current and rapidly taking share from GPU compute on inference workloads5 primary beneficiaries · 9 total
7

The Inference Architecture Shift

STRUCTURAL SHIFT

Training was a batch job. Inference is a 24/7 production workload — and it needs a completely different infrastructure.

Status: Emerging now → Dominant by 20274 primary beneficiaries · 10 total
8

The 800V Power Architecture

STRUCTURAL SHIFT

NVIDIA is mandating a complete rewrite of how power flows from the grid to the chip. Every part of the chain gets replaced.

Status: Emerging now → Mandatory for Rubin-era racks → Structural redesign of the DC electrical stack5 primary beneficiaries · 12 total
9

The Scale-Up Fabric Wars

STRUCTURAL SHIFT

The fight over how GPUs talk to each other inside a pod is the single most important standards war in AI infrastructure.

Status: UALink 2.0 spec released April 2026 → First products Q4 2026 → Structural by 20284 primary beneficiaries · 10 total

How to read this list

Constraint

A scarce input that is actively bottlenecking AI buildout today. Revenue follows capacity, and capacity is bound by a physical process.

Structural shift

A change in how AI infrastructure is built, financed, or operated. Creates new winners and displaces incumbents over multiple years.

The Big Picture

These nine trends are not independent. Packaging limits memory. Memory limits inference architecture. Inference architecture demands new fabrics. Fabrics demand optics. Optics demand power. Power demands cooling. The AI buildout is a chain of dependencies — follow the chain and you find the real winners.