Investment Thesis — March 2026

TheAIInfrastructureThesis

Why the picks-and-shovels play is the only play that survives every hype cycle.

● ○ ●

ThePatternRepeats

Every transformational technology follows the same arc: euphoria, overbuilding, consolidation, utility. The internet saw it. Mobile saw it. Cloud saw it. AI is seeing it now.

But in every cycle, one class of company survives regardless of which application wins: the infrastructure layer. Cisco shipped routers whether you used Yahoo or Google. AWS sold compute whether your startup succeeded or failed. NVIDIA sells GPUs whether the model that runs on them cures cancer or generates cat memes.

The thesis is simple: **don't bet on the application. Bet on the picks and shovels.**

TheNumbersTelltheStory

0%
of AI startups fail within 3 years
$0B
spent on AI infrastructure in 2025
0x
revenue multiple for infra vs. app layer
0%
GPU utilization at hyperscale
◆ ◇ ◆

In a gold rush, sell shovels. In an AI rush, sell the silicon, the inference APIs, and the data pipelines.

Daniel ShanklinDirector of AI, AIC Holdings
● ○ ●

ApplicationLayervs.InfrastructureLayer

Application Layer

Winner-take-most dynamics — brutal competition
Dependent on distribution and timing
Models commoditize; differentiation erodes
High churn, low switching costs for users
Margins compress as foundation models improve

Infrastructure Layer

Oligopoly dynamics — high barriers to entry
Revenue grows with total AI compute demand
Deep moats: silicon, data centers, edge networks
Sticky enterprise contracts and ecosystems
Margins expand with scale and utilization

InfrastructureWinsAcrossEras

1995-2001

The Dot-Com Boom

Cisco, Sun Microsystems, and Akamai sold networking gear to every startup. Most startups died. Cisco became a $500B company.

2007-2012

The Mobile Revolution

Qualcomm and ARM licensed chips to every phone maker. Thousands of apps failed. The chip designers printed money.

2012-2020

The Cloud Transition

AWS, Azure, and GCP sold compute to everyone. Most SaaS companies struggled to reach profitability. The cloud providers became trillion-dollar businesses.

2023-Present

The AI Buildout

NVIDIA, hyperscale data centers, and inference API providers sell the picks and shovels. The application layer churns. The infrastructure layer compounds.

◆ ◇ ◆

InfrastructureRevenueGrowth(Indexed)

Chart 1
+🖌
Infrastructure Revenue Growth (Indexed)
7205764322881440
100
145
230
380
520
710
2021
2022
2023
2024
2025
2026E
● ○ ●
💡

The Contrarian Angle

The most common pushback: "But what if a foundation model company wins everything?" Even in that scenario, they need inference infrastructure, training compute, and data pipelines. The infrastructure spend doesn't go down — it concentrates. Which is even better for the pick-and-shovel thesis.

WhereWe'rePlacingBets

Three layers of the AI infrastructure stack look most compelling:

**1. Silicon & Compute** — Custom AI accelerators, GPU clusters, and the companies building next-gen training hardware. NVIDIA is the obvious play, but watch for Cerebras, Groq, and custom silicon from hyperscalers.

**2. Inference & Serving** — The API layer that turns models into products. As AI goes from "training big models" to "serving billions of requests," the inference economics become the bottleneck. Companies solving inference efficiency will capture enormous value.

**3. Data Infrastructure** — The unglamorous plumbing: vector databases, feature stores, data pipelines, and observability. Every AI application needs this layer, and it's deeply sticky once embedded in workflows.

◆ ◇ ◆

ReadtheFullAnalysis

This is the public summary. The detailed analysis with specific positions, risk factors, and target allocations is available to AIC Holdings partners.

Request Access