A useful way to think about AI investing is to separate the stack into layers. The market often focuses on the obvious winners (NVIDIA, Microsoft, Amazon), but value can emerge in less crowded parts of the ecosystem.
| Layer | What drives value | Representative companies |
|---|
| Chips | Training/inference compute | NVIDIA, AMD, Broadcom |
| Cloud & infrastructure | Renting compute | Amazon, Microsoft, Alphabet |
| Networking | Moving AI data efficiently | Arista Networks, Cisco |
| Data centers & power | Physical infrastructure | Vertiv, Eaton |
| Software platforms | AI development tools | Databricks (private), Snowflake |
| Applications | End-user AI products | Hundreds of public and private companies |
The "picks and shovels" case
Historically, infrastructure providers have had three advantages:
- They benefit regardless of which AI model wins.
- Revenue tends to be more recurring and easier to forecast.
- Large customers often spend billions over multiple years once infrastructure is deployed.
That's why companies selling GPUs, networking, power equipment, cooling, memory, and optical components have generally captured a disproportionate share of AI profits so far.
However, infrastructure is becoming increasingly competitive. Margins may compress over time as custom chips, ASICs, and more efficient models reduce demand for premium hardware.
Application layer: harder, but potentially larger upside
Applications have a different risk/reward profile.
Many current AI apps have low switching costs. If every company can plug into similar foundation models, differentiation becomes challenging.
The winners are likely to have:
- proprietary data
- embedded workflows
- distribution advantages
- regulatory or industry expertise
Think less "AI chatbot" and more software that becomes indispensable to a specific profession.
Examples include vertical software for:
- healthcare documentation
- legal research
- industrial design
- drug discovery
- engineering
- cybersecurity
Non-obvious infrastructure plays
Some areas receive much less attention than GPUs.
1. Electrical equipment
AI data centers consume enormous amounts of electricity.
Companies involved in transformers, switchgear, breakers, and power management may see durable demand.
Examples:
- Eaton
- Schneider Electric
- ABB
2. Cooling
High-density racks require advanced liquid cooling.
Examples:
3. Optical networking
As GPU clusters scale, moving data efficiently becomes critical.
Examples:
4. Memory
AI models require massive high-bandwidth memory (HBM).
Examples:
- Micron Technology
- SK hynix
5. Foundries and packaging
Advanced chip packaging is becoming a bottleneck.
Examples:
- Taiwan Semiconductor Manufacturing Company
- Amkor Technology
Software infrastructure may be underrated
Rather than consumer AI apps, I think software infrastructure deserves attention:
- vector databases
- observability
- security
- inference optimization
- AI monitoring
- data governance
These businesses can develop sticky enterprise relationships without competing directly with foundation model providers.
Public-market examples include:
- Cloudflare
- Elastic
- MongoDB
A balanced approach
One framework is a "barbell" portfolio:
- 50–70% infrastructure: semiconductors, networking, power, cooling, foundries.
- 20–40% platform software: data, security, developer tools, enterprise AI infrastructure.
- 10–20% applications: selective companies with defensible distribution or proprietary data rather than general-purpose AI features.
This balances exposure to the near-term capital spending cycle with the possibility that long-term value shifts toward software as AI adoption matures.
One important risk to monitor is whether AI infrastructure spending continues to outpace monetization. If enterprise AI revenue grows more slowly than capital expenditures, infrastructure suppliers could face cyclical slowdowns. Conversely, if AI applications begin generating substantial economic value for customers, the application layer could eventually capture a larger share of industry profits—similar to how software ultimately became more valuable than much of the underlying internet infrastructure.
