#119- Where I’d invest in AI infra today
If I had $5mn and a clean slate
Over the last few weeks, I’ve written about how AI infra actually gets built and run:
Part 1: Data Preprocessing- clean input is the foundation
Part 2: Backup & Recovery- disaster-proofing AI systems
Part 3: Deployment & Monitoring- keeping models alive in the wild
Part 4: The 5-Point AI Infra Filter- how I spot weak wrappers vs real infra
But here’s the obvious follow-up… If I had to deploy $5mn today into AI infra startups, what exactly would I back?
Here are 3 wedges I’d bet on. Practical, fundable, and built for what real companies need right now.
1. Model monitoring that catches the problem before the KPI drops
What is it?
A dashboard that tells a company that your model’s going off chart, here’s where, here’s why. Right now, most teams realize this only when sales dip or a user complains
Why it matters?
AI is now in live systems, pricing, fraud, search. If it starts doing weird things, and you catch it late, the damage is real
What I’d back?
Plug-and-play drift monitoring
Clear alerts. No ML jargon
Works across both classic models and LLMs
2. Multi-model switchboard: let users route, compare, and control
What is it?
One tool to run multiple LLMs (GPT4, Claude, Gemini, etc.) side by side. Let teams decide which to use, switch on the fly, and track cost/performance
Why it matters?
Most companies are locked to one provider (usually OpenAI, which is now at $10bn ARR thanks to this). That’s dangerous. If price spikes or performance drops, they’re stuck
What I’d back?
Routing and fallback logic built-in
Cost and quality comparison dashboards
Easy enough for business teams to use without an engineer
3. Governance & guardrails for AI usage inside companies
What is it?
An internal control panel that shows who is using which AI tools, with what data, and under what guardrails. Imagine if IT, legal, and data teams could see, allow, or restrict usage in real-time, all without blocking innovation. I will write about this in greater detail soon.
Why it matters?
Most companies are blind to how their employees are using AI tools. Sensitive data is being sent to public APIs. Models are hallucinating without oversight. And once regulators start asking questions (which they will, soon), there’s no audit trail
What I’d back?
One pane of glass to track model usage and enforce access policies
Granular controls (team, app, data type)
Easy to deploy across tools like Notion, Salesforce, internal apps
What all 3 have in common
They solve today’s problems, not some 5 year futuristic science fiction
They support the ops, data, and IT users who actually run the business
They’re easy to starts, and very very hard to rip out once embedded
If you’re building something in this space, especially model access, monitoring, or control, I want to hear from you. These are the infra companies that last!