Have We Been Here Before? A Thought on AI Infrastructure

I belong to the generation that still saw cables everywhere :) and in the embryonic days of the Internet, we weren't building flashy applications. We were deep in infrastructure and architecture, the unglamorous layers most people didn't think about, but where everything either worked… or didn't.

Very early on, I learned something that has stayed with me ever since: optimization always starts at the foundation. If the infrastructure is wrong, no amount of clever software on top will save you.

Fast forward a few decades, and we're living through another major shift: AI.

A Familiar Moment

What strikes me is how familiar this moment feels. Once again, the conversation is dominated by what's visible—models, demos, prompts, and a growing number of AI wrappers*—while a lot of the real leverage sits underneath. And hardware is back with AI infrastructures as important as models.

With XY.ai building agentic AI systems for healthcare operations, we're dealing with systems that don't live in clean, controlled environments. They operate across fragmented software, inconsistent data, compliance constraints, and real human workflows. To make that work reliably in production, you need more than intelligence—you need the right infrastructure.

Building with the Right Foundation

In this article with Together AI, my co-founder and CTO Lamara De Brouwer explains how we have built customer-specific EOB parsers with serverless fine-tuning thanks to their platform, providing us with high-performance computing and tooling that makes it easier for us to build, fine-tune, deploy, and run models at scale. Think of it as the engine, tools, and workshop that let teams move beyond experimentation into real products.

For companies building practical AI solutions like XY.ai, this matters because healthcare workflows—claims processing, document management, scheduling—live in messy, real-world systems. To automate these reliably at scale, you need a combination of:

• Scalable infrastructure
• Flexible model support
• Observability and performance

Those foundations have to be in place before you can focus on solving real operational problems instead of reinventing the basics.

From Experiments to AI at Work

The big shift for me has been how quickly we've moved from "AI experiments" to "AI at work." We're no longer in abstract ML research. This is AI operating inside real business workflows. With the right infrastructure in place, our team has been able to build agents that automate complex, multi-step tasks and deliver measurable ROI. That's the difference between prototypes and production value.

This matters because the right infrastructure unlocks:

• Faster time to impact
• Lower engineering friction
• Better control and performance

Especially in domains like healthcare, where complexity is real and outcomes matter.

The Operating Layer for Real-World AI

AI infrastructure platforms are becoming the operating layer that allows organizations to build AI that works in the real world—not just in demos. That's a trend worth paying attention to if you're curious about how AI moves from buzz to business impact.

We all know the famous words of Marc Andreessen from Andreessen Horowitz about software eating the world. With the big comeback of hardware, the world needs to be ready to digest even more software with AI coming en masse, especially with the cost of software nearing zero.

*An AI wrapper is usually just a call to an LLM API behind a chat interface. The intelligence lives entirely in the external model. There's little domain logic, memory, governance, or execution capability. These systems break easily and they fall short quickly in regulated environments.