Recently, I was scrolling through TikTok when I saw a post about a Chipotle customer who pranked their ordering bot with a programming question. The bot answered it. It was such a LOL moment for me. Because obviously, when Chipotle first adopted AI, they probably didn’t think they had “a Nobel Prize Winner to take a Burrito Order.” From an economic point of view, Chipotle just paid the inference cost for something that wasn’t really meant to serve. While I laughed at the absurdity, there's a deeper issue here that reveals a fundamental flaw in how we build AI today.
Right now, we're in a moment where every entrepreneur building agentic systems is reaching for the same tool: frontier large language models. The logic is straightforward: use state-of-the-art models to solve your problem. And yes, some use cases genuinely need that power. But as I look at vertical AI platforms emerging across industries, I keep thinking the same thing: in five to ten years, many of these companies will realize they chose a wasteful, unwise approach. They're using a sledgehammer to hang a picture.
My point is, we don't need a Nobel Prize winner to operate the register at Chipotle. We don't need Frontier Intelligence to process an invoice, answer FAQs, or route a customer support ticket. Yet that's exactly what we're doing. We're deploying large general-purpose models trained on the collective knowledge of all humans to solve problems that require narrow, focused expertise. It's architecturally backward.
I would even argue that when entrepreneurs shift from general-purpose to purpose-driven architecture, it becomes a more natural way to solve problems. More of a first principles approach. With purpose-driven architecture, they don't address issues like privacy, efficiency, cost, and latency separately. They tackle them all simultaneously because they've addressed the fundamental design flaw. But there's something deeper: they can unlock proprietary intelligence. Instead of renting a frontier model that everyone else is also renting, they're creating systems that understand specific domains, workflows, and data. That makes their approach more defensible.
This isn't the first time I've seen this pattern. Computing has shifted between centralized and distributed architectures for decades: mainframes gave way to personal computers, and client-server moved to the cloud. Each change followed the same trend: centralization scales until it hits a limit, then distribution takes over. We're at that inflection point again. Today, the cloud architecture is powerful and dominant. But it quietly accumulates inefficiencies that will drive the next change. I've been noticing this for some time, which is why we've been investing in edge compute solutions for years. What's changed is that AI has raised the stakes significantly and made these inefficiencies more evident.
This doesn't mean the cloud architecture will become obsolete. Some workloads definitely belong in data centres, such as complex reasoning, multi-model orchestration, and tasks that require massive compute density. However, I believe a large portion of what we are currently running centrally will migrate closer to the edge, to devices and on-premises systems. Not because edge compute is trendy, but because purpose-driven models are smaller, faster, and make financial sense when deployed locally. The hybrid future isn't a compromise — it's the logical outcome of aligning architecture with real requirements.
I'm eager to see more vertical AI systems developed with purpose-driven models, proprietary enterprise-level intelligence, or smarter compute distribution. At least this is a more differentiated approach than the current methods used for agentic systems.
