2026 Is Not the Year of AI. It's the Year of AI Plumbing.

It’s January, which means my feed is full of AI predictions for 2026. AGI by December. Humanoid robots in every warehouse. Autonomous AI employees replacing half the workforce. World models that simulate reality in real time.

I have a different prediction. 2026 will be the year of AI plumbing.

Not because the flashy stuff won’t happen. Some of it will. But because the companies that win in 2026 won’t be the ones with the most impressive models or the most ambitious roadmaps. They’ll be the ones with the best infrastructure for AI to actually do work.

The Infrastructure Gap

When a new technology emerges, the first wave of investment goes into the technology itself: the models, the algorithms, the core capabilities. The second wave goes into the infrastructure that makes the technology usable in production: the tooling, the platforms, the connective tissue.

We’re at the transition point between wave one and wave two.

I’ve lived through this transition in other technology cycles. When I was building the IoT platform at Xively, the first wave was all about getting sensors connected and data flowing. Cool demos everywhere. But the companies that actually built lasting IoT businesses were the ones who invested in the boring stuff: device management, firmware updates, data pipelines, alerting, and monitoring. The plumbing.

Same thing happened with cloud computing. The first wave was about elastic compute and storage. The second wave, the one that generated real enterprise value, was about IAM, VPCs, logging, compliance tooling, and cost management. The plumbing.

AI is following the same arc. The models are good enough. Claude, GPT, Gemini can all do remarkable things. The question is no longer “can AI do this task?” For most enterprise tasks, the answer is yes. The question is “can I deploy this AI in a way that’s secure, observable, reliable, and governable?” For most enterprises, the answer is: not yet.

That gap is a plumbing problem.

The Plumbing Stack

Here’s what the AI plumbing stack looks like:

Authentication and authorization. When an AI agent acts on behalf of a user, how do you verify identity? How do you scope permissions? How do you ensure the agent can only access data and tools the user is authorized to use?

Context management. Models have finite context windows. Agents need access to vast amounts of information. How do you select, prioritize, and manage the context that flows into the model? How do you prevent sensitive information from leaking across contexts?

Observability. When an agent takes an action, can you trace exactly what happened? What context did it have? What tools did it invoke? What reasoning led to its decision?

Rate limiting and cost controls. AI inference costs money. Without controls, a runaway agent can burn through your budget in minutes. You need the same circuit-breaking infrastructure that you’d build for any production service.

Error handling and fallback. What happens when the model returns garbage? When the tool call fails? When the context is corrupted? You need graceful degradation paths that protect the user from AI failures.

Audit and compliance. Every action an AI agent takes needs to be logged, timestamped, and available for audit. This is a net-new engineering challenge that most organizations haven’t started thinking about.

Where I’m Investing

At Vestmark, this is where the majority of our AI engineering effort goes. Our internal platform handles authentication through Okta OIDC, runs workloads on ECS Fargate with scale-to-zero efficiency, and uses SQLite on EFS for simplicity and reliability. Every component was chosen for operational clarity, not architectural novelty.

For our AI products, we’ve built observability into every layer. Every inference call is logged. Every tool invocation is traced. Every AI-generated output is associated with the context that produced it. When someone asks “why did the system do this?” we can reconstruct the entire decision chain.

This work is invisible to the end user, and that’s exactly the point. Good plumbing is plumbing you never think about.

Why This Matters More Than AGI Predictions

The AGI debate is intellectually interesting but practically irrelevant to what most organizations need to do right now. Whether AGI arrives in 2026 or 2036 doesn’t change the fact that today’s AI models are useful and today’s infrastructure for deploying them is immature.

The companies that will lead in 2026 are the ones building the infrastructure that lets current models work reliably, safely, and at scale.

This is unsexy work. It doesn’t make headlines. But it’s the work that separates demos from products, experiments from deployments, and potential from impact.

2026 is the year of AI plumbing. And I wouldn’t have it any other way.