AI’s Center of Gravity Is Shifting, and So Is the Money

Picture an AI-enabled service center of one of the world’s largest automakers. The system works beautifully. The moment a car pulls into the bay, the AI mechanic has already done its homework, sifting through reams of vehicle data to surface the exact service configuration it needs.

This elegant edge application took a small army of engineers three years to develop. It ran at 10 dealerships.

When the time came to scale it to 12,000 locations, the team discovered what every enterprise deploying edge AI eventually learns: building the application is the easy part. Scaling it to thousands of locations requires infrastructure that most of the industry is still figuring out.

That gap between building something smart and actually deploying it where business happens is the defining challenge of AI right now. And it’s reshaping where the next wave of technology investment is headed.

THE CLOUD BET IS CHANGING

For roughly 20 years, the enterprise software playbook was simple: move everything into the cloud. Centralize data and applications and control. It worked brilliantly for managing workflows, and investors rewarded it handsomely.

But the economics are shifting. SaaS revenue multiples, which peaked near 18–20x during the pandemic, have settled to levels not seen since 2016. Median revenue growth for public SaaS companies fell to around 12% by late 2025. Investors still haven’t lost interest in software, but they’ve started asking a harder question: where does AI actually generate returns?

Increasingly, the answer is not in the cloud.

THE INTELLIGENCE TRAP

Global AI investment is projected to reach $2.5 trillion by the end of 2026. The four largest U.S. tech companies alone are on track to spend roughly $725 billion on capital expenditures this year, up 77% year over year, and the largest single-year infrastructure buildout in the history of technology. That money is overwhelmingly flowing into cloud infrastructure to support data centers, GPUs and training clusters.

But most of the world’s operational data isn’t generated in data centers. It’s generated at retail locations, on factory floors, at oil rigs, inside hospitals, on shipping vessels and across energy grids. In those environments, decisions need to happen in milliseconds and sending data to a distant server simply isn’t viable.

I see this constantly in my work with enterprises. An oil and gas company will spend months building a predictive model for well optimization, then package it into a massive data file, ship it to a remote site and hope it works. A retailer will develop AI for inventory management, only to discover that rolling it out across thousands of stores is a completely different problem than building the model itself.

Intelligence is effectively held captive in the cloud. The physical world doesn’t wait for a round trip.

WHERE THE MONEY IS ACTUALLY GOING

Jensen Huang calls it “physical AI.” Qualcomm frames it as AI at the edge becoming essential. When the two companies that arguably understand AI hardware economics better than anyone are both pointing in the same direction, that’s worth paying attention to.

This shift is already happening on the ground. Enterprises across oil and gas, automotive, retail and industrial sectors are already running AI workloads at the edge. They know that edge AI works, but struggle with how to scale it. Managing thousands of deployments across diverse hardware and environments, and updating models in the field without disrupting operations presents a tremendous challenge.

The investment case is straightforward: when AI needs to operate autonomously—inspecting products on a manufacturing line, managing energy output at a wind farm, coordinating autonomous systems across a logistics network—it needs to run where the action is.

FROM APPLICATIONS TO AGENTS

There’s another shift happening beneath the infrastructure story, and it may turn out to matter even more.

Agents are changing the traditional rigidity of enterprise applications. An agent isn’t static code that does one thing. It’s a system that can reason, interpret unstructured data and take action without being explicitly programmed for every scenario. Instead of building a bespoke application to monitor quality on a manufacturing line, you deploy an agent that interprets video feeds, identifies defects and flags issues, all the while adapting as conditions change. Instead of coding a predictive maintenance system from scratch for each equipment type, you deploy a reasoning model that processes sensor data and makes decisions in real time.

Where it once took years and teams of developers to build edge applications, agents can collapse that timeline dramatically.

WHY THIS ISN’T SIMPLE

But as in most instances, the devil is in the details. When a personal AI agent makes a mistake, it sends an awkward email. When an industrial agent makes a mistake, it can shut down a production line or create a safety hazard. The same quality that makes AI agents flexible and powerful is exactly what makes enterprises cautious about deploying them in safety-critical environments.

The IP sensitivity alone is staggering. An energy services company that has spent a decade gathering operational data to build an autonomous drilling optimization system has created something irreplaceable. If that agent is compromised, a competitor effectively acquires years of institutional knowledge overnight. The EU’s Cyber Resilience Act and evolving NIST frameworks are moving to address these risks, but the technology infrastructure to enforce guardrails at the edge is still being built.

Security and governance aren’t features for the edge AI era. They’re the foundation.

THE NEXT INFRASTRUCTURE BUILDOUT

Every major era of computing has been defined by an infrastructure buildout. The PC era built local networks and desktops. The cloud era built data centers and SaaS platforms. The AI era is building the infrastructure to run intelligence in the physical world.

The companies and investors who recognize this shift early will be positioned for the next decade. That means not just investing in AI capability, but in the orchestration, security and deployment infrastructure that turns a capable model into a reliable industrial system.

The value isn’t in the next chatbot or the next cloud-hosted model. It’s in the infrastructure that makes AI operational at scale in the factory, the oil field, the retail network and everywhere else the physical world generates data that can’t afford to wait.

Somewhere right now, another company is standing where that automaker stood, with a brilliant application running at a handful of sites, staring at the chasm between pilot and scale. The ones who bridge that gap will build the competitive moats of the next decade.