The Pendulum

07 Mar, 2026

Six years ago, a vice president at my company had a realization. If we were going to stay competitive in the artificial intelligence space, we needed GPUs. A lot of them. And the worst possible version of that procurement was letting every business unit across the corporation go buy their own, fragmenting a massive capital investment into a dozen underfunded efforts that would each independently struggle to reach critical mass.

So they centralized. They stood up an organization, literally called the Lockheed AI Center, LAIC for short, and gave it a mandate: procure the hardware, build the platform, attract the best and brightest ML practitioners across the entire corporation, and bring a shared vision to life.

Centralization made sense when the technology was immature and the talent was scarce.

That’s the first principle here. When GPUs are expensive, when ML expertise is concentrated in a handful of practitioners, when the tooling doesn’t exist yet and needs to be purpose-built, you want your best people in one room, mission-oriented, working together. You don’t want ten teams across ten business units each trying to figure out distributed training from scratch. Centralization was the right call.

That was five years ago.

I’ve spent a large part of my career inside this centralized org, providing AI development services to other teams and business units. A big portion of that job is being a power user of the internal MLOps platform that an infrastructure team adjacent to my AI consulting team developed. At times my role feels like a glorified software testing position. I’ve had a front row seat to both the brilliance and the dysfunction of concentrating this much capability under one roof.

And I think the pendulum has swung the other way.

As AI tools become more capable, the need for centralized expertise diminishes. As tools like Claude and Codex become more democratized, the power distributes itself. It shifts from the centralized org to individual contributors within the company who are equipped with the knowledge to use these tools effectively. This operational reality favors decentralization. I don’t really need a rock star team of AI consultants to help solve my business problem if I’m equipped with the knowledge to prompt Claude or Codex appropriately.

There have been many times where I have looked at the concentration of talent within my centralized org and wondered how much more effective we’d be, how much more value the organization would extract from our people, if we were broken apart and distributed amongst individual business units and teams throughout the company.

Many of the tools we tried to build over the last five years have been superseded by commercially available alternatives. That’s not a failure of the people who built them. It’s a reflection of how fast the landscape moved underneath all of us.

I do think there is a place for a centralized infrastructure team to monitor platform stability and ensure that the operation runs smoothly. The hardware still needs management. The platform still needs upkeep. But the consulting model, the “come to us with your AI problem and we’ll solve it for you” model, feels increasingly misaligned with where the technology actually is.

When you centralize development of a product, you are naturally constrained by the collective imagination of the team tasked with doing the development.

This was a source of real tension for me this past week. I submitted what I consider to be a key feature request to our internal platform team and was met with resistance. Not constructive pushback, not “here’s why that’s technically difficult,” but patronizing commentary about how I was not using the product as it was designed.

There’s great power in being a centralized organization. And that power can get to people’s heads.

I’ve seen it lead to inflexible ways of reacting to customer needs. The further the distance between the people building the product and the people with boots on the ground using the product, the further the customer needs get pushed away from the product’s core development priorities. Eating your own dog food has its advantages. When you don’t eat it, you lose the feedback loop that keeps you honest about what you’re actually serving people.

I find myself wondering if this is why so many ML platforms eventually become open source products, spun off with their own foundation to collect feedback from the masses. The reach of open source is undeniable, but that reach comes with access to a much wider range of diverse backgrounds and use cases for the thing you build. I believe that diversity of background and equity in access leads to more actionable feedback to create a better product. Especially when compared to heavily weighting the opinions of a few over the consensus of the many.

There’s a human element to this that I don’t want to gloss over.

I’ve observed time and time again that the easiest way to become irreplaceable as a leader is to begin managing a very large organization. As your budget grows, so does the hassle of trying to shake up the organization. There are leaders whose incentive structure rewards growing the team, not shrinking it. Not distributing the talent where it would be most effective, but consolidating it where it justifies the leader’s existence.

I’m a believer in Jeff Bezos’ rule about the two-pizza team: “if you cannot feed the team with two pizzas, the team is too large”.

Taking it a step further, giving the team real authority through a decentralized command structure is something I’ve gathered the elite special forces units within the United States military operate by. Small, autonomous teams with clear objectives and the latitude to execute. Not large, bureaucratic organizations that need seventeen approvals to change a button color on a dashboard.

I’ve written about this before, but I think there is a war being waged for the soul of AI.

It’s a fundamental tension. On one side, AI is the ultimate tool for centralized command to use to influence the masses. On the other, it’s the most democratized information-generating tool that humanity has ever created, something that levels the playing field in a way that has never been observed before.

That tension plays out at the macro level in society. It also plays out at the micro level inside a corporation. The centralized org wants to remain the gatekeeper. The individual contributors want the tools in their own hands. Both have legitimate arguments. Both have blind spots.

Experience has taught me that there is a time and a place for each of these approaches. One is not absolutely preferred over the other. Five years ago, centralization was the right call. Today, I think the evidence points towards distributing the capability closer to the people who need it.

The pendulum swings. The question is whether the people holding it are willing to let go.