Hi {{first_name | there}},

I run through Berkeley Hills most afternoons, and I never take the same route twice. Same hills, different turns every time. What changes is what's growing in front of every house. 

There's a yard where white flowers spill over the fence line, and if you catch it at the right angle you can see the Bay and the Golden Gate through the gaps. Two doors down, someone's been collecting succulents for years: black rosettes bred almost purple-black, planted right next to the plain green ones like a deliberate argument. A few blocks over there's a street sign that says HIDDEN, bolted to a post in the wildest, least hidden yard on the whole route. A deer stands in a yucca hedge some afternoons, ears up, watching me watch it, like it's paying rent there.

Hundreds of yards. I have never seen two that look the same. Nobody planned it that way. Hundreds of people made hundreds of specific decisions about their own small piece of land, most of them bad decisions by any professional landscaper's standard, and what came out the other side is more alive than anything one planner could have designed on purpose. That's abundance. Real abundance. And abundance like that only ever comes from diversity. Drive twenty minutes to a "PE slop" subdivision, the same five house plans copy-pasted down a hillside, and you get the opposite. Less alive, no matter how much money went into it.

I've built a company betting that AI compounds power into a small number of hands. Bigger models trained on more data get better faster than smaller ones can catch up. Whoever owns the compute and the distribution owns the default. That's a network effect, and network effects are not polite. They don't stop at "good enough consolidation." They keep going until one or two players own the substrate everyone else has to build on.

You can see a rough version of it in writing itself, if you look. Feed enough different people through the same handful of models and you start getting the same rhythms back: the same tidy contrasts, the same reflective openers, the same safe, polished voice showing up no matter who's actually typing. That's the stark version of the risk, not just fewer companies, but fewer ways of thinking and writing, dressed up as more of both.

The question I actually keep asking myself is more specific than "will AI make things generic," and harder to dodge: if the infrastructure keeps consolidating into a handful of hyperscalers the way it obviously is, does everything built on top of it get forced into the same shape too? Or is there a version of this where the substrate consolidates and the thing growing out of it still gets to be hundreds of different yards?

That's the actual fork in the road for the next decade of this industry, and most of the AI commentary I read skips right past it to get to the exciting part.

I went looking for an answer, and the picture turned out to be more specific than "diversity wins" or "consolidation wins."

The infrastructure layer really is consolidating. Hard, exactly the way the network-effects logic predicts. A handful of hyperscalers and model labs are going to own the substrate: the compute, the base models, the distribution rails everyone else's software has to sit on top of. Nobody's arguing that part. It's the shared hillside, the same soil and water table every yard on my street sits on top of, whether the owner ever thinks about it or not.

What actually surprised me, once I looked at where deployments are winning and where they're dying: no single deployment framework, not one, holds even 30% share of production AI deployments. The middle of the market, the generic, copy-paste, same-five-plans implementation, is exactly the segment losing ground fastest right now, not the segment consolidating. What's compounding instead is a long tail of operators who went deep on one specific domain, one specific kind of operational mess, and built something that only works because it was built for that mess specifically.

That's the actual shape of it: a barbell, not a funnel. A few horizontal giants own the shared substrate. A long tail of specific, deep operators owns everything that has to be built to fit an actual business instead of a generic one. The undifferentiated middle, the stuff trying to be broadly good enough for everyone, is the part that dies and the piece losing share fastest. The diversity on either side of it is what's compounding instead, and it's not a new pattern. Diverse systems, in ecosystems, in markets, in whole organizations, consistently out-innovate and outlast uniform ones. It's one of the more well-established findings there is, across way more fields than just this one. AI deployment is just the newest place it's showing up, and it won't be the last.

That's the closest thing to an answer I've found to the question I keep asking myself on that run. The hyperscalers get the hillside. Nobody gets to skip that, and honestly nobody should even want to, the same way nobody on my street is out there trying to dig their own well. But what grows on top of the hillside is still, stubbornly, hundreds of different decisions, because the alternative, one generic deployment forced onto every business regardless of what it actually does, is the thing that's provably losing right now, not winning.

I'll admit the self-interest here plainly. Future Works is built on getting the harmony between a business and its own AI right, specific to that business, instead of shipping one deployment plan copy-pasted across clients. That's the first of the two stages we've built the company around. What we do with the upside once that stage is actually working is a different conversation, for a different post. If the barbell picture is right, this first stage gets more valuable over time. If I'm wrong about it, I'd rather find out from someone arguing with me than from a client leaving.

The same PE playbook that turns five house plans into a subdivision shows up in this industry too: buying up fragmented service firms, standardizing them for margin. Same move. Different kind of fragmented land.

The barbell might not hold ten years out, at the scale the hyperscalers are playing at. It holds on my street this afternoon, in front of a sign that says HIDDEN, in the one yard on the block hiding absolutely nothing.

If you're deploying AI into your own business right now, here's the same test I run on my own street: if a competitor in your exact vertical could take your AI program, unchanged, and it would work just as well for them, what does that actually tell you? Is what you're running built for the specific mess your business has, the one nobody else has in exactly that shape? Or did you buy the five house plans and call it transformation?

The defense is the same whether you're running a business or writing a paragraph, my friend. Stay specific. Stay personal. Keep looking at the problem your own way instead of borrowing someone else's answer to it. That's what keeps a street, a company, or a piece of writing from turning into the same five house plans.

Tell me where you land. I'd genuinely like to know if you're seeing the same barbell from where you sit, or something else entirely.

Much Love,
Matt

At Lighthouse, we love featuring fresh perspectives from our community of AI, tech, and innovation leaders. Got insights to share? Just reply to this email—I’d love to hear from you!