In the Agentic Era, should we become a dumb pipe?

A couple years ago, I read a fantastic article from Matt Rickard. In the article he concludes that AWS has the might and will to specialize in, at least to a degree, applied services. Basically, while AWS may not build character.ai, they could have possibly built persona tuning ai platform, if they wanted to. There are many nuances to this though, as seen in Google Graveyards.

Today I had an opportunity to ask the CEO of TwelveLabs the exact same question (disclaimer: I work at TwelveLabs).

As a video intelligence foundation company, why do we start to investigate on applied services?

I basically connotated that

Wouldn't it be better to just become a dumb pipe company?

Here, I want to clarify that dumb doesn't necessarily mean bad business. Nor it means upper bounded. It just means that a sufficiently fundamental service is so ubiquitously desired that economically doesn't make sense to spend your resource on others. The ambition is basically

I don't care whatever the fuck you're building, but if you deal with video, you're certainly gonna need my API.

And these usually come in a universal API format, like S3 SDK supported by base URL override by others vendors, or OpenAI SDK supported by many open labs.

So in a technical sense, it becomes value creation v.s. value capture and when you own the primitives, it's always easier for you to create more value instead of competing to capture the same pie.

... and that was what I always thought. AWS provides myriads of applied services but their cashflow always come from their most fundamental services, and Anthropic and OpenAI try to compete with applied services but their primary source of income are B2B APIs, which, ironically builds coding harnesses.

It's like Samsung. They are the best component manufacturer for iPhone, but they also build Galaxy smartphones. It's a very delicate competitor-ship that, you still want some pie from Galaxy, but you definitely don't want to ruin your partnership with Apple.

And I was arguing that we become TSMC. Just become the best fucking pipe company (which TwelveLabs was good at), and why bother applications? They're all gonna use our APIs anyways

Though, there are some counterexamples, for example neoclouds, especially 1st gen neoclouds. Matt argued that Heroku failed because it struggled to supply the surrounding products as customer matured (lacking adjacent layers), thus, owning an stranded position.

Vercel is quite the opposite - they own AI gateway, AI SDK, workflows, sandbox, all sort of telemetry tools... (though I don't think that's why Vercel succeeded. They nailed owning the primitives so well, then later expanded)

Anyways.

The modern AI layering is particularly interesting and different.

AI is destructive, because it collapses abstractions. Previously layered software becomes one giant melting pot with the name "agent". If you squint, every sophisticated software system can be done by one agent given enough tokens and contexts...

So software services becomes extremely vulnerable, also known as SaaSpocalypse (which, I don't believe, still)

But still the modern version becomes

Become a primitive that agents compose

or

Own enough domain semantics that the agent cannot trivially reconstruct your value.

and that was the borderline I was drawing.

TwelveLabs's CEO, Jae, described something more delicate.

He described Uber, where Uber selectively applied its fleet management proficiency to high-reward verticals, like medicine delivery or food delivery.

They vertical'ed out, but it's all built on top of their fleet management software.

Although I think it's a little different, because, fleet management is not something you'd API-sell. It's something that only Uber could do.