Malcolm Leeming is CEO of Filtered. Filtered is building the learning infrastructure for enterprise AI to make enterprise learning content usable in the flow of work.

I landed on a really interesting piece this week on Spotify’s “narrow moat”. The argument, in the way Morningstar-style moat analysis tends to make it, is that despite 600 million users, a dominant brand and the strongest personalisation data in audio, Spotify’s defensibility is structurally limited.

The reason is not scale. It is shape. Spotify shares its catalogue with Apple Music, Amazon Music and YouTube Music. The same songs are available on every platform. Pricing power sits upstream with the major labels – Universal Music Group, Sony Music Entertainment and Warner Music Group – who hold the content most listeners will not switch platforms to lose. Switching costs for a user are five minutes. And the network effects, while real in social and playlist features, are modest compared to, say, LinkedIn or WhatsApp.

A lot of scale and a narrow moat. It is a useful frame. Not because Filtered is Spotify – we are not, by a factor of millions of users – but because the diagnosis is about shape, and the shape of the moats available in the next wave of enterprise AI infrastructure is fundamentally different from the ones that defined the consumer internet.

Most of what makes Spotify’s moat narrow is structurally absent from what the best enterprise AI infrastructure companies are building. That is worth saying clearly, because the prevailing mental model for what a technology moat looks like – scale, network, and brand – is a consumer-internet model. It is not the only model, and it is not the one that will define who wins in enterprise AI. I think the moats that matter in this category rest on four structural properties. They are not about being bigger than Spotify. They are about being differently shaped.

Non-fungible content

Spotify’s core input is fungible. Every song is on every platform. The content cannot be a source of differentiation because it is also the competitor’s content.

The equivalent in enterprise AI is not fungible at all. A bank’s learning content, mapped to that bank’s skills framework, sitting against that bank’s entitlement graph, wired into that bank’s workflow surfaces, is unique to that bank. A late entrant cannot obtain it, because it was never on a shared shelf in the first place. The per-tenant architecture that frustrates cross-client network effects is the same architecture that makes the content non-fungible. In a consumer category this would be a disadvantage. In enterprise AI it is the beginning of a moat.

Rights-aware delivery

Spotify licenses music. Enterprise AI infrastructure licenses usage against boundaries that change by the day: tenant, user, role, entitlement, provider policy, and regional rights. A platform that can enforce those boundaries at the object level, at retrieval, at rendering, and at launch creates a substrate that a music streaming service does not have the shape to build.

Ripping out a rights engine is not a migration project. It is a rewiring project across every content provider and every workflow surface in the enterprise. That is a very different kind of switching cost from choosing a different music app.

Multi-sided attribution

Spotify is two-sided, creators and listeners, but asymmetric. Most creators are small and replaceable; the major labels are not replaceable but already sit upstream, capturing the pricing power. The platform in the middle is squeezed from both directions.

The equivalent in enterprise AI infrastructure has genuine symmetry. Enterprise buyers pay for outcomes. Content providers are paid when their content gets used. The platform that brokers the transaction is the only party in a position to generate the attribution data that both sides need in order to decide what to renew, what to buy, and what works. Attribution becomes a flywheel: more launches produce richer data, richer data produces better recommendations, and better recommendations produce more launches. Providers cannot get this data directly. Buyers cannot synthesise it internally. Neither side has a shortcut around the platform.

Workflow embedding as switching cost

Spotify users can switch apps in five minutes. Enterprise AI platforms do not work this way, and the serious ones will not try to. An infrastructure layer wired into Teams, Copilot, LMS and LXP launch paths and the skills graph of a 60,000-person organisation cannot be unpicked in a quarter. That is not a Spotify moat. That is closer to an ERP moat ; which is why SAP and Oracle are still alive in a cloud-native world.

What this adds up to

The question I would put to anyone evaluating an enterprise AI infrastructure company is not “can this be as big as Spotify?” The honest answer for most of us is no, and the economics do not require it. The question is whether the moat is deeper per unit of scale than the consumer-internet moats we have been trained to compare everything to.

On the four properties above, I think the answer for the best-designed enterprise AI infrastructure companies is clearly yes.

Four properties applied to enterprise learning content;

• Non-fungible per-tenant content
• Rights-aware delivery
• Multi-sided attribution
• Workflow embedding

That is a different shape of moat from the one Morningstar was analysing, and it is a shape that is only now becoming possible because MCP, context-aware retrieval and the wiring of AI into the flow of work are giving us substrates that did not exist two years ago.

At Filtered, we have been articulating our own category as ‘rights-aware learning orchestration infrastructure’. Strip away the technical language, and what the category is really about is those four properties applied to enterprise learning content, the only category of enterprise data where the provider universe is already mature, the content is already licensed, and the rights complexity is already legally real. It is a very specific place to stand. It is also the place where the four properties compound most quickly.

Narrow moats and massive scale can still build extraordinary companies. Spotify is one. But depth without consumer scale is enough if the substrate is right. That is the opportunity enterprise AI infrastructure has to build.

And it is an opportunity music streaming never had.