For most of the 2010s, the IoT industry had one overriding strategic obsession: ecosystem ownership.

I know because I was in it. At ProSyst, then at Bosch, and in the standards bodies where these things get decided — OSGi, the Industrial Internet Consortium, Eclipse IoT. The technical debates were real. But underneath them was always the same commercial question: who controls the interface?

The Fragmentation Problem — And What Actually Caused It

When people talk about IoT fragmentation, they usually reach for the protocol list. And the list is real.

At the connectivity layer, you chose between Zigbee, Z-Wave, Thread, LoRa or a handful of cellular options. At the messaging layer, MQTT, AMQP, CoAP or plain HTTP. At the device management layer, OMA-DM or LwM2M. And at the platform and integration layer, OPC-UA for industrial environments, oneM2M as a telco-driven service layer, FIWARE pushing its own data models for smart cities, OSGi as a component framework for gateway runtimes. Just to name a few.

Every layer, competing standards. Every standard backed by companies with commercial interests in winning.

But here's the thing: the protocol layer was never actually the hardest problem.

Two devices both speaking MQTT can still be completely incompatible — because one sends temperature as {"temp": 21.5, "unit": "C"} and the other sends {"sensor_value": 21.5, "type": "temperature_celsius"}. You solved the transport. You haven't solved anything useful yet. The real problem was semantic interoperability — shared data models, shared meaning, shared context. And no protocol solves that.

What actually caused the fragmentation wasn't a missing standard. It was a governance failure. The economic incentive for every large company was to own the semantic layer — because whoever defines how data is structured and interpreted controls the application ecosystem above it. Openness was the marketing message. Control was the business model.

And it worked, for a while. For the large players.

Standards Existed. They Just Didn't Win.

This is worth saying clearly: the IoT industry was not short of standardisation attempts. MQTT has existed since 1999. CoAP was IETF-standardised. oneM2M was designed from the ground up as a universal IoT service layer. The W3C Web of Things working group tried to define semantic models. Eclipse IoT built open-source implementations of almost everything.

None of them really "won" in any decisive way. Not because they were technically inferior. But because adoption follows power, not merit. MQTT became dominant in cloud-connected IoT largely because AWS adopted it. OPC-UA became the industrial standard partly because Siemens and other large vendors backed it hard. The standard that wins is usually the one with the biggest commercial engine behind it.

Which brings me to MCP.

What MCP Does — And the Uncomfortable Question Behind It

In early 2024, Anthropic published the Model Context Protocol — a standardised interface through which AI models can connect to external tools, data sources and services. Clean, open, well-documented. The idea: one protocol, works everywhere, no vendor lock-in.

My first reaction as someone who spent years in IoT standardisation: this is elegant. My second reaction: I've seen this movie before.

But what makes MCP genuinely different from the IoT standardisation attempts is the speed of industry alignment. OpenAI has added MCP support to its Agents SDK. Microsoft, AWS and Google are moving in the same direction. In the IoT world, a competitor adopting your protocol within months of release would have been unthinkable — everyone was too invested in owning their own interface layer.

Whether this convergence holds as the commercial stakes rise is still an open question. But the early signal is different from anything we saw in IoT.

What IoT Actually Learned (Eventually)

I want to be fair here: the IoT industry did figure some of this out. Just slowly.

Matter, launched in 2022, is doing for smart home devices what a universal IoT protocol should have done a decade earlier. OPC-UA combined with Sparkplug B has emerged as a credible standard for industrial environments. The semantic layer problems are still unsolved at scale — but the connectivity layer has converged more than the early 2010s debates would have suggested.

The lesson wasn't that standards don't work. The lesson was that standards only win when the governance model is right and when the commercial incentives of the major players align — or when someone powerful enough forces alignment.

So What Does This Mean for Physical AI?

IoT and AI are converging. Edge AI, Physical AI, intelligent products that don't just collect data but act on it — this is the next wave. And it brings the same foundational question back to the surface: will we build on shared, open foundations, or will we repeat the platform wars?

MCP as a concept gives me genuine reason for optimism. The idea of a neutral interface layer — one that separates "how things connect" from "what value is built on top" — is exactly the right architectural instinct. Whether MCP itself survives as the standard, or whether it gets fragmented by the same commercial forces that fractured IoT, I honestly don't know.

What I do know is that the Physical AI layer — where AI models interact with the real world through sensors, actuators, devices and edge systems — needs this problem solved at the semantic level, not just the transport level. Connecting an AI model to a factory sensor via a standardised protocol is step one. Ensuring the AI model understands what that sensor is measuring, in what context, with what confidence — that's the harder problem. And it's the one the IoT industry never fully cracked.

That's the conversation I think we need to be having now — before the ecosystem wars start again.

I'd love to hear from the IoT and AI community on this — and I'm genuinely asking, not just posting.

Those of you who were in the platform wars of the 2010s: where do you think the real lesson is? Was it about protocols, governance, or commercial incentives?

And for those building Physical AI and connected intelligence today: are we heading into the same traps — or has something structurally changed that makes this time different?

Drop your perspective in the comments. I don't think there's a clean answer here, and I'm curious what the people who actually built this stuff think.