The IoT Platform Graveyard, and the Position Nobody Has Claimed

The Graveyard

#

Around 2015, at the peak of the hype, analysts were tracking the IoT platform market like a gold rush. IoT Analytics counted 260 platforms in 2015, 450 in 2017, and 620 at the peak in 2019. McKinsey had promised an $11 trillion opportunity. Everyone wanted a shovel.

Then the burials started.

• Google Cloud IoT Core — retirement announced August 2022, switched off August 16, 2023. Six years from its debut at Google I/O in May 2017, and only five and a half of those generally available.
• IBM Watson IoT Platform — sunset December 1, 2023, “without a direct replacement”.
• SAP Leonardo — the brand was quietly dismantled years before the underlying IoT service was retired. SAP now embeds Cumulocity — its ex-rival’s platform — inside its own asset management product.
• Bosch IoT Device Management — discontinued by mid 2024, citing “the market trend towards cloud providers”.
• Azure IoT Central — the strangest entry on this list, because it is not a burial. A February 2024 system message announced retirement for March 31, 2027, with no new applications from that April. Days later Microsoft’s general manager of Azure IoT called that message “not accurate” and “presented in error” — and said nothing about the roadmap. Two years on there is still no committed future for it, and Microsoft is simultaneously a Leader in Gartner’s 2025 Industrial IoT Magic Quadrant. A platform nobody can get a straight answer about is its own kind of dead.
• Relayr — the startup Munich Re bought for roughly $300M was wound down in 2025.

And the survivors are changing hands: Cumulocity’s founders bought it back from Software AG in January 2025, PTC sold ThingWorx and Kepware to TPG in late 2025, and SUSE acquired Losant in February 2026.

So, fifteen years after the hype: is there room for a new IoT platform, or is that graveyard telling everyone to go home? I spent a week doing the homework — pulling the actual pricing pages, reading the actual docs, checking what the AI features actually are. This post is the result. My conclusion up front: there is room, but only because the survivors converged on the same three weaknesses, and none of them can fix those weaknesses without rebuilding.

Full disclosure before we go on: I am the CTO of Amplía Soluciones, whose OpenGate platform lives in this market. That is why OpenGate is not in the table below — I’m not going to grade my own homework, and its commercial terms aren’t mine to publish. Every other number here comes from the vendor’s own public page as of July 2026, linked so you can check me.

What the Survivors Charge

#

Two absences in that table are deliberate. OpenGate is out for the reason above. And so are the mature enterprise platforms sold consultatively — the ones with industrial edge protocols, SCADA integration and a delivery team attached. Their price buys fifteen years of operational scar tissue in utilities and telco, and putting their terms next to Datacake’s would be dishonest in both directions. Everything from here on is about the segment underneath that league: the projects that have to work without an integration project around them.

Three patterns jump out of what’s left.

First: the dominant billing models punish volume. Per-message, per-payload, per-data-point — the moment your fleet actually starts producing telemetry at IoT scale, the meter spins. Losant’s overage is $100 per 100k payloads. Ubidots charges $5 per million data points ingested and then again on the way out. This is pricing designed for the demo, not for the deployment.

Second: opacity starts one tier up. Almost every vendor publishes exactly one price — the entry tier — and hides everything else behind “contact sales”. Cumulocity publishes €215/month and nothing else. Ubidots’ Industrial tier used to be a public $499; today it’s “Custom”. Azure IoT Operations’ pricing page literally displays $-/hour. If you cannot predict your bill from the public page, the public page is marketing, not pricing.

Third: features you already paid for come back as add-ons. Datacake — otherwise one of the most transparent vendors on the list — sells rules in expansion packs (€29 to €149/month) and extended retention separately (€39 to €165/month). ThingsBoard’s analytics live in a separate add-on called Trendz at $29/month. thinger.io doesn’t do overage at all: when you hit your device cap, provisioning simply stops until you migrate to a bigger instance.

What Their “AI” Actually Is

#

Every platform in that table now says “AI” somewhere on its homepage. Here is what the term resolves to when you read the docs, with dates:

• Cumulocity shipped an AI Agent Manager and an AI Agent block for its analytics builder in early 2026 — agent workflows layered on top of a platform whose core dates to ~2010.
• ThingsBoard added AI to its rule engine in 4.2 (2025); anomaly detection and prediction live in Trendz, the paid add-on, plus an “AI Credits Pack” at $5/month.
• Ubidots has simple built-in anomaly detection; anything more is Python you deploy yourself into their serverless runtime — on the Enterprise tier.
• Datacake deserves honest credit: its AI Assistant (May 2026) is the most developed copilot in the lean segment — OpenAI-backed, function-calling against their API, with write-confirmation guardrails. It costs €20 per user per month after a free token allowance.

Notice the shape of every single one of these: a copilot or an ML add-on, bolted onto a platform that was designed before any of this existed, and usually sold separately. The AI is a feature inside their walls, operating their UI, priced per seat.

I also went looking for the contrarian case — a post-2023 platform genuinely born AI-native, with agents or MCP in its core design. I could not find one. The closest things are Espressif’s private voice-agent platform for ESP32 devices and AI blocks on top of MQTT brokers. In 2026, “AI-native IoT” in vendor material almost always means edge inference on the device, not an AI-operable control plane. That position is, as far as I can verify, unclaimed.

Two Things That Cut the Other Way

#

Two findings argue against the comfortable version of this story, and they belong in the post.

ThingsBoard Community Edition is genuinely free — Apache 2.0, self-hosted, unlimited devices, and mature. If your only axis is license cost, you will not beat it, and I won’t pretend otherwise. What free software doesn’t eliminate is the cost of operating it: the cluster, the queue, the upgrades, the add-ons when you need analytics, the person who owns all of that. That is a real bill; it just doesn’t arrive as an invoice.

Datacake’s copilot is good. The lean segment is not asleep on AI. But a captive per-seat copilot inside one vendor’s walls is a different animal from a platform any agent can operate end to end. Which brings me to the actual argument.

What “Born for AI” Would Actually Mean

#

If you designed an IoT platform today, from scratch, knowing that AI agents exist — what would be different? Not “it has a copilot”. I think it resolves to three properties, and the useful thing about them is that they’re checkable from the outside.

1. The whole control plane is agent-operable, by construction. Not a curated copilot with seven skills and a per-seat price: every operation the platform has, reachable by whatever assistant the user already runs, under their own role and tenant, with confirmation gates on anything destructive. You get there by generation rather than by discipline — one operation registry from which the REST API, the CLI and an MCP server are all produced, with parity checks in CI, so a capability cannot exist in one surface and be missing from another. I’ve written about what that design buys and what it costs in The CLI Wasn’t Built for Me. It Was Built for the AI. That one was a tool built on top of an existing industrial platform, which is exactly the point worth making here: retrofitting gets you a good tool. It doesn’t get you an agent-operable platform.

2. The intelligence lives in the data path, not beside it. Telemetry scored inside the ingest pipeline, the verdict landing as ordinary fields that rules react to like any other — and never a naked score: a verdict with calibrated uncertainty, an out-of-distribution flag, per-feature attribution. No new query language, no analytics add-on. The distinction isn’t cosmetic. If the model sits beside the platform, somebody has to own the export, the training, the deployment, the call and the merge back, plus keeping the schemas in sync and noticing when it silently stops working. That somebody bills by the day, and their invoice is the real cost of AI in IoT today.

The second arrangement isn’t a cheaper version of the first. The pipeline the first one needs simply doesn’t exist.

3. The training loop closes on the platform’s own data. Your telemetry is already in there. So the platform’s own job engine should be able to export it, train a model on your normal, and hot-reload the result — scheduled, audited, repeatable, an operation like any other. Not “integrates with SageMaker”. This is the property I found missing absolutely everywhere, and it’s the one that decides whether the intelligence is yours or rented.

And underneath the three, the property that can’t be retrofitted at all: operational footprint. Volume-punishing pricing isn’t greed, it’s a pass-through — brokers, queues, coordinators and the cluster to run them get paid for by somebody. A platform whose runtime is small enough to serve a modest fleet from a modest machine is the only kind that can publish a low entry price and a complete public rate card without lying about one of them. That’s an architecture decision made on day one, not a pricing decision made later.

The Verdict on the Hypothesis

#

Fifteen years after the hype, the giants have retreated, the survivors keep changing hands, and the ones left standing share three structural weaknesses: volume-punishing prices, opacity one tier up, and AI as a bolted-on, separately monetized copilot. Meanwhile the position “IoT platform designed for the age of agents” sits unclaimed — not because it’s hard to describe, but because it isn’t reachable from where the incumbents stand.

That’s the part worth sitting with. Pricing model, operational footprint, and an operation registry that feeds API, CLI and MCP from a single source are not features on a roadmap. They’re foundations: you either start from them, or you spend a decade retrofitting toward them while your rate card defends the old shape.

So it is worth asking what a platform built for that unclaimed position would actually have to be. Three things, and not one of them is a feature you could add later.

It cannot be a monster. What kills a small deployment is rarely the licence — it is the operational surface. A broker, a queue, a coordinator, the cluster underneath them, and the person whose job is keeping all of that alive. That person is the most expensive line item in the whole project and never appears on any pricing page. A platform a workshop or a municipal team can actually run is one that runs on ordinary hardware, survives being ignored for six months, and can be upgraded by someone who has other work to do. Everything else on this list depends on that one, because a small footprint is what makes a small price arithmetically possible instead of merely promised.

It has to be economically honest. Not cheap — honest. A price you can predict from the public page before you talk to anyone. Metering tied to something that tracks the value you get rather than the traffic you generate, because “many cheap sensors, little value each, value only in aggregate” is precisely the shape the per-message models punish hardest. No cliffs where provisioning simply stops. No rules sold back to you in an expansion pack. And no per-seat charge on the intelligence, which is the current industry answer and the one that guarantees AI stays a thing the largest customer buys.

And the AI has to live in both planes. Everything shipping today is one plane: an assistant on top, operating the vendor’s console, at a price per user. The second plane is the one nobody sells, because it cannot be bolted on — the model trained from the data the platform already holds, scoring each measurement as it arrives, its verdict landing as an ordinary field that rules read like any other. Those two planes are what “AI-native” would have to mean for the word to carry weight: a control plane an agent can drive end to end, and intelligence inside the data path rather than beside it. One without the other is half a claim.

None of that is speculative, exactly. I have been writing about the parts separately for most of this year, because that is the order in which I ran into them: Keystone for the edge runtime, OTA updates over NB-IoT for keeping a fleet alive over a link that drops mid-transfer, and a CLI whose primary reader is an LLM for what an agent-operable surface really costs to build once you stop treating it as a demo. Three posts about three problems that look unrelated until you put them side by side, and then it turns out there is one shape behind all of them.

So: is a new IoT platform in 2026 a sane idea? Fifteen years late, and I’d argue precisely on time.

Hero photo by Peter Herrmann on Unsplash.