Something interesting is happening inside the companies that make physical things — the industrial firms, the fleet operators, the smart building managers, the manufacturers running IoT platforms at scale. They’re all quietly becoming AI companies. Not because they want to be. Because the economics are forcing their hand.

A recent analysis by Andreessen Horowitz, based on spending data from over 200,000 companies, revealed that AI application spend is surging — and it’s no longer confined to software startups in San Francisco. The top 50 AI applications by enterprise spend now include everything from creative tools and coding platforms to compliance automation and accounting software. The pattern is clear: AI is moving from experimental side project to core operational cost.

For anyone working in IoT and industrial technology, this shift has a very specific implication: the cloud bill is changing shape. And if you’re not paying attention, it’s going to catch you off guard.

IoT platforms have always been cloud-dependent. Sensor data flows into Azure IoT Hub, AWS IoT Core, or Google Cloud IoT — where it gets stored, processed, and analysed. That’s been the architecture for a decade. But the workloads running on top of that data are changing fast.

Predictive maintenance models that used to run on simple threshold logic are being replaced by machine learning pipelines that need GPU compute for training. Anomaly detection across thousands of sensors is moving from rule-based alerting to AI-driven pattern recognition. Natural language interfaces are being layered on top of industrial dashboards so that operations managers can query their data conversationally instead of writing SQL. Every one of these upgrades adds AI inference costs on top of existing IoT infrastructure costs.

The a16z data backs this up. Nearly 70% of the top AI applications they tracked can be adopted by individuals and brought into teams without enterprise procurement — meaning AI tools are spreading through organisations bottom-up, often without centralised budget oversight. In IoT-heavy companies, this means engineers, data scientists, and even field technicians are spinning up AI workloads that nobody in finance is tracking until the invoice arrives.

Here’s the part that doesn’t get enough attention. Many IoT and industrial technology companies signed cloud commitments — reserved instances, enterprise agreements, or startup credits — based on projections that assumed their workloads would grow in a certain way. Then AI happened, and everything shifted.

Some teams moved workloads to a different provider that offered better AI services. Others consolidated platforms and ended up with surplus capacity on their original cloud contract. Startups that received large Azure or AWS grants through accelerator programmes pivoted their product and never consumed the credits. The result is a growing pool of unused cloud capacity sitting on balance sheets, quietly expiring.

If your organisation has unused Azure capacity — whether from a startup grant, an enterprise agreement you’ve outgrown, or credits from a programme you no longer participate in — you can sell Azure credits through brokers who match sellers with buyers looking for discounted cloud capacity. It’s a straightforward way to recover cash from credits that would otherwise expire worthless. Given how fast cloud strategies are shifting right now, this is worth checking at least quarterly.

The companies handling this transition well share a few common practices. First, they’re separating IoT infrastructure costs from AI inference costs in their cloud budgets — treating them as distinct line items with different growth trajectories and different optimisation levers. IoT data ingestion and storage is predictable. AI inference is not.

Second, they’re pushing inference to the edge wherever possible. Running an anomaly detection model on a gateway device at the factory floor costs a fraction of running it in the cloud — and it’s faster. Edge AI is not new, but the economic pressure from rising cloud inference costs is making it a financial necessity rather than a technical preference.

Third, they’re auditing their cloud commitments regularly. The days of signing a three-year cloud deal and forgetting about it are over. When your workload mix is changing every quarter — more AI, less traditional compute, different providers for different tasks — your cloud procurement strategy needs to be equally dynamic.