For years, I have said blockchain is not just money.
Then crypto took over the conversation.
Blockchain became tokens, exchanges, wallets, speculation, and charts people watched like heart monitors. Even serious enterprise projects were forced to explain themselves through the language of financial transactions.
That was always too narrow.
The more consequential use was sitting in the physical world.
A sensor detects heat. A machine changes pressure. A door opens. A container moves. A vehicle leaves its route. A medical device produces a reading under a particular firmware version.
These are small facts. Most are uneventful. Together, they describe what actually happened.
IoT gave machines the ability to observe those facts. What it did not give them was a reliable, shared memory.
Now AI has arrived, and that missing memory matters.
Most business systems are designed to maintain state.
A database tells us that a pump is offline, a truck is at a location, a shipment is delayed, or a piece of equipment is unavailable. That current state is useful, but it does not necessarily preserve the path that created it.
The pump did not simply become offline.
It began vibrating differently. Pressure changed. A warning was generated. Someone dismissed it. A technician inspected the machine. It returned to service. Two more abnormal readings appeared. Then it stopped.
The current state tells us where the machine ended up.
The event history tells us what happened.
That history is usually scattered across sensor platforms, maintenance systems, text messages, vendor databases, dashboards, invoices, emails, and handwritten notes. Some records are overwritten. Some are summarized. Some are stored without enough context to interpret them later. Some can be changed without leaving a meaningful trace.
Then we give the fragments to AI and ask it to explain reality.
The problem is not always that the AI lacks intelligence.
It lacks evidence.
In Humanity’s Ledger: The Trust Protocol, I described blockchain as a digital counterpart to human memory. Its value was not simply that it could process transactions. It could create a persistent record of interactions and preserve that record across time.
I was primarily thinking about human trust, identity, contribution, and the social systems we were building around AI and automation.
The same idea becomes more concrete when the writers are machines.
A device does not need to produce a complex blockchain transaction every time it observes something. It can create a small, signed JSON object:
{ "device_id": "pump-18",
"event": "pressure_change",
"value": 91.4,
"unit": "psi",
"observed_at": "2026-09-03T03:18:22Z",
"firmware": "3.2.1",
"location": "site-04",
"previous_event": "0x8c41...",
"signature": "0xa72e..."
}One object means almost nothing.
That is what makes the idea practical.
It is simply a machine making a small statement:
This is what I observed.
This is when I observed it.
This is the software I was running.
This is the event that came before it.
This is my signature.
Millions of those statements create a chronological, attributable, machine-readable account of the physical world.
The important unit is not the transaction.
It is the event.
We have been able to collect machine data for years. We have also been able to build immutable ledgers for years.
What we lacked was something capable of continuously interpreting billions of small events without requiring a custom application for every possible question.
AI changes that.
An AI system can examine the history and determine what changed before a failure. It can compare one device with nearby devices. It can identify whether an abnormal reading was isolated or part of a larger pattern. It can connect equipment behavior with maintenance history, environmental conditions, operator actions, software updates, and financial consequences.
Consider a piece of construction equipment.
A sensor records that it stopped operating. A maintenance system records that it was pulled from the field. A dispatch system still shows it assigned to a project. A billing system continues treating it as active.
Those are not four unrelated records.
They are one event moving through four parts of the business.
The equipment stopped operating, which changed its availability, which changed the dispatch plan, which changed what could be billed to the project.
A person familiar with the company may understand that relationship immediately. The systems usually do not.
An AI reasoning over a continuous event history can.
That is the difference between using AI to summarize a dashboard and giving AI enough memory to understand the operation.
The idea falls apart if every temperature change, vibration reading, and GPS coordinate requires a public blockchain transaction.
That would be expensive, slow, invasive, and unnecessary.
The practical architecture is more restrained.
Devices sign events locally. Full telemetry flows into inexpensive event or object storage. Important state changes, device identities, permissions, hashes, and cryptographic commitments are written to a private ledger. Groups of events can be committed through Merkle roots or similar proofs rather than recorded individually.
A company can periodically anchor a checkpoint to a public network when independent verification is useful.
The detailed data remains private.
The integrity of the history remains provable.
“Free” does not mean that storage, computing, networking, and security cost nothing. It means the machine does not face a separate transaction toll every time it speaks. The marginal cost of preserving another small event should be close enough to zero that the system does not need to decide whether the truth is worth recording.
There does not need to be a token.
There does not need to be a public marketplace.
There may not even need to be a conventional blockchain.
A permissioned append-only log, supported by device signatures, cryptographic commitments, access controls, and external anchoring, may provide everything a single organization needs.
The architecture matters more than the label.
What matters is that the history is persistent, attributable, ordered, and difficult to alter without detection.
A normal database may be enough when one company owns every machine, every system, and every decision.
The problem becomes more interesting when several parties own different pieces of the truth.
A manufacturer builds the device.
Another company operates it.
A contractor maintains it.
A software vendor controls the firmware.
An insurer covers the risk.
A customer pays for the output.
A regulator may later ask what happened.
Each party has its own records. Each record may be accurate within its own system. None of the parties independently possesses the full history.
A private shared ledger does not require everyone to expose all of their data. It gives them a common evidentiary timeline to which each party can make signed commitments.
The ledger does not replace the manufacturer’s database, the operator’s maintenance system, or the insurer’s claims platform.
It binds their records to a history they cannot quietly rewrite later.
That distinction matters in supply chains, healthcare, insurance, energy, construction, aviation, logistics, manufacturing, and any other environment where the physical event crosses organizational boundaries.
In Humanity’s Ledger, I argued that the ledger should be part of a larger system of trust, not merely a mechanism for managing economic transactions. The challenge was to design systems that recognize the broader human and societal meaning around the record.
The same principle applies here.
A ledger is not valuable because it stores data.
It is valuable because it allows separate parties, systems, and intelligent agents to reason from a common history without requiring blind trust in whoever currently controls the database.
There is an obvious limit.
A compromised sensor can sign false information.
A ledger can prove that a record was not changed after it was written. It cannot independently prove that the observation was correct.
Record integrity and truth are not the same thing.
A serious system must account for device identity, secure hardware, calibration, firmware integrity, trusted time, physical tampering, network conditions, maintenance history, and the possibility that a device is simply wrong.
This is another place where AI becomes useful.
The system can compare a sensor’s statement with surrounding evidence. Did nearby sensors report the same condition? Was the device recently calibrated? Is the reading physically possible? Did the equipment behave in a way consistent with the event? Did another machine observe the same change from a different position?
Instead of declaring each signed event true, the AI can evaluate confidence.
The ledger preserves the claim.
The evidence surrounding the claim determines how much the claim should be trusted.
That is a more honest architecture than pretending immutability creates truth. It also gives AI something much better than a clean-looking row in a database. It gives the model the event, its origin, its context, its relationships, and the contradictory evidence around it.
Once this memory layer exists, AI agents can act on something stronger than a current-state API.
A maintenance agent can detect that a failure pattern is forming and schedule service before the machine stops.
A logistics agent can determine that a shipment’s temperature history violated its handling requirements, even if its current temperature has returned to normal.
An insurance agent can reconstruct the conditions surrounding a loss rather than relying only on statements gathered after the fact.
A medical system can connect a specific reading to the exact device, firmware version, calibration state, and operating conditions that produced it.
A construction agent can recognize that equipment removed from service should also be removed from dispatch availability and project billing.
The intelligence is not limited to identifying anomalies.
It can understand consequences.
That is where the idea becomes larger than blockchain for IoT.
IoT creates observations.
The ledger turns those observations into persistent memory.
AI interprets the memory.
Agents carry the interpretation back into the operation.
We spent the blockchain decade focused on whether machines could hold money, exchange value, and settle transactions.
Those capabilities may still matter.
But the larger breakthrough may be giving machines a trustworthy way to leave behind evidence of what they saw, what they did, and what happened next.
The physical world already produces the data.
What it has been missing is a memory.