Skip to content
Free public beta until December 4, 2026: every course is open, no card needed. After that, the first chapter of each course stays free and full access is a one-time $39 ($19 for up to 100 founding members chosen for helpful feedback). Your progress is kept. How the beta works

Every way to send less data

Fewer messages, smaller messages, or a smaller question: batching, reporting by exception, compact formats, summaries, edge filters and on-device models, and what each one costs.

Ask first which number you're trying to cut, because the levers are different.

  • The message count (it sets most per-message bills and much of the radio's energy) → send several readings per message, or report only when a value changes enough, with a heartbeat so silence still means something.
  • The bytes per message (they matter on LoRaWAN, NB-IoT and metered links) → CBOR or Protobuf instead of JSON, shorter topics.
  • What gets stored → a retention policy with downsampled summaries.
  • What leaves the device at all → filter, summarise or classify at the edge, and send the result instead of the raw stream. Reach for rules and features before a model.

Every lever has a cost: latency for batching, a breaking change for a new message shape, lost detail for summaries. Write down which one you paid.

Fewer messages

Section

Cost and scale

Halving the reporting interval costs as much as doubling the fleet.

Term

batching

See section →
Saving up several readings and writing them in one go, which is much cheaper than one write each.

Reporting by exception, and why it needs a heartbeat.

Term

heartbeat

See section →
A small message a device sends on a schedule just to say it is still alive.

Smaller messages

66 bytes of JSON, 55 of CBOR, 21 of Protobuf, for the same reading.

Term

CBOR

See section →
A binary format that holds the same kinds of data as JSON in fewer bytes.
See section →
A compact binary format whose messages are described by a schema both sides share.

Example

The greenhouse reading's IP packet falls from 159 bytes (JSON) to 114 (Protobuf): 28 % smaller, though the payload is 68 % smaller.

Less stored

See section →
The rule for how long each kind of data is kept before it is deleted.
See section →
Replacing many detailed readings with fewer summary ones, such as hourly averages.

Less leaving the device

See section →
Computing on the device so that much less data has to leave it, such as a few features instead of every sample.

A window of samples becomes a few numbers.

A threshold or a rule first; a model only when they fail.

Section

What TinyML is

Send the class, not the waveform.