Total cost of ownership (TCO) is the cost of running a technology choice as a whole, not just its initial price or a single usage fee. For a digital twin, count ingestion, storage, compute, and data transfer. For AI, add training, ongoing inference, and API fees. Compare that cost with the value the choice produces. section 4.6 section 8.5.3
Estimate costs from the workload before selecting an architecture:
- Check data volume, growth, retention, access patterns, and query needs. Storage technologies have different costs under different patterns. section 4.6
- Include integration complexity. A simpler architecture can reduce operational overhead; a polyglot architecture may improve optimization but increases the management burden. section 4.6
- For AI, estimate training compute and the cost of each inference at its planned frequency. Rule systems can run on existing hardware at negligible compute cost, while deep learning may need GPUs or costly edge hardware; managed generative AI adds per-token fees. section 8.5.3
An electricity-saving decision
Suppose a home digital twin considers an LLM API decision every 15 minutes. At about $0.05 per decision, it costs $4.80 per day. If turning off the air conditioner can save at most $2.00 per day, that approach loses money even if its decisions are more sophisticated. A rule using solar production and temperature can instead run on existing edge hardware or a low-cost cloud function. section 8.5.3
Table 4.2 Typical sources of different data types in a digital twin and how they are typically structured, integrated, and stored
|
Data type
|
Typical sources
|
Integration method
|
Structure
|
Storage technology
|
Considerations
|
|---|---|---|---|---|---|
|
Reference data
|
ERP, asset management
|
Batch, API
|
Relational
|
PostgreSQL, Memgraph
|
Update frequency, consistency
|
|
Time-series data
|
IoT sensors, SCADA
|
Streaming
|
Columnar
|
InfluxDB, Timestream
|
Volume retention, query patterns
|
|
Spatial data
|
GIS, surveys
|
Batch, API
|
Specialized
|
PostGIS, Memgraph
|
Complexity, query types, accuracy requirements
|
|
Derived data
|
Analytics, ML
|
Batch processing
|
Various
|
Feature stores, S3
|
Refresh frequency, serving requirements
|
|
Unstructured data
|
Cameras, documents
|
API, batch
|
Binary/text
|
S3, blob storage
|
Size, access patterns, processing needs
|
Use the data-type table to identify likely integration and storage choices before estimating ingestion, storage, and processing costs; the choices differ for sensor streams, reference records, spatial data, and unstructured files. section 4.6
The cheapest component is not necessarily the cheapest system. Balance cost against latency, scale, consistency, and integration overhead, then check whether the expected benefit still exceeds the cost at the intended operating frequency. section 4.6 section 8.5.3
