The difference is purpose, not size.
An Inference Point and a data center share equipment: racks, chips, cooling, power. They share almost nothing else. One is built to hold. The other is built to answer. Treat them as one thing at two sizes, and you get both wrong.
Side by side
| Data center | Inference Point | |
|---|---|---|
| Exists to | House and store | Compute and deliver, locally |
| Defined by | What it contains | What it reaches |
| Counted in | Square feet and megawatts | Service radius and response time |
| Serves | Customers anywhere in the world | The people, businesses and machines within its reach |
| Sited | Wherever power and land are plentiful. Its work does not depend on where it sits. | Near the people using it. Moved away, it stops working. |
| The data | At rest: kept, backed up, retrieved | In transit: a request in, an answer out |
| Arrives as | Typically a purpose-built campus on new land | A tenant inside a building that already exists |
| Nearest relatives | Warehouse, archive | Switching office, cell site, exchange |
The test is location.
Ask one question of any facility: would it still do its job if you moved it 500 miles away? A data center would. Its customers would never notice. An Inference Point would not. The answers would arrive too late to use.
That dependence on place is what makes it local infrastructure. It belongs with the systems a community already hosts because they only work nearby.
Cell tower
Stands where its signal can reach you.
Fire station
Sits where its crews can reach you in time.
Substation
Steps power down for the streets around it.
Inference Point
Computes where its answers can reach you in time.
Two more ways to say it
A warehouse and an exchange
A warehouse is judged by what it holds. An exchange is judged by what passes through it, and how fast. A data center is a warehouse. An Inference Point is an exchange.
Data in transit
An Inference Point is not where records are kept. A request arrives, is computed, and leaves as an answer within milliseconds, the way a call passes through a switch.
A category is only as good as its limits.
Naming a category is easy. Holding it to limits is what makes it mean something. An Inference Point comes in three scales, each sized to the area it serves and each with a ceiling.
Local Inference Point
Up to 7 MW
Community Inference Point
Up to 12 MW
Regional Inference Point
Up to 20 MW
What holds at every scale
- Inside an existing building
- On existing distribution power
- No water consumed for cooling
- Serving the area around it
A facility that goes past these limits is not an Inference Point, and should be reviewed as whatever it actually is.
Questions people ask
Isn't this just a small data center?
No, and size is not the reason. A small warehouse is still a warehouse. The two share equipment and very little else: a data center is defined by what it contains, an Inference Point by what it reaches. Ask what a facility is for and where it has to be, and the two come apart.
Can't a data center run inference too?
Not all inference. A data center can run the kind that can wait: summarizing a report overnight does not care about distance. Real-time inference is different. A conversation, a production line or an emergency call needs its answer in milliseconds, and a distant data center cannot deliver that. That work needs compute within reach, which is the job an Inference Point exists to do.
Why does the category matter?
Because rules follow categories. Communities are writing standards for data centers, built around what data centers are: large campuses on new land, sited for power and far from the people they serve. An Inference Point should be reviewed on what it actually is, and held to limits that fit it.
See what one means for the place that hosts it.
What neighbors see, what it draws, and where the local growth comes from.