What is a data center?
Understand the physical facility.
← Back to Compute College
Compute College
What is Data Center Interconnection?
Data center interconnection links AI capacity to networks, clouds, data sources, and buyers who need to use it.
Plain-English definition
Data center interconnection is the physical and network connectivity that links a facility to carriers, cloud networks, other data centers, data sources, and customers. For AI compute, interconnection determines how effectively inputs reach a cluster and model outputs reach buyers and users.
Memory trick: Power turns GPUs on; data center interconnection makes them reachable and commercially useful to buyers.
Why it matters
Accelerators and power do not make a useful market product in isolation. Training data may need to enter a site, checkpoints may need to leave it, cloud services may coordinate with it, and serving traffic may require predictable latency. Weak connectivity can reduce the value of otherwise available compute.
- Data movement time and egress expense can alter total buyer cost beyond the GPU-hour rate.
- Well-connected locations can be more usable for hybrid cloud, distributed, or latency-sensitive workloads.
- Network expansion may improve the market value of regional capacity already built or planned.
Simple example
Suppose two providers offer the same 48-hour accelerator job and the remote option appears cheaper per GPU-hour. If that site adds hours of staging delay, expensive data transfer, or unacceptable output latency, the buyer may finish later or spend more overall despite paying less for the GPU time.
- Compare project completion time and transfer fees together with hourly capacity pricing.
- A training workload and a real-time serving workload may value the same connection very differently.
- Treat network-cost or latency assumptions explicitly rather than implying every location is equivalent.
Example figures are illustrative calculations, not current quoted market prices.
Common mistake
Do not assume the lowest-cost location is the cheapest completed workload. A buyer may lose time moving data, pay additional transfer costs, or fail latency requirements. A facility also can be physically near demand yet lack the needed carrier or cloud connection.
Practical takeaway
What you can do with this
Procurement teams should gather connectivity terms with any compute quote. Analysts evaluating a region should consider network reach as one contributor to usable supply, while operators should identify the connections required before promising capacity to a customer.
- Buyers: ask about carriers, cloud connections, egress charges, throughput, latency, and delivery timelines.
- Product managers: separate batch-training connectivity needs from interactive serving requirements.
- Analysts: follow network investments together with facility and power announcements.
- Operators: identify which buyer workloads the site can support economically, not just physically.
Decision check: normalize provider comparisons using GPU cost, transfer cost, data-staging time, latency needs, and availability before selecting a site.
Compute College track
Power & Data Centers
Step 11 of 17: What is data center interconnection