Facebook builds new Express Backbone network, to split the cross-data center server-to-server traffic from regular Internet-facing user traffic
Facebook has multiple data centers in the U.S and Europe. For the past ten years, they were connected by a single wide-area (WAN) backbone network known as the classic backbone or CBB.
Context & Ripple Effects
For a decade, all traffic between Facebook's U.S. and European data centers rode one wide-area network, the classic backbone or CBB, with machine-to-machine replication competing against Internet-facing user traffic for the same links. The Express Backbone ends that by giving cross-data-center server-to-server flows their own dedicated network.
The move fits a run of infrastructure work at the company: the Big Basin GPU server it debuted weeks earlier trains larger ML models, which multiplies internal synchronization traffic, and within a year Facebook would donate its Fabric Aggregation Layer data center architecture to the Open Compute Project — the same open-spec playbook this backbone follows.
First-order effects
- Facebook's own replication and ML training jobs between its U.S. and European data centers no longer contend with user-facing traffic, letting each class scale independently on hardware tuned to its needs.
- The CBB's role narrows to Internet-facing traffic only, so capacity planning for that backbone can be driven purely by user growth rather than internal workload spikes.
Second-order effects
- Network equipment vendors face a shrinking market for general-purpose WAN gear as hyperscale buyers like Facebook define their own backbone specifications and publish them through Open Compute, shifting purchasing power toward commodity, spec-driven hardware.
- Donating architectures like the Fabric Aggregation Layer pressures rival cloud operators to match the openness or risk their proprietary designs being locked out of the emerging vendor ecosystem built around OCP specs.
Third-order effects
- If the pattern holds, large-scale network operators increasingly split their networks by traffic class — user-facing versus machine-to-machine — and treat internal backbones as self-designed infrastructure rather than purchased products, a structural shift in who defines networking equipment.
- Open-sourcing timekeeping tools like Time Card alongside backbone and fabric designs points toward an industry where the coordination layers of distributed infrastructure — timing, routing, topology — are standardized publicly rather than owned by equipment makers.
The trend: Hyperscale operators are carving machine-to-machine traffic onto self-designed, openly specified networks, converting backbone infrastructure from a vendor product into a buyer-defined buildout.