Zscaler Tunnel Types: Z-Tunnel, GRE, and IPsec, With the Numbers

Three tunnels carry traffic to the Zero Trust Exchange, and they are not interchangeable: the Client Connector's Z-Tunnel for devices, GRE and IPsec for locations. The published capacity figures - 1 Gbps, the 250 Mbps NAT penalty, 400 Mbps per source IP - the scale-out shapes the documentation itself uses, the HA pattern, and the MTU habit that keeps all three healthy.

Forwarding chose the coverage domain; this article prices it. Three tunnel families carry traffic to the , split cleanly by who builds them: the Client Connector builds Z-Tunnels for a device, and the location's edge builds or for a site. Every figure below is Zscaler's own published number, live-verified 2026-07-21, and the differences between the three are exactly the fundamentals this series laid down first.

The device tunnel: Z-Tunnel 1.0 and 2.0

is the transport Zscaler Client Connector (and Cloud Connector) establishes to a . Version 1.0 is an HTTP CONNECT tunnel - a proxy connection, in essence, shaped for web traffic. Version 2.0 upgrades the transport to TLS or its datagram sibling (the same split this series covered for the protocols themselves), and with it the coverage: 2.0 is built to carry all ports and protocols from the device, not only the web pair, which is why the vendor's guidance steers new deployments toward it. The practical reading: Z-Tunnel is a per-device, authenticated transport whose capacity question is the device's, not the site's - sizing math lives with the location tunnels below.

The location tunnels: GRE first, for a reason

A site forwards through GRE when three preconditions hold: a static public egress address, an edge device that speaks the protocol, and no requirement to encrypt in transit - the plain envelope's whole trade. In exchange comes the highest published ceiling: 1 Gbps per tunnel. One documented rule bites hard enough to memorize: if the internal tunnel endpoint addresses are source-NATed, the supported figure drops to 250 Mbps per tunnel, because the service load-balances GRE traffic on those inner addresses and NAT collapses them into one. Removing that is usually the cheapest capacity upgrade in the entire design.

Scaling past a gigabit is additive and explicit: more tunnels, each from a different public source IP. The documentation's own worked examples set the shape - 2 Gbps is two primary tunnels plus two backups; 3 Gbps is three plus three - and the backup halves point at Public Service Edges in different data centers, which is the platform's standing high-availability pattern.

The location tunnels: IPsec, when GRE cannot

Where the egress address is dynamic, the device lacks GRE, or the transit must be encrypted, the site forwards through IPsec - IKEv2 negotiating, ESP carrying, peers identified by when addresses will not sit still, Dead Peer Detection watching liveness. The published capacity is 400 Mbps per public source IP address, and the two documented ways past it are pure protocol: additional tunnels from different public source addresses, or multiple tunnels from the same address using NAT-Traversal with source-port randomization under - the port-varies property doing exactly what the fundamentals article promised it could. The documentation's sizing examples again: 800 Mbps as two primaries plus two backups, 1200 as three plus three, backups in different data centers.

The honest comparison, then, is not "which is better" but "which constraints do you actually have": GRE buys ceiling and simplicity at the price of preconditions; IPsec buys tolerance and secrecy at the price of megabits.

The habit all three share: MTU

Every one of these is an envelope, and envelopes charge bytes. Zscaler's field guidance is the same habit that article teaches: set the tunnel to the lower of the appliance MTU and the measured path MTU, fall back to 1400 when trouble appears anyway, and clamp TCP to match - because the signature of a wrong tunnel MTU (handshakes fine, real data stalls) looks like a Zscaler problem and never is.

Choosing, deterministically

The whole decision - eliminations, per-tunnel figure, minimum primary and backup counts - is mechanical once the constraints are named, which is why it is a tool on this site: six answers in, the sized recommendation out, every step and source shown. The exam-shaped summary it encodes: encryption, dynamic IP, or a GRE-less device means IPsec at 400 per source IP; otherwise GRE at 1000 (250 under internal-endpoint NAT); primaries by ceiling division; backups mirror primaries into a different data center.