rickmur

5 min read

Simple edge, not a closed network

Moving routing and security from the branch box into the cloud makes sites easier to deploy. Bundling every layer with 1 provider again is a different story.

Deploying SD-WAN or SASE at a branch still needs an engineer on site who knows networking. Someone has to install the box, connect the right cables, check the ISP config and troubleshoot. For a retail chain opening 40 stores in 6 countries, that means a lot of visits.

This week, Orange presented its Alternative WAN concept: a compact, standard CPE built on the open source prpl stack, with routing, security and policy running centrally. First customer PoCs are planned for mid 2027. An operator backing open source CPE software instead of another proprietary box is good for the industry. It also promises very simple installs on site.

Thin edge, thick cloud

The thin branch is not new. Cato’s socket does just enough to get the packet to the PoP: application classification, link selection and QoS. Decryption, firewalling and IPS happen centrally. Zscaler’s Zero Trust Branch splits it the same way. Carriers are heading there too. BT’s Global Fabric is an API first NaaS platform for IP VPN, Ethernet and internet on demand, and Lumen sells connectivity online with a SASE service that runs at the customer edge or in the cloud.

The split makes sense. Forwarding, path selection, local breakout and failover at the site. With policy and apps deployed centrally.

The industry tried a version of this before. Between 2014 and 2019, operators pushed vCPE and NFV, with the router and firewall running in their data centers. Traffic tromboned, latency grew and the stack moved back on premises as uCPE. So my first questions for any thin edge are simple. Does it break out trusted SaaS locally? Does it keep forwarding with its last known policy when the PoP is unreachable? And how far is the nearest PoP from a store in rural Vietnam or Brazil?

Self driving pulls the same way

Vendors take another route to the same goal. (Disclosure: I used to work at Juniper, now part of HPE.) HPE extended its self driving stack. Mist and its Marvis AI engine now also run HPE’s CX switches, with zero touch provisioning and agents that fix ports by themselves. Its EdgeConnect based SASE puts SD-WAN, firewalling and cloud delivered security in 1 console. In October, Marvis took over SD-WAN and SASE operations too, which I wrote about earlier this month.

Both approaches remove the engineer from the site. The thin edge makes the box so simple there is nothing to configure. The self driving edge lets agents configure and repair it. HPE keeps a capable box that routes and enforces policy locally and moves the operations brain to the cloud. The thin edge moves enforcement itself into the provider’s PoPs.

Where they meet? At lock in. A closed loop works best when 1 party sees all telemetry and controls all devices. HPE builds that loop across campus, data center, SD-WAN, SASE and compute. A carrier builds it across underlay, overlay and security. The more autonomous the network becomes, the stronger the pull to buy every layer from whoever runs the loop.

It looks a lot like MPLS L3VPN

It does sound familiar, because we built it before, 20 years ago with MPLS L3VPNs. Your CE talked BGP to the PE, your network lived in a VRF. The carrier added managed CE routers, central firewalls and internet breakout. It worked well. It was also 1 package: access, routing and security from 1 provider. Leaving meant migrating every site in every country.

SD-WAN broke that, you pick the best ISP per site and swap it without touching the overlay. Underlay and overlay became separate decisions, and a large part of the price and performance gains of the last 10 years came from that separation. GNX is built on it.

A carrier cloud that supplies the CPE, the access, the routing and the security brings those layers back under 1 contract. No labels or VRFs, but the shape is the same: the PE becomes a PoP, the VRF becomes a tenant and the managed CE becomes a thin CPE.

Why AI raises the stakes

The rest of the industry is pushing intelligence outward. Gartner expects more than 2 in 3 enterprises to deploy edge AI by 2029, and IDC expects half of enterprise inference to run on endpoints or edge nodes by 2030 (Network World). Small language models now run on a single GPU in a store or a metro edge node.

That changes things. First, traffic shifts from users calling a few SaaS platforms to agents making many small, latency sensitive calls to models and tools. Second, agents that operate the network need telemetry and flow data. If that data lives only in 1 provider’s cloud, your agents can only do what that provider’s API allows.

What I would love to see

Yes to the simple edge. For any thin or self driving edge, I would check:

  • Open CPE platform, so hardware and software can be replaced independently.
  • A standard data plane: IPsec or WireGuard and standard routing protocols, with nothing proprietary you can’t terminate elsewhere.
  • Local survival: local breakout and headless forwarding when the cloud is unreachable.
  • Free choice of underlay and security: any ISP, any access type, and the option to steer traffic to a third party SSE.
  • Portable policy and telemetry: policy exportable as code, and telemetry over open interfaces.

Simple sites, open layers

The thin branch will win in retail and SME, where there is no engineer on site. But simplicity should come from automation and good defaults, not from collapsing every layer into 1 provider. Whether the intelligence ends up in a carrier PoP or in a vendor’s AI engine, the layers underneath should stay replaceable. We spent 10 years unbundling the WAN. Bundling it again, just as AI makes flexibility count most, would be a step back.

Sources