The pitch is that AI has changed what a network is for. Sakhnov put it this way: the network is "no longer about connectivity" and is becoming the intelligent control layer for the agentic world.
Why AI traffic changes the design
As agentic AI grows, so do the demands on networking. Kerner explains that AI workloads shift traffic inside the data center from north-south to east-west, because GPUs exchange data with other GPUs, servers and storage. AI agents add another layer. They span compute, GPUs, the network and data centers, and demand can move from one place to another overnight.
A flatter, two-tier data-center fabric
Microsoft is replacing its three-tier data-center network with a flatter two-tier design. Per Sakhnov:
- In the old design, servers sit under fixed branches with shared uplinks.
- A path that took five hops can now take three.
- Servers can push full throughput without the network limiting them.
- The new fabric is more than 20% more power efficient than the old one.
The power angle is the one to notice. Sakhnov said power is the scarcest resource in the AI era, so a network that uses less of it frees capacity for compute and GPUs. He described that as "a huge deal."
Read the numbers carefully. These are Microsoft's own figures about its own infrastructure. The reporting gives no benchmark methodology, no definition of "full throughput," and no baseline for the 20% claim. It also doesn't say how widely the new fabric is deployed. Fewer hops is not a promise of lower end-to-end latency for every Azure workload. Azure customers can't switch this on. It is an infrastructure change inside Microsoft's data centers.
Hollow-core fiber outside the data center
Sakhnov said that outside the data center, "physics becomes the next constraint." Conventional fiber sends light through solid glass. Hollow-core fiber sends it through air.
Microsoft says this:
- cuts latency by about one-third;
- needs fewer amplifiers;
- lets data centers sit farther apart without a latency penalty.
That last point matters because power allocation heavily drives where sites get built. Microsoft runs about 15,000 kilometers of hollow-core fiber today. It plans to have 10 times more by the end of 2028. That is a plan, not a delivered result. The reporting doesn't say which routes the fiber covers or whether every Azure link will use it.
GPU capacity beyond Azure regions
Microsoft says GPU capacity is also moving outside Azure regions, to partner data centers, extended zones and enterprise locations. It connects that capacity to Azure through AI points of presence. Private traffic runs from each PoP over ExpressRoute, and security devices attached to the PoP add a security overlay. In Sakhnov's words, outside the region doesn't mean outside the network or outside of control.
AI agents running the network
Microsoft operates about 1.6 million network-attached devices. More than 100 AI agents already work on that network. Microsoft calls them digital coworkers, each with its own identity and a manager. Over the past year they handled more than 30,000 cases and cut the backlog by about three quarters.
Sakhnov gave one example, an agent named Miles Dyson that specializes in fiber cuts. (The name is also that of the Skynet founder in the Terminator films.) On June 7, an excavation cut a fiber east of Jakarta. Per Sakhnov:
- About 20 minutes later, the agent had used device telemetry to locate the fault to within a meter.
- It arranged two-hour updates with the local repair crew.
- When the crew didn't return in the expected time, it brought Microsoft engineers onto the bridge.
- After the repair, it verified the link from both ends and closed the ticket.
This is a vendor anecdote, not independent validation. Still, it shows what Microsoft means by agents handling repetitive operations work as the network grows.
Customer-facing networking announcements
Several services at Ignite are aimed at Azure customers:
| Service | What it does | Status reported |
|---|---|---|
| Azure Multicloud Interconnect | Provisions cross-connects between Azure and other clouds through APIs | AWS generally available; GCP in preview; AT&T coming soon |
| Azure Last Mile Interconnect with Lumen | Last-mile connectivity | Preview |
| Cross-cluster networking for AKS | Peers clusters through Azure Kubernetes Fleet Manager, with failover and one network policy | Not specified |
| Azure Route Server route maps | Filter and summarize routes, apply AS path prepending | Not specified |
| Agent Grid | Identity-based microsegmentation, isolation and traffic policies for AI agents | Public preview, per Microsoft's Ignite materials |
| Azure Firewall AI add-on | Runs two inline small language models on Firewall Premium to detect malware and exploit requests | Not specified |
| Azure Adaptive Virtual Patching | Maps published vulnerabilities to VM workloads and applies inline mitigations until the patch arrives | Not specified |
Narayan Annamalai, vice president of product management for Azure Networking, framed the customer problem as operating networks rather than setting them up. Many customers already run multiple VNets that peer with each other and connect on-premises through ExpressRoute and VPN gateways.
Annamalai said a cross-connect between clouds that once took days or weeks now takes minutes. Setup runs through APIs, and each connection is quad resilient, with two sites and two connections at each site. Last August, The Register reported that Microsoft's Azure-AWS link was then in preview.
Other points from the session:
- IPv6 peering: Virtual network peering with IPv6 is in public preview. Teams can reuse the same IPv4 ranges in every VNet and give each dual-stack VNet unique IPv6 addresses. Traffic between VNets runs over IPv6. Annamalai said Microsoft runs its own network this way.
- Route maps: Teams can filter routes, summarize on-premises routes into fewer prefixes, drop routes, and use AS path prepending to select a source.
- AKS: Cross-cluster networking removes the manual stitching between clusters. Traffic shifts to the other cluster if backends fail.
Agent Grid
Agent Grid is a security control. Microsoft's Azure Networking blog describes it as identity-based, Zero Trust secure network access for agents. It brings network isolation, micro-segmentation, inspection, and granular Layer 3–7 policies into Azure's network. Kerner reports that agent code doesn't change, and that each agent connects out to other agents, models, MCP servers and the internet. Annamalai's argument is that protecting each of those connections with separate gateways has become very complex. Microsoft's Book of News lists Agent Grid among the preview capabilities, so it is not yet a production-ready control. Treat it as something to evaluate, not deploy.
What admins should take from this
- The fabric and fiber changes are background context. They are not settings you can enable.
- The most actionable items are the customer-facing ones: Multicloud Interconnect, route maps, AKS cross-cluster networking and IPv6 VNet peering. Check each one's status before planning around it, since several are in preview.
- The security items (Agent Grid, the Firewall AI add-on, Adaptive Virtual Patching) matter if you are putting AI agents or exposed VMs on Azure. Pricing, regional availability and exact release timing weren't covered in the reporting.
- Microsoft's performance and efficiency numbers are self-reported. Wait for independent measurement before using them in capacity planning.
Kerner also noted that Microsoft previewed Windows Server 2028 at Ignite, which it calls its largest platform update since Windows Server 2025. That is a separate story.
Annamalai's closing line summed up the stakes: you can have the best AI model in the world, but if your network goes down, it's an expensive way of saying 404 not found.
References
- Microsoft Ignite news: What’s next for Azure networking? - Network World Network World · 2026-11-20T21:47:39+00:00
- Introducing Agent Grid techcommunity.microsoft.com
- How Azure Networking is ready for the next generation of AI solutions | BRK433 - Tech Hub tech.hub.ms