Three different IT directors mentioned the same problem to me in the last fortnight: egress costs are running well over forecast and the FY27 budget cycle is going to be awkward. This isn’t a surprise to anyone watching the cloud cost trajectory, but the magnitude is bigger than what most teams modelled when they signed their current cloud agreements.

A look at what’s actually driving the pain and what’s working to control it.

Why egress is hurting more in 2026 than in 2022

A few structural shifts have stacked on top of each other.

Cross-region traffic patterns have grown faster than baseline workload growth. The shift to multi-region active-active architectures, the rise of cross-region data replication for compliance, and the AI workload patterns that pull data from one region for inference in another have all pushed cross-region egress sharply higher.

Inter-cloud traffic is now real. Five years ago, most Australian enterprises were single-cloud. Multi-cloud was an aspiration, not a reality. In 2026 it’s reality for most mid-market and large organisations — Microsoft for productivity, AWS for digital services, Snowflake or Databricks for analytics, OpenAI or Anthropic for AI. Every connection between these has egress implications, and the architectural choices made three years ago were often made without thinking carefully about the resulting traffic patterns.

The AI workload story makes this worse. Inference workloads typically pull substantial amounts of context data — embedding lookups, document retrieval, knowledge base queries — and the architectural separation between where the data lives and where the inference happens has predictable cost consequences.

Where the cost concentration tends to sit

When I’ve looked under the bonnet of mid-market Australian IT cost stacks in 2026, the egress concentration usually clusters in a few places.

Data analytics platforms calling source systems. The classic pattern is a cloud-hosted analytics platform pulling production data from databases hosted in a different region or a different cloud. Six years ago this was a few gigabytes a day. In 2026 it’s often terabytes, and the egress line on the cloud bill reflects that.

AI inference pulling from RAG knowledge bases. Retrieval-augmented generation workloads at scale are egress-intensive. Every query involves embedding lookups, document fetches, and round trips back to the inference endpoint. The architectural decisions about where to host the knowledge base, the embedding service, and the inference all matter a lot.

Backup and DR replication. This used to be a quiet line item. With more data, longer retention requirements, and stricter compliance expectations, it’s not quiet any more.

Internet egress for user-facing workloads. Browser-based applications serving large media assets — video, high-res images, downloadable documents — generate user-egress costs that scale linearly with traffic. CDN strategy and caching design make an enormous difference here.

What’s actually working

A few patterns I see giving real cost relief in 2026.

Architectural co-location of related workloads. The single biggest impact comes from putting workloads that talk to each other a lot in the same region (and ideally the same availability zone where it makes sense). The savings from this kind of architectural rework can be 30-50% of the egress bill once it’s done.

Cross-cloud private connectivity (Direct Connect, ExpressRoute, dedicated interconnect). For organisations with predictable high-volume inter-cloud traffic, the dedicated connection options are now meaningfully cheaper than internet egress. The break-even point is lower than it used to be.

Egress-aware data architecture for AI workloads. Putting the vector store in the same region as the inference endpoint. Caching embeddings where it makes sense. Designing the retrieval flow to minimise round trips. The teams doing this carefully are getting AI cost-per-query 40-60% lower than the teams that just deployed the reference architecture and accepted the egress consequences.

Honest CDN configuration review. Most enterprises have a CDN. Most enterprises haven’t reviewed the configuration in 18-24 months. The defaults that made sense in 2022 don’t always make sense in 2026; cache hit rates and origin egress are often substantially improvable with a couple of weeks of focused work.

What’s not working

A few approaches that look like they should help but mostly don’t.

Hoping the cloud provider will discount. Egress pricing is one of the stickier areas of cloud pricing. Even at enterprise commitment levels, the egress component doesn’t move much in negotiations. The price card is the price card, and the conversations to have are about architecture, not discounts.

Repatriating workloads to on-premises infrastructure. The hyperscaler economics still beat private cloud for most enterprise workloads even when egress is included. The break-even point for repatriation has shifted, but it’s still in the territory of large, predictable, IO-heavy workloads — not the general case.

Blaming the developers. The egress patterns in a typical enterprise cloud architecture reflect choices made by infrastructure architects, not application engineers. Trying to fix egress by asking engineers to “be more careful” tends to produce frustration and not much else.

What I’d budget for

For FY27 budgets in mid-market Australian IT, I’d expect egress to grow 25-40% year-on-year even after sensible optimisation efforts. The growth in AI workloads alone justifies the upper end of that range.

The IT directors who’ve forecast this honestly and have a 12-month plan for architectural optimisation alongside it are getting more credibility with their CFOs than the ones who promised flat spend and are now coming back asking for an unforeseen budget top-up.

It’s not glamorous work but it’s the work that matters. Cloud cost discipline in 2026 is mostly about being honest about the cost trajectory and disciplined about the architectural choices that drive it.