Q1 2026 cloud infrastructure spending hit $129 billion - up 35% year over year. The hyperscalers are collectively pouring over $600 billion into capex this year, most of it going to AI infrastructure. And that bill is getting passed downstream.
If you run enterprise infrastructure, you've probably already noticed. Instance prices creeping up. Reserved instance discounts getting thinner. Energy surcharges appearing on invoices that didn't have them last year. This isn't a blip. The structural cost of running cloud compute is increasing because the data centers hosting your workloads are also hosting GPU clusters that consume enormous amounts of power.
The squeeze for non-AI workloads
Here's the thing that doesn't get talked about enough: most enterprise compute isn't AI. It's virtual desktops, databases, web apps, batch processing. Regular workloads that haven't gotten cheaper to run even though they haven't gotten more demanding. The cloud providers are investing in AI infrastructure because that's where the growth is, but they're not lowering prices for everything else.
For DaaS specifically, this hits hard. Desktop workloads are compute-intensive but low-margin. You can't just absorb a 10-15% cost increase across hundreds of thousands of sessions. The math breaks. And unlike a SaaS product where you can pass cost increases to customers through price hikes, enterprise DaaS contracts are typically multi-year with fixed pricing.
What we're doing about it
Three things are working for us at Citrix right now:
Aggressive right-sizing. This sounds obvious, but most orgs are still massively over-provisioned. We run weekly right-sizing analysis across our entire fleet and have automated recommendations that flag instances running consistently below 40% utilization. The hard part isn't finding the waste - it's getting teams to act on it. I've started including right-sizing compliance in PI planning as a standing commitment, which gives it visibility and accountability.
Spot and preemptible for everything that isn't session-critical. Build systems, staging environments, batch analytics, image baking pipelines - none of these need guaranteed uptime. Moving them to spot instances cut that portion of our bill by roughly half. The engineering cost was designing for interruption (checkpoint, resume, retry), which took a PI to get right but pays for itself every month.
Reserved instance arbitrage across clouds. We run multi-cloud anyway for resilience, but we've gotten more intentional about shopping the pricing. AWS and Azure don't move their RI pricing in sync. When one offers better rates for a given instance family in a given region, we shift baseline capacity there. It requires flexible deployment tooling (which we have) and a TPM who watches the pricing dashboards (me, every Monday morning).
The repatriation question
Cloud repatriation is having a moment. Basecamp did it publicly, DHH wrote the blog posts, and now every CFO asks whether we should "bring it back on-prem."
For most enterprise workloads, the answer is still no. The operational cost of running your own hardware only makes sense at very large, very stable, very predictable scales. If your workload is bursty - and DaaS is extremely bursty - you need cloud elasticity. The morning login storm means you need 3x your average capacity for 90 minutes, then you scale back down. Owning hardware sized for peak means it sits idle 85% of the day.
But I do think hybrid is getting more attractive. Baseline capacity on owned or colocated hardware, burst to cloud. The economics are shifting. If your baseline load is big enough and stable enough, the break-even point for owned compute is getting lower as cloud prices rise. We're running the numbers on this for some of our most predictable workloads.
What this means for planning
If you're a TPM running infrastructure programs, cost is now a first-class planning input alongside performance and reliability. It's not something finance handles separately. Capacity plans need cost projections. Architecture decisions need cost modeling. Every PI planning should include a line item for cost efficiency work.
The orgs that treat cloud cost as somebody else's problem are going to have a rough year. The ones that bake it into engineering culture - where teams own their cost metrics the same way they own their latency metrics - will come out ahead.