GenAI Won't Just Eat Software Jobs—It's Coming for Public Cloud
For ten years, the narrative felt unstoppable. Build on public cloud or watch your competitors leave you in the dust. Managed services promised infinite scale and global reach. Why bother fighting gravity? Then generative AI arrived and flipped the script in a way nobody saw coming. The same tech everyone feared would replace software engineers is now quietly tearing down the operational moat that made hyperscale cloud feel inevitable.
Think of it as a great rebalancing. Agentic SRE systems write runbooks automatically. Auto-remediation fixes problems before humans notice. Copilots generate Terraform configurations and Kubernetes manifests in minutes instead of days. Once the operational burden drops, the economics shift dramatically. Many large businesses run predictable workloads. For them, owning capacity or using cheap global VPS suddenly makes financial sense again.
(This article focuses on non-AI public-cloud workloads other than Gen AI workloads)
The Old Deal Is Dying
Public cloud never just meant "someone else handles the servers." Companies paid premiums for speed, safety, and scale without building massive platform teams. However, AI is closing that gap faster than most people realize.
Agentic operations eliminate grunt work entirely. Large language models generate infrastructure code on demand. They propose safe rollbacks during incidents. They summarize complex problems and execute standard runbooks within policy boundaries. Tasks that consumed entire afternoons for platform teams now finish in minutes. Better yet, documentation stays consistent and audit trails remain complete.
Open-source tools have become "managed enough" for serious production use. Postgres operators, vector databases, model servers, feature stores, and observability stacks once seemed too risky to self-host. Now agents assemble them using policy-as-code frameworks. Customization beats vendor lock-in every time. The AI prevents glue code from deteriorating over time.
Predictable workloads outperform elastic ones economically. Most mature enterprises see steady traffic patterns above eighty percent. High utilization on owned infrastructure or VPS crushes pay-by-the-hour pricing. Egress costs amplify this advantage significantly.
Here's the kicker: operational premiums are collapsing industry-wide. Unit economics become the only scoreboard that matters—euros per thousand inferences, euros per thousand tokens, euros per gigabyte processed. Cloud's convenience markup gets exposed under this harsh light.
Bursts Still Need Cloud—Just the Top Ten to Twenty Percent
Cloud advocates immediately mention spiky workloads. Marketing events cause traffic surges. Model evaluation farms spin up temporarily. They're absolutely right about these scenarios. Renting capacity for unpredictable bursts makes perfect sense. That doesn't justify parking your entire infrastructure in hyperscale data centers forever though.
A smarter pattern is emerging across the industry. Run your baseline on owned infrastructure or VPS. Size it for seventy to eighty-five percent utilization. Create burst lanes in the cloud or with secondary bare-metal providers. Use them for experiments, traffic spikes, and disaster recovery drills. Maintain the same control plane and policies everywhere. Placement becomes a scheduling decision rather than a painful migration project.
Cloud isn't disappearing from this picture. It's becoming the overflow valve instead of the default home for everything.
Last-Ditch Arguments Are Losing Steam
Three claims used to shut down this debate immediately. Today they're just speed bumps on the road to hybrid infrastructure.
Global failover sounds critical until you examine it closely. Reliability emerges from system design, not from premium SKUs. Agentic operations automate region health monitoring, DNS steering, cross-region snapshots, and disaster recovery testing. The game-changer is instant evidence generation. Agents compile recovery time objectives, backup logs, and change approvals into auditor-ready packages. Discipline still matters. You don't need a hyperscaler to achieve it though.
Enterprise identity management and guardrails have open-source equivalents now. Keycloak, Ory, SPIFFE, SPIRE, OpenFGA, Cedar, Vault, OPA, and Gatekeeper provide robust identity, policy, and secrets management. Agents wire these controls together end-to-end. They keep policy documentation synchronized with actual infrastructure. The organizational consolidation public clouds offer is convenient. It's no longer irreplaceable.
Special network fabrics like NVLink or InfiniBand matter for specific workloads. Training models across hundreds of GPUs demands exotic interconnects. Ultra-low latency inference needs careful topology planning. But that's workload classification, not a universal trump card. Most inference pipelines, data processing jobs, and application backends run fine on standard networks. Put specialized workloads on appropriate hardware. Move everything else off expensive hyperscale platforms.
Small Teams Just Got a Platform Rebate
Small companies historically chose cloud to avoid cognitive overload. Now agents dramatically reduce that burden instead.
You can bootstrap a complete platform in roughly a week. Terraform provisions infrastructure automatically. Talos or K3s provide lightweight Kubernetes. Argo CD handles deployments. Vault manages secrets. SPIRE handles service identity. Keycloak provides authentication. Cilium networks everything securely. Postgres operators run databases. MinIO offers object storage. Prometheus, Loki, and Grafana monitor the whole stack. AI writes the configuration for all of it.
Operations run on autopilot after that. Runbook agents execute safe remediations automatically. Compliance agents assemble access reviews, backup verification, and disaster recovery metrics. Cost agents report spending per thousand requests and flag sudden increases.
When infrastructure work reduces to writing prompts and clicking approvals, VPS economics crush on-demand pricing for most small business workloads. Managed services still help with genuine minute-scale bursts or heavy compliance requirements. They're optional now instead of mandatory.
The New Math: Price Outcomes, Not Instances
This transformation isn't ideological posturing. It's basic arithmetic taking over decision-making. Stop comparing virtual machine specifications. Start tracking what actually matters for your business.
Calculate cost per outcome across environments. Measure euros per thousand inferences, euros per thousand tokens, euros per gigabyte processed. Time-to-capacity shows agility—minutes versus days to provision GPUs or double a service tier. Reliability proofs include timestamps of successful restores and verified recovery objectives from recent drills. Change velocity tracks how quickly issues become production deployments with agent-generated pull requests. Operational toil counts pages per week and mean time to remediation.
Put these metrics on a unified dashboard. You'll discover exactly how much cloud capacity you genuinely need. It's rarely "all of it" anymore.
Will GenAI Devour Public Cloud?
Not through direct competition or hostile takeover. Through slow starvation instead.
Agentic programming and AI-powered operations are maturing rapidly. The premium you once paid for fully managed infrastructure becomes harder to justify with each passing quarter. Baseline workloads migrate to owned capacity or low-cost VPS providers. Bursts and edge cases still rent hyperscale capacity briefly. Cloud bills shrink from "foundational platform" to "pressure release valve."
GenAI already writes application code for developers. Now it's writing the operations playbooks, security policies, and compliance evidence that enable confident self-hosting. Software engineering jobs won't vanish overnight. Their fundamental nature is changing though. Public cloud won't die either. It will shrink toward its most defensible positions—genuine burst capacity, exotic network fabrics, and compliance automation you cannot or will not build yourself.
Everything else moves elsewhere. Steady workloads with proven patterns and cost sensitivity will get eaten by a new triumvirate. AI-powered operations, mature open source tooling, and cheap global VPS capacity are rewriting the economics of infrastructure. The cloud era isn't ending. It's just getting right-sized.