Twenty years in. One interview. One blank whiteboard. What a missed answer taught me about where architecture is really heading.
Today I sat across from a VP for a role I'm qualified for on paper and in practice. Twenty years of building enterprise systems. I've been out of work for a while, and I walked in confident.
Then she asked the question: How would you architect this? Not in theory. Step by step. Users, identity, platform, integration, systems of record, data. The whole stack, in order.
I drew a blank.
I knew every piece of it. I've built every piece of it. But I answered the way I've worked for two decades: in fragments, in context, jumping to the interesting problem. She kept rephrasing the question, and each time I gave her pieces instead of the path. By the end I felt small. And somewhere in that feeling, I realized something else: I didn't want this job. Before we wrapped up, I told her so.
Afterward, I kept turning it over. She wasn't curious about how I'd approach the problem. She had one way of thinking about architecture, a single top-down template, and any answer that didn't arrive in that shape simply didn't register. That rigidity is what turned me off. Good architecture is collaborative: you reason through trade-offs together, and there's more than one valid path into a design. A leader who only accepts one shape of answer is practicing exactly what's dying: architecture as a fixed template. And today I learned the value of fit. It cuts both ways, and this time I was the one deciding that table wasn't one I wanted to sit at.
The industry data says the same thing my interview did.
Cloud migration was sold as a one-way trip. In 2026, it's a revolving door.
The Barclays CIO survey found roughly 83–86% of enterprise CIOs planned to pull at least some workloads back from public cloud. Read carefully, that number is less dramatic than the headlines: one repatriated app satisfies it, and analyst estimates of full exits sit closer to 8–9%. But that nuance is exactly the point. Nobody is "leaving the cloud." Everybody is placing workloads, one at a time, based on utilization, egress, latency, and law.
The economics forced it. Flexera's 2026 State of the Cloud report puts wasted IaaS and PaaS spend at 29%, reversing a five-year decline. 37signals cut its annual cloud bill from $3.2M to about $1.3M after moving onto roughly $600K of its own servers. Meanwhile, Uptime Institute's latest survey still counts about 45% of IT workloads in corporate data centers. The "everything in the cloud" end state never arrived. And sovereignty finished the job. In the EU, DORA now requires financial firms to assess concentration risk and maintain tested exit plans. Data residency is no longer a checkbox in a region dropdown. It's a question of who can legally compel access.
A discipline built on picking one platform and designing natively for it cannot survive a world where every workload needs a defensible exit.
Classic cloud architecture was designed around general-purpose compute: stateless web tiers, elastic VMs, managed databases. AI doesn't fit that shape.
Gartner expects AI providers to spend nearly $1 trillion on AI infrastructure in 2026 alone, with another $6 trillion forecast through 2030. That money isn't buying more of the old cloud. It's buying liquid-cooled GPU campuses, high-density power, and AI fabrics, much of it from neoclouds like CoreWeave and Nebius that barely existed as cloud providers five years ago.
That splits the map. Training wants massive, bursty, specialized capacity. Inference wants to sit close to users and close to data. Regulated data wants to stay home. No single hyperscaler region is the right answer for all three, and the architect who memorized one provider's well-architected framework has little to say about any of them.
The ground is also unstable. Gartner predicts more than 80% of neocloud providers will lose independence by 2030, and warns that over $3.5 trillion of AI infrastructure could face write-downs. Designing a ten-year architecture on top of a provider landscape that may consolidate in four is not architecture. It's speculation.
The last refuge of the cloud architect was the design itself: the Terraform modules, the landing zones, the golden paths. That's being automated too. According to env0 research cited by DevOps.com, 71% of cloud teams have seen a measurable jump in infrastructure-as-code volume directly attributable to generative AI. Agents now read provider API specs, generate Terraform or Kubernetes manifests from a plain-language request, open the pull request, and reconcile drift on their own.
The output isn't trustworthy yet. Veracode's Spring 2026 testing found only about 55% of AI code-generation tasks produce secure code by default, even as syntax correctness passed 95%. But notice where that leaves the human. The job is no longer to design the VPC. It's to write the policy that decides which VPCs an agent is allowed to create.
That's not cloud architecture. That's governance, platform engineering, and risk management wearing a cloud badge. Gartner projects 80% of large engineering organizations will have dedicated platform teams by 2026, up from about 45% in 2022. The architect's seat is being absorbed into the platform.
What replaces cloud architecture is something less glamorous and far more useful: deciding, workload by workload, where compute should run, for how long, and under whose law.
In this model the hyperscaler, the neocloud, the colo cage, the edge site and the on-prem rack are all just venues. The unit of design isn't the platform. It's the workload, scored on utilization, growth, egress, coupling, latency and regulatory tier, and revisited every year or two as prices and rules change.
| Cloud architecture (2010–2024) | Workload placement (2026→) | |
|---|---|---|
| Starting question | Which provider? | Where does this workload belong? |
| Unit of design | The platform and its services | The individual workload |
| Default venue | Public cloud | Whatever wins on cost, latency and law |
| Exit plan | Rarely written | Required, and tested |
| Who writes the infra | Architects and engineers | Agents, under human-written policy |
| Core skill | Knowing one vendor's catalog | Economics, governance and portability |
Portability becomes a first-class design constraint: open formats, loosely coupled services, identity that spans venues. Not because multi-cloud is fashionable, but because the cheapest or safest venue for a workload today may not be the same one in 2028.
If the vendor catalog no longer makes you an architect, what does? The ability to explain a system end to end, in order, from the person using it to the data it produces. Agents can generate the Terraform. They can't yet tell a VP why the layers sit where they do.
This is the answer I couldn't give in the interview. Here it is for a typical enterprise process platform:
Users: employees, managers, partners (web, mobile, Teams)
│ SSO (Entra ID / SAML / OIDC)
┌──────────▼──────────────────────────────────────────┐
│ LOW-CODE PLATFORM (process engine) │
│ forms/UI · workflow & case mgmt · rules · approvals │
│ reusable component library · audit log │
└──────────┬──────────────────────────────────────────┘
│ APIs + events
┌──────────▼──────────────────────────────────────────┐
│ INTEGRATION LAYER (API gateway / iPaaS / event bus)│
│ sync APIs for lookups · async events · batch loads │
└───┬─────────────┬──────────────┬─────────────┬──────┘
▼ ▼ ▼ ▼
HRIS ERP/Finance CRM Doc mgmt
(systems of record: own master data)
│
┌──────────▼──────────┐
│ DATA PLATFORM │ process metrics, reporting, AI
└─────────────────────┘
Cross-cutting: identity · DLPNotice what's missing: no provider name. Every box could run on a hyperscaler, a neocloud, or a rack down the hall. The design holds regardless of venue.
The pattern generalizes into a seven-step walk that works for almost any "how would you architect this?" question:
Who and how. Who uses it, and through which channels?
How they get in. Identity, SSO, roles and permissions.
Where the work happens. The core platform: UI, workflow, rules, approvals, audit.
How it talks to everything else. Sync APIs for lookups, events for async, batch for bulk.
Who owns the truth. Systems of record, and which one masters which data.
What we learn from it. The data platform: metrics, reporting, AI.
What wraps all of it. Security, DLP, observability, governance, scale.
Start with the user and work inward. If I'd had that sequence ready, I'd have spent the interview showing twenty years of judgment instead of searching for an entry point. That's not cloud architecture. It's just architecture, and it isn't going anywhere.
To be fair to the corpse: cloud spending is still growing. Most repatriation is partial, and full exits remain rare. Elasticity is still unbeatable for bursty demand, and managed services like serverless warehouses have no easy on-prem equivalent. Plenty of teams pulling workloads home are really just paying for years of skipped rightsizing.
So the cloud is fine. What's dead is the architecture of default: the assumption that public cloud is where everything goes, that one provider's framework is a career, and that designing infrastructure is a human-only craft.
If you carry the title "cloud architect" today, the move is clear:
Learn the economics. Build break-even models that include egress, refresh cycles, staffing on both sides, and the cloud bill that keeps running during migration.
Own the policy layer. If agents write the Terraform, the person who writes the guardrails holds the real design authority.
Design for the exit on day one. Open formats, portable pipelines, identity that spans venues.
Get fluent in AI infrastructure. GPUs, inference placement, power and cooling are now architecture topics, not facilities trivia.
I walked away from that role, and I don't regret it. But I'm keeping the lesson: in that moment, I was still answering for an era that's over. Cloud architecture is dead. The architects who can walk the whole stack, out loud and in order, are just getting started. I plan to be one of them.
Share this article
The most effective leaders don't wait for clarity — they create it. Here's how to build decision-making frameworks that thrive in ambiguity.
Read More
Building AI strategies that actually stick. How to move beyond proof-of-concept to scalable adoption that transforms how your organization thinks and delivers.
Read More
Understanding coordinated inauthentic behaviour in the age of AI — and why every Canadian leader needs threat literacy now.
Read More