Every analyst firm is telling you to build converged AI infrastructure. They're all wrong.
I've spent 20 years as a CISO, CIO, and CTO, making technology decisions that either propelled organizations forward or locked them into regret. Now I run Phronia, an independent analyst firm, which gives me a clear view of both sides: what practitioners actually live with, and what the industry tells them to do.
On AI infrastructure, those two perspectives are completely misaligned.
WHAT THIS MEANS FOR THE CIO/CTO/CISO
Your architecture must assume model churn every 90 days. Any infrastructure decision that doesn't account for this will become technical debt before implementation is complete.
Every integrated AI platform decision must be justified against the switching cost. If you can't quantify what it costs to leave, you can't evaluate whether you should enter.
Your north star architecture is fungibility, not optimization. In stable markets, you optimize. In rapidly evolving markets, you preserve optionality. AI is not a stable market.
THE INSIDE PERSPECTIVE
When I was in the chair, I learned that the worst technology decisions were the ones that locked me in. Every time I built something rigid, I regretted it within 18 months. Sometimes within six.
I remember deploying what was supposed to be our strategic platform for the next five years. Eighteen months later, the vendor was acquired, the roadmap was abandoned, and we were stuck with infrastructure that couldn't adapt to what we actually needed. The switching costs were enormous. Not just financial, but organizational. We'd built processes, trained people, and integrated systems around something that no longer served us.
That pattern repeated throughout my career. The more integrated the solution, the more painful the eventual migration. The more strategic the partnership, the more leverage the vendor had when our needs diverged from their roadmap.
THE OUTSIDE OBSERVATION
Now I sit on the other side of the table. I see the vendor decks. I hear the analyst recommendations. I watch enterprises make decisions based on narratives that have no grounding in operational reality.
Every vendor is selling integrated AI stacks. Converged infrastructure. Unified platforms that promise to handle everything from data ingestion to model deployment to inference at scale. It sounds great in a PowerPoint. It photographs well for the trade publications.
It's a disaster waiting to happen.
Every vendor selling an integrated AI stack has a business model that depends on lock-in. They don't want you to be able to swap components. They don't want you comparing their inference engine to a competitor's. They want you committed, contractually and architecturally and organizationally, to their ecosystem. That's great for their revenue predictability. It's terrible for your ability to adapt.
THE UNCOMFORTABLE TRUTH
AI is too new, too poorly understood, and too unpredictable for integrated stacks.
Let me be specific about what I mean. I use five different AI tools right now. Not because I'm indecisive or because I enjoy complexity, but because the technology is evolving so fast that last month's best choice is this month's second-tier option.
I switch between models every three weeks to a month. Different company. Different architecture. Different strengths. The model that was best for code generation three months ago has been surpassed. The model that was best for analysis has new competitors. The model that was best for creative work keeps getting leapfrogged.
Anyone making a model choice today needs to be prepared to make a different choice in three months. Six months at the absolute outside. That's not pessimism. That's the observed reality of how this technology is developing.
You cannot adapt that quickly with an integrated stack.
When your inference engine is tightly coupled to your data pipeline, which is tightly coupled to your orchestration layer, which is tightly coupled to your model serving infrastructure, changing any component means changing everything. The integration that was supposed to be a feature becomes a constraint. The unified platform becomes a unified prison.
THE ECONOMICS NOBODY DISCUSSES
The cost of inference is dropping precipitously. The cost of training is dropping. The cost of fine-tuning is dropping. What costs a dollar today will cost a dime next year and a penny the year after.
When you lock into an integrated stack, you lock into today's economics. You commit to price points and performance characteristics that will be obsolete before your implementation is complete. You lose the ability to arbitrage between providers as costs drop unevenly across the market.
The organizations that will win in 2026 and beyond are the ones that can move workloads fluidly between providers based on current economics. The ones that can swap a model component the moment something better emerges.
THE ARCHITECTURE VIEW
The lock-in isn't in the compute. It's in four places:
- SDKs and APIs. Vendor-specific code throughout your stack.
- Data formats. Proprietary schemas that don't export cleanly.
- Orchestration. Workflow definitions that assume specific components.
- Contracts. Terms that penalize migration.
Build your abstraction layer to isolate all four.
SIGNS YOU'RE ALREADY IN TROUBLE
Use this diagnostic. If three or more apply to your organization, you're building the integrated stack that will sink you:
- Your architecture team insists on consistent vendors across the AI stack. They're optimizing for comfort, not adaptability.
- Model-specific SDKs are required to integrate anything. Every SDK is a lock-in vector.
- You cannot explain the swap cost for any component. If you can't quantify it, you can't manage it.
- AI workloads are tied to a single cloud region or provider. Geographic diversity is strategic optionality.
- Your AI budget assumes stable pricing for 12 to 18 months. Economics change quarterly.
- Your team measures integration depth, not swap speed. You're measuring the wrong thing.
Scoring: 0-2, monitor closely. 3-4, immediate action needed. 5-6, critical risk.
WHAT I'D TELL MY FORMER SELF
If I had known then what I know now:
I would never again select a platform that required vendor-specific SDKs. Every SDK is technical debt with a vendor's logo on it.
I would treat model selection like financial portfolio management. Diversify. Rebalance regularly. Never go all-in on a single position.
I would budget for churn, not stability. The stable AI infrastructure budget is a fiction. Plan for constant motion.
I would measure my architecture team on swap speed, not integration depth. The metric you measure is the behavior you get.
ZERO COMMITMENT ARCHITECTURE
This is not Zero Trust. This is Zero Commitment.
Your infrastructure should be so loosely coupled that swapping any component is an operational decision, not an architectural crisis. You should be able to change your inference provider while your applications remain completely unaware. You should be able to swap models without redeploying anything. You should be able to shift workloads between clouds based on cost or performance without modifying code.
That's the architecture that survives 2026. Everything else is technical debt waiting to explode.
THE BOTTOM LINE
Integrated stacks are a vendor strategy, not an enterprise strategy. Convergence is an optimization play in stable markets. AI infrastructure is not stable. It's in constant motion.
Build for fungibility. Assume churn. Preserve optionality.
Your career might depend on it.