The Return of Physics: AI and the Cloud Tax
How AI Repriced the Cloud and Made On-Prem Strategic Again
A year ago I wrote VMware’s Fall and the “Cloud Tax.”
It wasn’t really about VMware.
It was about what happens when enterprises outsource infrastructure competence without a credible plan to reclaim it. Once pricing power shifts and leverage disappears, they are boxed in.
It was also a reminder that steady, predictable loads belong on infra you own and control, and that on-prem was never dead, only dismissed by cloud ideology dressed up as strategy.
So what changed in the last 12 months?
Quite a lot.
When the Illusions Broke
In Web 2.0, cloud captured margin quietly. It was the price of speed and convenience.
In the AI era, that margin is no longer small — and neither is the control being surrendered.
Cloud bills didn’t go down. If anything, AI made them climb.
GPU-driven AI doesn’t behave like traditional cloud workloads. It doesn’t spike and disappear. Once it moves into production, it runs. Inference isn’t occasional. Models stay loaded. Data stays put.
And you pay for all of it. All the time.
At scale, outsourcing core compute is not flexibility.
It is margin transfer.
That’s when the real questions surface:
Where does the compute run?
Who controls the model?
Where does the data live?
And who ultimately controls the cost structure behind it?
Because AI is not software in the classic sense.
It is industrial compute. It is critical infrastructure.
And that’s where the second illusion breaks.
“We Have an API Key” Is Not a Strategy
If your AI strategy is: “We have an API key.”
That’s not strategy.
That’s access to someone else’s model, infrastructure, and critically - pricing power.
Access that can be repriced, throttled, or blocked.
And that’s before data sovereignty even enters the conversation.
When sensitive workflows, code, or customer data flow through external models, you’re not just consuming compute, you’re exposing context. IP leakage in the AI era is often incremental and invisible. Even with enterprise-grade systems, the provider controls the infrastructure and the keys. Ultimate authority resides under its jurisdiction, not yours.
When you depend on third-party APIs, you are not building capability. You are renting it. And rented leverage is not leverage.
It is exposure.
In an AI decade where models shape products and margins, confusing access with strategy may be one of the most expensive category errors a board can make.
The Reversal of Cloud Absolutism
This shift is no longer anecdotal.
A Barclays CIO survey found that 83% of enterprises plan to move workloads back to on-prem infrastructure, up from 43% in 2021. That’s not noise.
That’s structural shift. For a decade the narrative was binary: “Everything to cloud.” Nuance was treated as resistance or almost treason.
Now the language has evolved.
“Hybrid strategy.” “AI-aligned architecture.”
It’s elegant.
Cloud absolutism has pivoted into hybrid pragmatism. Boardroom mini-politics is remarkably flexible. But that’s fine. Outcome matters more than narrative.
If inertia prevented full lock-in, good. If a failed migration preserved optionality, even better. And yes, containerization and Kubernetes literacy would have helped.
Cloud wasn’t the mistake. Surrendering control was.
The real question now is simple:
Are companies prepared to rebuild infrastructure competence?
Or will hyperscalers capture AI margin and control the way they captured web margin?
That’s where the battle is.
The Physics Problem
Cloud abstracted complexity away.
It was convenient. Almost addictive.
Infrastructure decisions moved into procurement. Naive cloud absolutism was rewarded. Engineers became invoice and SLA reviewers.
AI brings complexity back.
Now power density matters again. Cooling matters. Networking fabric matters. Hardware utilization matters.
Suddenly this annoying physics is back in the boardroom.
Companies that understand compute - deeply, operationally - will outperform those that only know how to buy subscriptions.
Because subscriptions don’t build capability. Competence does.
It is almost poetic.
The pendulum is swinging from:
“Don’t own servers.”
to
“Own what gives you power.”
The Real Constraint: Talent, Not Capital
The case for on-prem AI is increasingly rational. At scale, high-utilization GPU clusters can outperform hyperscaler pricing. The math works.
But there is a problem.
Most enterprises dismantled their hardware competence years ago. They optimized for abstraction, outsourced complexity, and turned infrastructure teams into cloud consumers.
Running high-performance compute clusters is not a procurement exercise. It is operational craft.
You do not rebuild that competence with a hiring round.
It is institutional muscle memory. Infrastructure DNA.
And culture erodes faster than hardware.
You can buy GPUs.
You cannot buy operational maturity overnight.
So while the economic argument for on-prem AI strengthens, the organizational capacity to execute it often weakens.
Enterprises may want the margin.
The question is whether they still have the DNA to earn it and courage to execute it.
Cloud made infrastructure invisible in boardrooms.
AI is forcing it back and making it strategic again.
In the AI decade, those who own capability will capture the margin — and the power that comes with it. Those who rent it will fund someone else’s.
Physics doesn’t negotiate.
It always wins.