A funny thing happens when a trend gets big enough: people who rarely agree start arriving at the same conclusion.
Recently, Chamath Palihapitiya asked whether “on-prem is the new cloud,” and Michael Dell replied: “True.” That short exchange captured something we’ve been hearing quietly and consistently from large enterprises for years:
If AI is going anywhere near your most sensitive information, the default architecture can’t be “ship it to a vendor’s cloud and hope for the best.”
For a lot of organizations, that’s not just a preference. It’s the line between possible and not happening.
Cloud AI breaks down where enterprises care most
When people say “cloud AI doesn’t work for enterprises,” what they usually mean is: cloud-first AI doesn’t work for the hardest enterprise requirements.
The moment you’re dealing with:
- corporate strategy
- board materials
- product roadmaps
- M&A plans
- financial records
- pricing models
…you run into the same set of constraints:
- Privacy & security that can’t be “best effort”
- Data sovereignty requirements across regions and business units
- Compliance that demands provable controls, not marketing promises
- Permissioning that has to map to real enterprise identity and access rules
- Vendor risk concerns that show up in every serious procurement process
Cloud can be a great operating model for many workloads. But for “crown jewels” data, enterprises don’t want a debate. They want a guarantee.
And the cleanest guarantee is simple:
Keep the AI inside the walls.
We heard this before we built anything
Before we wrote a line of code at Atolio, we spent time talking with leaders across the Fortune 1000. Security, IT, data, and the business owners who actually live in the mess of modern enterprise information.
The pattern was clear:
If an AI solution is going to touch our most sensitive internal data, it must run on our infrastructure, not in a vendor’s cloud.
That wasn’t paranoia. It was experience.
Enterprises have already seen what happens when tools that were built for convenience get asked to carry the weight of regulated, high-stakes data. The architecture that looks “fast” at the beginning becomes a liability at scale, especially when the questions shift from “Can it demo?” to “Can it be trusted?”
SaaS-first was easier. It just wasn’t acceptable.
Most enterprise AI search products went SaaS-first, and it's not hard to understand why. It's...
- easier to build
- easier to scale
- easier to deploy
- easier to sell
But for organizations with strict regulatory and security mandates (think banks, defense contractors, healthcare, government, and other public sector organizations with critical infrastructure and highly-sensitive IP), a SaaS-first architecture is often a dealbreaker before the conversation even starts.
Not because those teams are anti-cloud as a philosophy.
Because their risk model doesn’t allow “external by default.”
On-prem doesn’t mean “old.” It means “controlled.”
There’s a misconception that on-prem equals slow, clunky, and outdated. In reality, modern on-prem (or private infrastructure) can be:
- cloud-native in design
- containerized and orchestrated
- integrated with enterprise IAM and permissions
- deployed in your VPC, data center, or secure environment
- compatible with the models you choose and the controls you require
The point isn’t nostalgia.
The point is control.
Who has access, where data flows, how it’s processed, what’s logged, what’s retained, and what can be audited. Those are enterprise requirements, not “nice-to-haves.”
The market is catching up and retrofits are expensive
What Chamath and Dell are saying publicly is what many enterprise teams have been saying privately for years.
The difference is timing.
A lot of SaaS-native platforms are now trying to “retrofit” their products for enterprise-grade requirements. Private deployments, strict permissioning, security controls, data locality, model choice, auditability.
Those are not features you bolt on.
They’re architectural decisions.
At Atolio, we started there.
We designed around the hard constraints from day one because that’s what enterprises told us they needed if AI search was going to be trusted with real internal knowledge.
Where this goes next
I expect the next wave of “enterprise AI” will look less like consumer AI packaged for business, and more like:
- AI that runs where the data already lives
- AI that honors enterprise permissions by default
- AI that can be audited, governed, and proven
- AI that doesn’t require compromising on sovereignty or security
In other words: AI that behaves like enterprise infrastructure.
If you’re building or buying AI systems right now, here’s a simple test:
If your AI can’t live inside your environment, can it really be trusted with your most important knowledge?
For many enterprises, the answer is increasingly obvious.
