Enterprise AI Is Leaving Public Cloud, and This Week's Vendor Response Sells Anywhere, Not Cloud

Cloud & Edge Computing 7 OCT 2026 1,601 words 8 min read By news.jualin.id

Enterprise AI Is Leaving Public Cloud, and This Week's Vendor Response Sells Anywhere, Not Cloud

Introduction

For most of the last decade the direction of travel in enterprise IT was one-way. Workloads moved outward, from the server closet to the data center to the public cloud, and the industry's answer to almost every infrastructure question was a variation on "get it into a public cloud and stop thinking about it." That era is ending, and not because enterprises have suddenly rediscovered a love of cold rooms and CRAC units. It is ending because AI workloads have a different shape than the transactional workloads that were migrated before them.

AI workloads are sustained rather than bursty, they are governed rather than merely provisioned, and their most valuable input is frequently the enterprise's own regulated data. Those three properties break the economics that justified public cloud in the first place. When Cloudera published its Great AI Re-Architecture survey on 11 August 2026, the headline figure was arresting on its own: 95% of surveyed organizations had delayed or canceled at least one AI initiative in the past year because of data governance, compliance, or regulatory problems. The same release recorded that 66% of respondents had moved AI workloads back from public cloud to private cloud or on-premises infrastructure in twelve months, and that 84% reported AI-driven increases in infrastructure costs.

Those August numbers set up what has been a remarkable five days of announcements. On 1 October IBM opened a self-hosted deployment option for its agentic development platform. On 5 October PwC and Cohere announced a global alliance whose explicit deployment menu reads private cloud, on-premises, and air-gapped. On 6 October Zetaris launched a cloud-based data harness that exists specifically to stop enterprises from centralizing their data. Three different companies, three different layers of the stack, and one shared commercial strategy: sell the customer control over where the AI runs, and stop treating public cloud as the default destination.

Main Content

The most common mistake in covering this shift is to call it repatriation. The data does not support that framing, and the vendors themselves are careful not to use it. Cloudera's survey found that 97% of respondents still move data between environments at least monthly, and that only 25% plan to prioritize a hybrid-first architecture over the next two years — a number that sounds decisive until you notice it is a plurality rather than a mandate. Reporting on the same survey data, Businessworld quoted Biswajeet Mahapatra of Forrester noting that enterprises are now evaluating infrastructure "through the lens of business risk and workload characteristics rather than simply infrastructure cost." The unit of decision has changed. It is no longer the organization that picks a cloud; it is the workload.

That distinction matters because it resolves the obvious objection. Public cloud is not being abandoned, and the announcements this week make that explicit. Cohere's own framing in its announcement of the PwC partnership is not that the public cloud is unsafe but that enterprises want to "maintain control over sensitive data and infrastructure" while still scaling. IBM's release makes the same argument in the language of software development rather than data center real estate: customers can run supported models locally or connect to external model services through hybrid configurations, choosing per workload. Aidan Gomez's framing in the PwC release is the cleanest statement of the position — enterprises need AI that delivers capability "without forcing them to give up control of their data, infrastructure or security."

Note who is not in that sentence. Nobody argues the public cloud is going away. What is being renegotiated is the default.

The economics underneath this are straightforward, and they are worth stating plainly because they are less ideological than the discourse suggests. As EPAM's cloud practice lead put it in the reporting, "running continuous, high-throughput inference 24/7 on cloud GPUs can cause operational budgets to spiral out of control." Sustained baseline load is the single worst fit for a consumption model that prices for peaks. Add unpredictable data egress to that — a real cost for any enterprise that moves proprietary corpus material into a cloud region repeatedly — and a predictable owned or private capacity line starts to look rational for the workloads that never sleep. Meanwhile the workloads that are bursty, the large-scale pre-training runs and the research sandboxes, remain firmly public cloud territory, because no enterprise justifies building a GPU cluster purely for occasional experiments.

This is why the workloads are dividing rather than the strategy flipping. Sensitive data, regulated workloads, low-latency processing, proprietary intellectual property, and high-volume production inference all move toward dedicated infrastructure. Foundational training and early experimentation stay in the cloud. That is a two-market split, not a migration, and it explains why the Cloudera release reports 66% moving workloads back while simultaneously finding that enterprise architects are now planning hybrid-first rather than committing to a single destination.

IBM's specific move is instructive about where the pressure actually lands. The self-hosted option for IBM Bob targets organizations "whose sensitive source code, regulated data or critical systems limit their ability to use externally hosted AI development services." The SiliconANGLE report on the announcement notes IBM cited its own June Institute for Business Value research in which 68% of executives said meeting data residency and sovereignty requirements across geographies is challenging. Source code is the asset. Development context and build artifacts are the payload. This is not a data center capacity story — it is a story about the fact that the code a company writes is among the most sensitive things it owns, and it now travels through an AI development tool whether the enterprise intended that or not.

The Zetaris announcement on 6 October attacks the same problem from the opposite direction, and the contrast is the interesting part. Rather than moving compute toward the data, it removes the need for the data to move toward the compute. The company's press release frames the problem as a data-plumbing problem — enterprise data "scattered across cloud, on-premises and edge systems, often in formats too inconsistent to trust once pulled together" — and sells federated querying that lets agents read data in place, with policy and access control travelling alongside it. It claims a reduction in total cost of ownership of up to 67%, and envisions more processing happening close to where data is created, including across edge device networks.

Both approaches are commercially convenient, and that is the honest caveat. IBM is selling the option to keep development workloads in-house. Zetaris is selling a cloud service that reduces the number of reasons to leave it. PwC and Cohere are jointly selling governance consulting attached to a platform that can be deployed in all three models. Every one of these moves increases the vendor's surface area and none of them reduces cloud spending in the customer's favour without a counterweight somewhere else. The re-architecture is happening, but so is a repricing of who gets paid for it.

The 95% delay figure is the most under-discussed number in this whole story. Nearly all surveyed organizations are actively using AI, and nearly all of them have stalled something. Governance, not compute, is the binding constraint, and the resolution of that is organizational work before it is a procurement decision — which is precisely the gap the PwC alliance is structured to sell into.

Conclusion

The pattern across these announcements is not that the public cloud lost. It is that "the cloud" stopped being a place and became a default, and defaults break when the workload class that inherits them changes. Continuous inference, regulated data, proprietary source code and air-gapped requirements are all workload classes the default never fit well, and AI made them mainstream.

The direction of travel is toward placement decisions made per workload, with the vendor's job shifting from providing a single destination to supporting several. Whether that produces genuinely better infrastructure economics or merely a more sophisticated form of lock-in depends on decisions enterprises have not made yet. For now, the notable thing is that the industry's loudest cloud-era assumption — that more workloads in public cloud is the success metric — has quietly stopped being the assumption.

Images

Enterprise storage and compute racks in a dark server room. This is a research-institute facility photographed in Tucson, Arizona — illustrative of the kind of owned capacity enterprises are building toward, not any installation operated by the companies discussed in this article. No physical security controls are visible in the frame.

What "on-premises" frequently looks like in practice: one mixed-generation rack in a converted office closet, running tape backup, file and network services from a single cabinet, cooled with flexible duct rather than precision air conditioning. A third-party corporate local site, photographed as documentation of its own setup.

Staff working at open racks in a third-party telecommunications data center. The badges, cabling and populated cabinets illustrate the operational labor that stays with infrastructure wherever it is deployed — this is not the facility of any company named in this article.

References

Related coverage on news.jualin.id tracks the buildout side of this shift; the AI category covers the model and tooling layer that is driving it.