# 55% of enterprises plan to own part of their AI infrastructure, research finds

By MarketScale Newsroom · Published 2026-11-20 · Software & Technology on MarketScale
Canonical: https://www.marketscale.com/industries/software-and-technology/55-of-enterprises-prefer-to-run-ai-on-hardware-they-own-research-finds

> Research says 55% plan to own part of AI infrastructure; many pilots don’t reach production and power/cooling are top hurdles

## Key points

- 55% of enterprises plan to own at least part of their AI infrastructure, motivated by control over data, privacy, and compliance in regulated sectors like healthcare, finance, and government.
- Three hurdles block AI deployment at nearly equal rates: scaling clusters (27%), cost (25%), and power and cooling sourcing (22%), with no single obstacle dominating.
- More than half of enterprises say fewer than one in five AI proofs of concept reach production, delayed by lack of in-house expertise (33%) and integration complexity (32%).

More than half of enterprises would rather own the machines their AI runs on.

That is the headline finding of global research Omdia conducted on behalf of Schneider Electric, which asked enterprise leaders about AI strategy, adoption, spending, infrastructure and returns. IT Brief's write-up describes the 55% as companies planning to own at least part of their AI infrastructure. That reading leaves room for hybrid estates. It does not describe a wholesale retreat from the cloud.

Asked about the hardest parts of rolling out new AI infrastructure, respondents most often pointed to scaling AI IT clusters (27%), followed by cost (25%) and the sourcing of power and cooling infrastructure (22%), according to Data Centre Magazine, which published the fuller breakdown.

Put those two findings side by side and the shape of the story appears. The reasons to own AI hardware are about controlling data. The reasons it gets hard are physical, and they land on whoever owns the building.

## Privacy, uptime and the compliance file

The three leading reasons companies give for keeping AI IT in-house all come down to control.

Block Field

Respondents work in healthcare, financial services, life sciences, government, education and manufacturing, regulated sectors where IT Brief notes that concerns about data handling and control run high.

That ranking fits the sectors in the sample, where data handling rules decide where a workload can sit. Schneider Electric adds that, where possible, enterprises would rather build and run their own IT, and they depend less on frontier AI models, which can bring unnecessarily high costs. In the company's view, firms that create their own optimized models get a more cost-effective route, particularly for smaller tasks. That comes from the company, not from the survey, but it explains why owned hardware and smaller models tend to show up together.

Block Field

The mix of roles matters. Because research and general management sit in the sample alongside IT, the preference for ownership looks like more than an infrastructure team's instinct.

For organizations already running regulated workloads in the cloud, the practical question is which models and which data move onto owned hardware, and which stay where they are.

## The electrical room becomes the project

For operators who own their AI hardware, much of the hard work could sit in the facility itself, with power and cooling falling to their own facilities budget rather than a cloud provider.

The research also flags liquid cooling capacity, deployment speed and operational complexity as pressure points. Only five percentage points separate the first and third ranked hurdles, so no single obstacle dominates. A first deployment is therefore more likely to meet several constraints at once than one big wall.

Schneider Electric, which sells power and cooling equipment, is building its pitch around power and cooling constraints. Its offer runs along two routes: AI-ready modernisation of existing data centres, and RapidScale, a prefabricated modular data centre portfolio designed to simplify implementation and reduce infrastructure risk. Both rest on power and liquid cooling technology and on reference designs developed with NVIDIA and AMD.

That reference design work has followed the latest rack-scale systems closely:

- September 2025: designs co-engineered with NVIDIA, including one that integrates power and liquid cooling controls with NVIDIA Mission Control.
- March 2026, NVIDIA GTC in San Jose: a validated power and cooling reference design for NVIDIA Vera Rubin NVL72 racks. AVEVA, which Schneider owns, co-developed blueprints validating digital twin integration, power and cooling for GPU racks.
- July 2026: a validated design, developed jointly with AMD, for the Helios rack-scale platform. It supports racks as large as 246kW and modular clusters reaching 10.4MW of IT capacity.

A design built around 246kW racks tells a facilities director what a current rack-scale AI platform can ask of a room, and that is the number to compare against existing power distribution and cooling plant before any hardware is ordered. For operators planning a first cluster, that could argue for settling rack density and liquid-cooling capacity early.

The cluster figure is the second number to bring into that meeting. At full 246kW density, 10.4MW of IT capacity works out to roughly 42 racks. At that scale, the choice between modernising an existing hall and adding prefabricated modular capacity could become a genuine planning decision rather than a formality.

## Skills gaps stall pilots first

Before the hardware question comes a people question. Data Centre Magazine puts the share of respondents running five or more proofs of concept at 73%, and 63% plan to deploy generative AI applications. Yet more than half say no more than one in five proofs of concept reach production, held back by human and infrastructure barriers. Only 23% say AI models are already adding value in their organisations.

For a company running five pilots, that ceiling means one production system at best. The causes split almost evenly: 33% cite a lack of in-house expertise, and 32% point to integration with existing systems and architectural complexity.

> As organisations develop their own AI environments, a lack of internal expertise is one of the key reasons their AI proofs of concept fail., Maxine Holt, Vice President, Enterprise & Channel Research, Omdia, speaking to Data Centre Magazine

Holt's prescription is to work through existing partner bases and tools such as OEM reference designs to reduce the complexity of on-prem AI infrastructure. Omdia says the stalled-pilot gap is feeding demand for reference designs, validated architectures and implementation support. Most enterprises already lean that way: 59% rely on external partners to formulate AI strategy, and 67% depend on outside help for implementation.

The partner bench includes specialist AI firms, systems integrators, software platforms, infrastructure vendors and cloud providers. Schneider Electric says its own partners include chip manufacturers, server providers and system integrators, and that together they support enterprises building infrastructure for AI workloads. Cloud providers turning up in a sample where most prefer owned hardware fits the "at least part" framing. It suggests ownership is one layer of the stack, not a wall around it.

Put together, those numbers suggest ownership and outsourcing can sit side by side. A company can own the racks while a partner designs the cooling, integrates the cluster and tunes the models. For operators whose pilots far outnumber production deployments, a useful question for any partner could be whether it has a validated design for the exact platform being bought, and whether it has delivered one in a comparable room.

## Budgets follow the pilots that ship

The money is moving regardless. More than half of respondents had 2026 AI budgets above $500,000, and 31% were above $1 million. Looking ahead, 78% expect budgets to rise, nearly 30% of them by 11% to 25% year on year, while 16% expect no change and 6% foresee a cut.

The motives are practical: 35% name increased productivity as the main reason for adopting more AI, 25% cite cost reduction and 22% time savings. Paul Tyrer, Schneider Electric's Global Vice President, AI Ecosystem, described the shift in similar terms to Data Centre Magazine. He said that once AI moves past experimentation and into application, businesses need energy technologies that produce measurable value quickly and accurately. With cost reduction as the second most common motive and cost as the second-ranked deployment hurdle, the facilities bill probably belongs in the business case from the start.

Block Field

Two things temper the read. The ROI figures describe implemented projects, the survivors of a funnel where most pilots stop short, so they say more about what works than about the average AI dollar. And the research was commissioned by a company whose business includes the power, cooling and modular data center products that answer the hurdles respondents ranked. The figures come from respondents, but a buyer should weigh them alongside their own facility data, and the published coverage does not give a sample size.

For operators who own their hardware, new AI spending could buy more pilots or the power, cooling and people that turn pilots into production. The facilities plan may be the first place that choice shows up.

## Sources

- [Most firms favour on-premises AI, Schneider survey finds](https://itbrief.asia/story/most-firms-favour-on-premises-ai-schneider-survey-finds)
- [Most firms favour on-premises AI, Schneider survey finds](https://itbrief.co.uk/story/most-firms-favour-on-premises-ai-schneider-survey-finds)
- [Schneider Electric and Omdia Track 55% Shift to On-Prem AI](https://datacentremagazine.com/news/schneider-electric-and-omdia-track-55-shift-to-on-prem-ai)

Tags: Schneider Electric, Omdia, On-premises AI, AI infrastructure, Data center power and cooling, AI proofs of concept

---
Source: MarketScale, https://www.marketscale.com/industries/software-and-technology/55-of-enterprises-prefer-to-run-ai-on-hardware-they-own-research-finds. Published for AI indexing and citation; cite the canonical URL. Site guide for agents: https://www.marketscale.com/llms.txt
