Enterprise AI Infrastructure: Joining the Queue at the Back

Enterprise want to own AI – are they joining at the back of the queue?

Enterprise want to own AI – are they joining at the back of the queue?

Enterprise AI Infrastructure: Joining the Queue at the Back

New research says more than half of enterprises plan to own at least part of their AI infrastructure – the research suggests they will find it hard to buy, hard to build and harder still to run.

We have previously examined how AI infrastructure is effectively being rationed – the largest hyperscale buyers secure equipment slots first, grid connections sit behind multi year queues, and a powered shell without a hardware delivery schedule is an idle asset. New research from Schneider Electric and Omdia adds a fresh perspective: mainstream enterprises, 55% of whom say they intend to own at least part of their AI infrastructure.

On paper, this looks like a wave of distributed enterprise demand – in practice it may look more like an enterprise sized crowd joining the queue at the back.

Is Control, not Just sovereignty driving the shift?

The Omdia survey covered 540 enterprise leaders, split 36% EMEA, 33% North America and 31% APAC, with fieldwork completed between January and March 2026. Respondents came from sectors including healthcare, financial services, life sciences, government, education and manufacturing.

The survey leads on sovereignty and compliance, but the underlying numbers tell a more practical story; Respondents cite data privacy and security (30%), reliability (22%) and regulatory compliance (21%) as their main reasons for keeping AI infrastructure in house. Omdia confirmed that the survey didn’t ask specific questions about data sovereignty, so the findings speak to these practical concerns rather than to sovereignty as such.

That points to a control story – these organisations want direct authority over who touches the hardware, how workloads are isolated and when systems go down – those operational guarantees are far harder to outsource than a regional cloud availability zone.

Wanting the Hardware Is Not the Same as Being Able to Run It

The same survey data exposes a gap between ambition and operational readiness; Respondents list scaling AI clusters (27%), cost (25%) and sourcing power and cooling infrastructure (22%) as their top deployment hurdles – a third point to a lack of in house expertise, and 32% struggle to integrate AI stacks into existing enterprise systems.

The consequence is visible in how few projects graduate from the lab – more than half of respondents report that at most one in five proofs of concept reach production, and only 23% say their AI models are delivering measurable business value today. Unsurprisingly, 59% rely on outside partners for AI strategy and 67% lean on external specialists for implementation. Enterprises want to own the kit, but they are still calling in third parties for the know-how to make it work.

The survey doesn’t say how many of the 55% intend to run their infrastructure themselves, and how many through a partner. Steven Carlini of Schneider Electric offers an expectation rather than a statistic: “Enterprises today are using a hybrid approach, combining owned hardware with colocation and the cloud. As AI workloads mature, we may see higher levels of AI IT become self-managed by enterprises, but initially more will be managed by a partner or colo.”

Is Infrastructure the Bottleneck?

It’s worth asking whether physical infrastructure is really the first thing holding buyers back – the release attributes stalled pilots to a mix of human and infrastructure challenges – Omdia’s Maxine Holt says in the release that “the lack of internal expertise remains a significant barrier and a leading reason proofs of concept fail.”

In reality, the physical kit is rarely the first thing that breaks an enterprise AI rollout; Power and cooling are necessary conditions, not sufficient ones – a rack of GPUs in a room that cannot cool it is not an AI strategy, but neither is a perfectly cooled, high-density rack with nobody on the payroll who knows how to run it.

Why Enterprises May Sit at the Back of the Supply Queue

The budgets reported in the survey sharpen the supply chain picture; Most respondents’ AI budgets run from hundreds of thousands to a few million dollars, with just over half exceeding $500,000 and 31% exceeding $1 million. Omdia told us it can’t provide a breakdown of spending on physical infrastructure, so these figures cover overall AI programmes, not hardware alone. Even so, they look small next to the multi-billion dollar purchase commitments that the largest buyers have already made with chip and component suppliers.

Enterprises sit downstream of those allocations – they buy through channel integrators and OEMs whose supply pipelines are governed by upstream contracts the enterprise buyer never sees. When manufacturers have to ration liquid cooling distribution units, switchgear and high-density racks between gigawatt-scale operators and mid-market buyers, capacity is unlikely to be shared equally. Time matters too. Carlini notes that “new builds will require more planning and going through permitting which will be lengthy.”

Who Fills the Execution Gap?

The most realistic path for most enterprises is a hybrid model: own the compute kit, rent the room and buy the operational skills. Colocation operators can provide high-density, liquid ready white space housing customer owned servers. Managed private AI providers can handle day two stack operations, and systems integrators can supply validated reference designs to strip out deployment risk.

Asked how demand might split between new builds, retrofits and colocation, Carlini was careful not to forecast: “While we can’t speculate on customer demands, we might anticipate there will be a choice between available grid power capacity and the cost and speed of retrofitting existing facilities.”

His view of who retrofits is practical: “Some enterprise organisations are already evaluating the feasibility of retrofitting, especially if they have in-house data center expertise. If not then they will go to colo.” The questions that push them that way are technical: “the compatibility of existing IT racks, power distribution, density, back up time needed, and if liquid cooling is required, may also drive them to colo.”

Schneider is clearly positioning for that middle ground, leaning on prefabricated modular data centres and reference architectures it is developing with Nvidia and AMD. It’s a logical play for a major infrastructure supplier.

The Operator Takeaway

For colocation providers, managed service operators and integrators looking to capture this demand without overbuilding on optimism, four practical things stand out:

Package small, high-density, owned hardware footprints. Offer pre-engineered blocks of a few liquid ready racks with transparent power, cooling and cross connect terms, tailored to enterprise budgets. Be ready to answer the retrofit questions Carlini lists: rack compatibility, power distribution, density, backup time and liquid cooling.

Bundle day-two operations with white space. Internal skills gaps are a leading reason pilots stall, so build managed monitoring and stack support into the commercial offer.

Qualify demand before committing capital. Ask prospects whether they have a funded production workload or only an exploratory pilot, and whether they have the staff to run it – base capacity expansion on signed commitments, not survey intent.

Use early allocation slots as a sales asset. Equipment and switchgear lead times remain a major differentiator. Operators that can guarantee delivery schedules and power certainty hold an advantage enterprises cannot secure on their own.

Disclosure: the research was conducted by Omdia on behalf of Schneider Electric, which supplies power, cooling and modular data centre infrastructure to this market.