search
close-icon
Data Centers
PlatformDIGITAL®
Partners
Expertise & Resources
About
Language
Login
banner
Blog

AI is leaving the lab. Here’s what it needs.

04 November 2026
author-photo
Chris Sharp

Chief Technology Officer

There’s a big difference between getting an AI application to work in a controlled environment and making it work reliably across an enterprise.

Once AI moves into production, the demands on infrastructure change. Compute must be available in the right locations, with sufficient power and cooling. Models must connect quickly and reliably to enterprise data and applications. And businesses must be confident that the entire system will perform as expected at scale.

We're seeing this reflected in how customers approach their deployments. Large campuses provide the concentrated compute required for training, while inference runs closer to the data and end users. That means planning capacity across multiple locations, with the connectivity to bring it all together.

Three things stand out in making that transition successful: power, connectivity, and a clear path to production.

Power: Supporting the workload

Finding enough power is difficult. Making it usable for AI brings another set of challenges.

AI systems consume considerably more power per rack than traditional enterprise equipment, generating heat that conventional air cooling can't manage efficiently. For high-density deployments, liquid cooling is becoming the default.

We've now delivered more than 115 MW across 25+ live liquid-cooled deployments in Digital Realty data centers, with contracted capacity approaching a gigawatt. That includes a 200 MW deployment earlier this year and another 78 MW this quarter (3Q26).

These aren't capabilities reserved for future developments. Our existing data centers and solutions are supporting AI workload at scale, giving customers options beyond waiting for new facilities to come online.

Power availability will continue to influence where AI capacity gets built, but the location of that capacity also determines how effectively enterprises can use it.

Connectivity: Every millisecond matters

Consider what happens when an AI agent completes a task. It might make 20 to 50 model calls, pulling information from different sources before producing a response.

If each call takes 50 milliseconds less, that's a saving of one to two-and-a-half seconds across sequential calls. When we multiply that across thousands of interactions, the network becomes a significant factor in application performance.

Enterprise AI must also operate across environments. This means connecting data held in private infrastructure, models running on dedicated hardware, and applications spread across multiple clouds.

Digital Realty’s ServiceFabric® connects these environments, giving customers an open, programmable way to manage how their workloads communicate. ServiceFabric spans more than 875 connected data centers and 375 cloud on-ramps, with programmable capabilities available since June 2026.

Blackfuel's deployment with Digital Realty in Barcelona shows this in practice. Its AI infrastructure combines liquid cooling, Dell systems, and AMD GPUs, with ServiceFabric providing connectivity.

We're also seeing customers bring large campuses and distributed capacity together. One recent deployment combines a 64 MW campus with 14 MW across global locations, supported by 1,600 cross-connects.

Power gives AI somewhere to run. ServiceFabric gets it where it needs to go.

Path to production: Test, validate, and optimize the infrastructure environment

An AI pilot can demonstrate that a model performs well on a particular task. It won't necessarily show how the application behaves with live enterprise data, sustained demand, or the power and cooling requirements of a production environment. These are critical variables I'd want to establish and test before committing significant CapEx on a deployment.

At the Digital Realty Innovation Lab (DRIL), we co-design deployments with customers, validate the architecture, prove the economics, and prepare for deployment at scale. The labs sit inside operating data centers, so testing takes place under production conditions.

Every lab uses liquid cooling, with no air-cooled test benches. ServiceFabric is fully integrated, allowing customers to test private AI and hybrid cloud connections as they'd operate in production.

Our partners bring much of the technology into the labs. Customers can test their own workloads on AMD EPYC processors and Instinct GPUs, Cisco's full-stack AI Pods, Lenovo's Neptune liquid cooling and Supermicro systems.

Vcinity provides data movement over ServiceFabric, ePlus supports AI design and integration, and Vultr offers GPU capacity. ORCA Computing also has photonic quantum systems running in London.

With compute, networking, cooling, and the facilities already validated together, customers can focus on testing their own workload.

Our DRIL sites are live to customers in Northern Virginia, London, Tokyo (through MC Digital Realty, our joint venture with Mitsubishi Corporation), and São Paulo. With an additional DRIL expected to open in Singapore in November 2026.

Digital Realty has hosted more than 400 tours, supporting engagement that has contributed to production AI deployments at our facilities. Many of the enterprises deploying tests in DRIL are building focused clusters. Validating and optimizing how to run specific applications alongside existing systems, often across several locations.

DRIL gives customers and partners a space to validate solutions in a live colocation environment. This also means that any infrastructure constraints, which often hold production AI back, can be addressed.

Our 2026 Global Data Insights Survey found that 40% of IT leaders identified infrastructure as their biggest AI constraint, up from 9% in 2024. This reinforces the need to test components together before deploying advanced workloads.

From pilot to production

We expect hundreds of megawatts of AI demand to come online in the next 12 months, along with the connections that'll make that capacity accessible.

For enterprises, the immediate priority is understanding what their applications require and proving that the infrastructure can deliver it.

If you're preparing to move an AI workload into production, bring it into the DRIL. Test it using the partner technology you're considering, with real cooling, power, and connectivity.

Already deploying at scale? Contact us to talk about where that infrastructure should sit and how to connect it.

Learn more about the DRIL in this video.

play
Tags