How to Solve for Distributed Data in the AI Era

As enterprises work to prepare their AI strategies, there are still many unknowns they have to account for:

• They don’t know exactly which AI use cases they’ll pursue in the future.
• They don’t know how they’ll scale up each AI workload, how much it will cost, or where the resources to scale AI will reside.
• They don’t know which ecosystem partners they’ll work with to execute their AI strategy.
• They can’t know how technology will evolve and improve in the future.
• Most importantly, they don’t know exactly where all their data is, what data is appropriate or eligible for AI use, and what’s needed to make that data AI-ready.

These days, traditional compute is essentially a commodity you can get anywhere, but high-performance compute for AI is still very specialized. It often requires advanced liquid cooling capabilities and scaled power, which makes it very complicated and expensive to deploy.

Since many enterprises are still experimenting with AI and trying to get answers to the unknowns above, they’re reluctant to commit to the up-front CAPEX required to deploy infrastructure themselves. Instead, they’ll likely begin testing use cases in the public cloud, and then move these workloads to GPU as a Service (GPUaaS) providers or other neoclouds to meet their AI compute needs.

No matter what your AI compute landscape ends up looking like, that compute can’t function without the right data to process. You’re going to need agile interconnection capabilities to move that data to the GPUs quickly and securely, regardless of where it resides.

The work you do today to start building that AI-ready network architecture will continue to pay off whether you stick with public cloud or GPUaaS, or eventually graduate to your own AI hardware once you‘ve verified the return on capital investment. In fact, it’s far more important to get your data AI-ready first than it is to worry about what hardware you’ll use to process it, as many organizations who have suffered through failed AI projects can attest. As I’ve often said, GPUs don’t matter if your data isn’t ready.

If you build the right data architecture today, you can ensure you’re prepared to support massive datasets from distributed data sources in the future, regardless of the exact nature or location of that data.

How has AI changed the infrastructure equation?

The advent of generative and agentic AI is forcing enterprises to reevaluate how they manage their IT infrastructure, and that certainly includes their data platforms and network architecture.

For years, IT leaders have understood that a centralized data architecture makes it more difficult to utilize the right datasets in the right places. But thanks to growing AI adoption, support for distributed data is no longer a “nice to have.” It’s absolutely critical.

Many AI use cases rely on large volumes of data, and we can safely assume those datasets will grow even larger in the future. AI workloads are inherently distributed, and enterprises need to move data quickly between systems of record, unstructured data corpora, and inference compute in different locations. If that data must cross unreliable, low-bandwidth network connections, there’s a high risk of unacceptable latency, resulting in poor application performance and a negative user experience. To prevent this, enterprises must build and maintain an optimized AI data platform.

Different subtasks involved with running an AI data platform include:

• Redacting sensitive data so that it’s not improperly trained into AI models.
• Deduplicating datasets to avoid unnecessary network traffic.
• Ingesting data into the right systems in the right places.
• Applying governance principles to avoid compliance penalties and other risks.
• Selectively consolidating datasets to avoid data sprawl.
• Federating data processing where data must remain distributed.

To build an effective AI data platform, enterprises must maintain an agile, secure, high-performance network and then regionally position key inference-related components within distributed AI hubs. This ensures that AI data is ready to move between distributed endpoints, including AI ecosystem partners, without latency becoming an issue. It also places AI-ready datasets in the optimal locations for AI agents to quickly use this data in reasoning and grounding, thus greatly reducing the overall time required for multi-step tasks.

What does an ideal data platform look like?

Building a future-ready AI data platform includes strategizing for data sovereignty. Enterprises must be able to connect with partners, securely and compliantly consume or share data wherever it resides, and leverage different infrastructure models depending on the maturity of the workload. They must do all this while maintaining full custody, control and observability over their data.

As enterprises increasingly pursue agentic AI, they’ll likely need to work with many different AI ecosystem partners. This could include GPUaaS and public cloud providers, as mentioned earlier, but it could also include foundation model providers, SaaS platforms, customers, supply chain partners, data marketplaces and more. This can create new sovereignty challenges, as any time data moves into an external provider environment, there’s risk of losing control over that data. For instance, moving datasets into cloud-native storage must now be viewed in the context of regulatory compliance, with these regulations changing every year. Any data placement strategy must include flexibility and mobility to meet this challenge.

To overcome these challenges and ensure a sovereign AI data platform, enterprises can deploy a private, interconnected storage environment and data governance platform that they maintain complete control over. When the need arises to move data into the cloud or to other AI ecosystem partners, they can replicate temporary copies of that data, while leaving the authoritative datasets in their own storage environment.

They can simply delete the temporary copies when they’re no longer needed, instead of worrying about how to get them back out again or the high cost of egress. Therefore, they’re free to move their data across their AI ecosystem while maintaining control over it.

We call this concept an authoritative data core. It’s not a monolithic storage silo that resides in one specific place; rather, it’s made up of a series of interconnected storage environments adjacent to cloud providers and other ecosystem partners. Enterprises have increasingly been moving to this methodology to address the challenges of maintaining a holistic data governance platform while optimizing for the costs of redundant data copies and cloud transit. This is a dependency for enterprise AI success.

The example above shows an enterprise just beginning its AI journey. The authoritative core can connect to the enterprise’s on-premises classical IT infrastructure, as well as multiple clouds, neoclouds and SaaS providers. This shows that AI sovereignty does not have to mean isolation.

Equinix Distributed AI Hub helps address distributed data challenges

We recently announced the Equinix Distributed AI™ Hub, a new framework that serves as a convergence point for distributed AI assets, including AI-ready data. Now, enterprises can bring together their public and private datasets on a robust, connected governance foundation and use that data across their distributed AI environments. It’s built upon Equinix connectivity solutions, such as our virtual interconnection offering Equinix Fabric® and our network functions portfolio Equinix Network Edge. These solutions make it possible for enterprises to use their AI-ready data in all the places they need to, today and in the future.

The Equinix Distributed AI Hub also brings together the best of the Equinix AI ecosystem. This includes storage providers that offer AI-ready data solutions. For instance, the Dell AI Data Platform is available at Equinix to help our joint customers break through silos and unleash the full power of their AI datasets on a global scale.

Maintaining global control and governance over data, including the ability to store and access it in all the right locations and share it with all the right partners, is an important part of what it means to achieve data sovereignty. To learn how leading organizations are balancing data control and AI acceleration, read the analyst report from Enterprise Strategy Group: Optimizing AI in the Sovereign Era.