The infrastructure that looks optimal for an AI workload today may be the wrong place to run it a year from now. As models, costs and regulation change faster than traditional infrastructure planning cycles, the ability to move workloads may prove more valuable than making the perfect initial choice.
There is an uncomfortable mismatch at the heart of enterprise AI. Organisations are being asked to make infrastructure decisions with consequences measured in years for a technology whose economics can change in months. A model can be replaced, inference pricing can fall, a new accelerator can alter the performance equation and regulation can suddenly determine where data is allowed to travel. An architecture optimised around today’s assumptions can therefore become tomorrow’s constraint.
That matters because AI is escaping the relatively controlled environment in which many enterprise experiments began. Production systems increasingly draw upon proprietary data while connecting to external models, specialist infrastructure and cloud services. Agentic systems complicate matters further because a single business process may involve several models and data sources, with actions passing between different environments before a task is completed. The infrastructure beneath AI is consequently becoming less like a destination and more like a constantly changing set of relationships.
Kevin Egan, Senior Director, Technical Solutions at Equinix, argues that enterprises should therefore stop treating infrastructure selection as an attempt to predict the eventual shape of AI. “What really matters is not getting every decision right today; it’s preserving the flexibility to change as models, providers, regulations and economics inevitably evolve,” he says. That changes the objective. Instead of asking where AI should live, organisations increasingly need an architecture that allows the answer to keep changing.
The workload no longer has one natural home
Cloud computing encouraged enterprises to think about infrastructure largely through abstraction. Applications could move away from corporate data centres and consume resources when required, while standardisation promised to remove much of the complexity beneath them. AI is exposing the limits of that simplicity because different parts of the same workload can have very different infrastructure requirements.
Training may gravitate towards an environment where GPUs are available at the right price. Inference serving European customers could need to remain within a particular jurisdiction, while retrieval may work best close to the proprietary data being interrogated. Another workload could move between model providers as token costs or performance change. None of those decisions is necessarily permanent.
The scale of the existing fragmentation is striking. Research conducted by Foundry and sponsored by Equinix surveyed 1,643 senior IT leaders and found that half of enterprise application workloads already operate in multicloud or hybrid environments. Respondents reported using 36 cloud and SaaS providers on average, while data and analytics processing was the workload most commonly prioritised for hybrid or multicloud deployment. AI is arriving, in other words, into an enterprise landscape that was already distributed rather than creating that complexity from scratch.
It is also exposing weaknesses within it. Cost management was cited by 45 per cent of respondents as a challenge when integrating on-premises systems with cloud and SaaS environments, while 40 per cent identified security or compliance gaps. The detailed survey data also shows network performance becoming more problematic as multicloud adoption deepens: high latency was reported as an issue by 41 per cent of organisations with heavier hybrid or multicloud adoption, compared with 32 per cent among the remainder.
AI makes those infrastructure decisions more consequential because data movement itself can become part of the cost and performance equation. Inference prices vary between providers and locations, while routing data unnecessarily through cloud environments can create egress charges that are insignificant during experimentation but difficult to ignore at production scale. Performance can also change according to physical proximity and network congestion, making the cheapest compute instance a poor bargain if reaching it introduces another bottleneck.
“The ‘best’ environment for any one AI transaction will not be the best environment for the next,” Egan argues. That is particularly relevant to agentic AI, where chains of inference, reasoning and action can alter connectivity requirements as work progresses. Infrastructure optimisation therefore becomes continuous rather than something completed during deployment.
Flexibility moves below the software layer
Multicloud flexibility has traditionally been discussed as a software problem. Orchestration platforms abstract underlying infrastructure so applications can consume different resources without developers having to understand every component beneath them. AI does not invalidate that approach, but it introduces physical constraints that software cannot simply wish away.
Data still has a location. Networks still introduce delay. Moving large datasets still costs money, and sovereignty cannot always be satisfied merely by choosing a region from a cloud provider’s menu. As AI becomes embedded in sensitive enterprise processes, organisations may need to understand not only where information is stored but the path it takes between their data, a model and the application consuming its output.
That is why Egan sees distributed AI infrastructure emerging across cloud, edge and on-premises environments rather than consolidating around a single destination. “AI infrastructure constraints increasingly emerge below the software layer,” he says. “Enterprises must account for network proximity, data gravity, sovereignty requirements and ecosystem adjacency.”
The distinction becomes important when sovereignty moves beyond data residency. Keeping information in a particular country provides one form of control, but an application interacting with several providers can still send traffic across infrastructure outside the desired jurisdiction. Network-level controls that constrain those paths can therefore become part of sovereignty architecture rather than connectivity being treated simply as transport.
This is also where Equinix’s role becomes more prominent. Egan argues that carrier-neutral colocation can provide a meeting point between enterprise infrastructure and an AI ecosystem that now stretches well beyond the hyperscalers. Model providers, neoclouds and networks increasingly sit alongside the major cloud platforms, giving enterprises greater choice but creating another layer of interconnection to manage.
Equinix says eight of the top ten AI model providers and four of the top five neoclouds are deployed within its ecosystem. Its Distributed AI Hub is intended to provide a reference architecture for connecting enterprise data with those distributed AI resources, while Equinix Fabric provides private software-defined connectivity between environments. Fabric Intelligence extends that proposition towards greater visibility of network conditions and more informed decisions about routing and workload placement.
The important point is not simply that enterprises can connect to more providers. Choice has little value if exercising it requires rebuilding the architecture each time. Neutrality becomes strategically useful when it gives an organisation the ability to change provider, location or workload placement without unravelling everything around it.
The network becomes part of the AI decision
The network has often been treated as plumbing beneath the more interesting layers of enterprise technology. Distributed AI makes that increasingly difficult because model performance cannot be separated from the movement of the data feeding it. A powerful accelerator in the wrong location can still produce a poor application.
The Foundry research provides some evidence of that tension. Among organisations integrating or planning to integrate generative or agentic AI, cost and cloud or hybrid network complexity were the joint-leading networking pain points, each cited by 38 per cent. Data access control and compliance followed at 34 per cent, while data synchronisation also emerged as a significant concern. These are not model problems; they are consequences of connecting increasingly complicated AI systems to existing enterprise environments.
Enterprises are responding by becoming more deliberate about interconnection. Almost half of respondents access network endpoints through carrier-neutral data centres, while 27 per cent use specialised multicloud networking services. More importantly, organisations with intentional hybrid or multicloud networking strategies reported tangible benefits: 55 per cent cited faster data transfer and 51 per cent improved support for AI workloads.
The direction of travel is towards networks that become more programmable as the applications above them become more dynamic. Egan points to automated capabilities such as Equinix Fabric Intelligence as an early step towards infrastructure able to respond to changing network conditions rather than relying entirely on static configuration. As agentic systems themselves become capable of making decisions about tools and resources, the logical extension is infrastructure that can respond with similar flexibility while remaining within policies established by the enterprise.
That does not mean allowing AI to move workloads indiscriminately in pursuit of milliseconds or marginal cost savings. Governance becomes more important as infrastructure becomes more dynamic because organisations need to define what may move, where it may move and which considerations take precedence. Flexibility without control simply replaces lock-in with unpredictability.
Designing for decisions that have not been made
The temptation during a period of rapid technological change is to search for certainty. Enterprises want to know which models will dominate, whether workloads should run in public cloud or private infrastructure and where the economics of inference will eventually settle. The difficulty is that committing too heavily to any one answer may create precisely the constraint they are trying to avoid.
Egan’s argument is that infrastructure strategy should instead preserve options. “The next phase of AI infrastructure will be defined by architectures that can adapt continuously as models, economics, regulations and ecosystems evolve,” he says. The competitive advantage does not necessarily belong to the organisation that selects today’s cheapest provider or fastest model, but to the one that can respond when either ceases to be the right choice.
That represents a subtle but important change in infrastructure thinking. Standardisation once promised efficiency by reducing variation, while cloud offered abstraction from much of the physical technology beneath enterprise applications. AI is introducing variability back into the equation because workload placement increasingly depends on circumstances that refuse to remain fixed.
The consequence is that infrastructure flexibility can no longer be treated as insurance against a bad technology decision. It becomes part of the architecture itself. Enterprises cannot know exactly where their AI workloads will need to run several years from now, and increasingly they do not need to. The more important decision is whether the infrastructure they build today leaves them free to choose again tomorrow.
