Garima Kapoor is the Co-founder & Co-CEO of MinIO, the only pure-play exascale object store for secure, enterprise AI data.
getty
Here is the uncomfortable truth the industry keeps avoiding: buying GPUs is only half the battle. The harder challenge is extracting real value from every cycle. Many organizations are now GPU-rich but data infrastructure-poor, and that imbalance has become one of the biggest barriers to enterprise AI performance.
GPUs are no longer a scarce resource. Organizations can provision compute almost anywhere. What remains scarce and genuinely differentiating is a data foundation that is fast enough, open enough and scalable enough to keep those GPUs fully utilized. The race for AI advantage has shifted upstream, from silicon to storage architecture, and many enterprises have not adjusted.
AI changes the physics of a technology stack. It requires constant data movement, persistent context and real-time feedback loops that legacy file systems were never designed to support. When enterprise AI projects stall, the bottleneck is often not the model or the compute infrastructure. It is the data layer underneath.
The numbers reflect that reality. More than half of enterprise practitioners cite storage and networking bottlenecks as their top AI constraint. Nearly two-thirds of organizations have not yet begun scaling AI across the enterprise despite significant investment. Many companies made the compute investment but neglected the architectural foundation required to make those systems perform.
The cloud-first era created a false assumption that choosing a cloud provider was equivalent to choosing an architecture. AI does not operate in one environment. It runs across edge deployments, private clouds, neo clouds and public cloud environments simultaneously.
Increasingly, the best economics for AI infrastructure are found at the edge or in private cloud environments, as the cost of moving large volumes of data through public clouds compounds quickly. At AI scale, data gravity becomes unavoidable.
That is why object storage has become foundational. It creates a common data layer across environments, so workloads can move without forcing teams to rebuild pipelines or redesign access models whenever compute shifts.
It also aligns with the data mix AI depends on. AI workloads consume structured data in tables alongside unstructured data such as text, images, video, logs and sensor streams. Object storage can support both at scale without the complexity penalties that make traditional file systems difficult to manage as data expands from petabytes to exabytes.
That does not mean object storage replaces every system. Databases, file systems and specialized low-latency platforms still serve important roles. The shift is that AI requires a durable, scalable foundation that can unify mixed data types, preserve context and operate consistently across environments.
AI infrastructure spending exceeded $300 billion in 2025, with storage and networking investments growing almost as quickly as compute. The capital is moving in the right direction, but the architectural thinking often is not.
Too many organizations continue layering AI onto infrastructure designed for a different era, treating AI as another workload rather than as a force that fundamentally changes how the technology stack must operate.
That is why the gap between organizations successfully scaling AI and those struggling to move beyond pilots continues to widen. The winners are not necessarily the companies with the most GPUs. They are the organizations that recognized early that AI infrastructure must be built differently: with an object-native data foundation; structured and unstructured data coexisting in a unified high-performance layer; and storage operating as an active part of the AI stack rather than a bottleneck at its edge.
Before expanding GPU investments or increasing model spending, enterprise leaders should pressure-test the data foundation supporting those systems.
1. Where does the data actually live? If training data, inference data, logs, embeddings, model outputs and governance metadata are scattered across clouds and application silos, the architecture will create friction before the model ever runs.
2. Can compute move easily to the data? AI increasingly spans edge, private cloud, neo cloud and public cloud environments. If every change in compute location requires rebuilding pipelines or paying large egress costs, the architecture is not cloud-neutral. It is cloud-constrained.
3. Does the system preserve context? AI performance depends on more than data volume. Related signals must remain connected so models can preserve meaning rather than process information.
4. What happens at scale? A pilot can hide architectural weaknesses. Leaders should evaluate how systems behave as data grows from terabytes to petabytes, more teams access shared datasets and more workloads require real-time or near-real-time performance.
5. Is storage treated as part of the AI stack or as a passive back end? In traditional enterprise architecture, storage existed behind the application layer. In AI, storage is an active performance layer. If the data foundation cannot feed compute, preserve context and operate consistently across environments, AI ambitions will eventually outpace the architecture that supports them.
Organizations that answer these questions honestly will make better infrastructure decisions. They will know whether they need to modernize the data layer, re-architect around an object-native foundation or limit certain AI workloads until the underlying systems can support them.
The question is no longer whether an organization’s AI strategy is ambitious enough. The real question is whether the data foundation can execute it. Architecture is no longer the back half of the AI conversation. For organizations serious about AI performance at scale, it has to be the starting point.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。