AI infrastructure has spent the last several years racing to solve a relatively straightforward problem that happens to be brutally difficult at scale: feed increasingly large models with enough data and compute to train them, then make the resulting systems available for inference.
That problem is not going away. But VAST Data CEO and co-founder Renen Hallak sees another one arriving behind it. Models themselves are becoming infrastructure resources, agents will increasingly use and modify those models, and the systems underneath them will have to understand far more than where data lives and which accelerator has capacity.
“Models are the new data,” Hallak says. More specifically, the weights are becoming the valuable resource. In an enterprise filled with AI agents, there will not simply be one model sitting above a pile of corporate data. Different models will be suited to different jobs, carry different costs and have access to different information.
He explains that agents will continually fine-tune them as new information arrives, creating derivatives that need to be tracked along with what data was used, when the changes occurred and which policies governed access. An infrastructure layer will need to know which model should receive a prompt, which GPU already has that model loaded, whether the necessary context is nearby and whether the data involved is even permitted to touch that model.
That turns model management into something much closer to a systems problem. Hallak describes organizations eventually operating with agent employees alongside human employees, with those agents learning continuously through inference and fine-tuning. At sufficient scale, model lineage becomes operationally important.
An organization may need to reconstruct which derivative of a model produced a result months earlier, what information went into that version and what the agent was authorized to see at the time. This is part of why VAST has continued moving upward from storage into database capabilities, orchestration, functions and now deeper awareness of the models and agents consuming those resources. The infrastructure cannot simply schedule compute. It increasingly has to understand what that compute is doing.
Security becomes considerably more complicated in this model because there are valuable assets on both sides of an AI transaction, Hallak adds.
An enterprise may want to run inference against proprietary or regulated information that cannot leave its environment. At the same time, a model provider has little interest in handing an enterprise unrestricted access to the weights that represent its own intellectual property. Hallak points to confidential computing as the mechanism for bridging that gap: keeping enterprise information within an environment it controls while protecting model weights through encrypted memory and cryptographically controlled execution.
These ideas existed in CPU-centric enterprise computing, but the challenge now is extending them across GPUs, AI systems, models and weights, with the infrastructure layer coordinating the process end to end.
The implications stretch well beyond securing today’s enterprise inference workloads. Hallak expects enterprises themselves to become model owners as their agents learn from proprietary information and accumulate specialized intelligence. Those models and weights become corporate IP, potentially something an organization can use internally, lease or monetize without giving up control of the underlying asset. From there, the infrastructure problem expands again. Agents will need mechanisms to interact, transact and potentially pay one another through enormous numbers of small transactions.
Systems designed around human activity suddenly have to support machine activity occurring at radically greater frequency and scale.
Hallak puts that eventual difference at six or seven orders of magnitude. That is the part of the AI infrastructure discussion that may ultimately matter most. Today’s AI factories are already exposing physical limits around power, datacenter construction, supply chains and compute.
The next phase adds another scaling dimension above all of that: potentially millions of agents, proliferating model derivatives, continuous fine-tuning, distributed inference and machine-to-machine economic activity. Storage, databases, schedulers, security systems and transaction infrastructure were built around human-scale assumptions. If agents become a meaningful part of how organizations operate, many of those assumptions disappear.
The interesting question, then, is no longer simply how large the next AI factory can become. It is how the infrastructure underneath AI evolves when models are dynamic resources, weights are valuable IP and agents become active participants in the enterprise.
Hallak’s view is that the software infrastructure layer sitting between applications and models above and compute, data and physical infrastructure below becomes the natural place to manage that complexity. If that view proves right, the next phase of AI infrastructure will be defined as much by controlling, securing and understanding models as it has been by supplying the GPUs needed to run them.
