The Autonomous Data Infrastructure Imperative: Paul Speciale on Why Enterprise AI Needs a New Operating Model
Q1. Scality ADI introduces the concept of “autonomous operations” through Scality Guardian, but with a deliberate human-in-the-loop principle where agents surface recommendations and humans approve every decision. That is a meaningful architectural and philosophical choice at a moment when many vendors are racing toward full automation. What is the reasoning behind that boundary, and where do you see the risks of infrastructure autonomy that moves faster than human oversight can follow?
A: Scality ADI’s human-in-the-loop approach is a deliberate safeguard rather than a limitation on autonomy. Scality Guardian can analyse workloads and recommend changes, such as shifting data across performance tiers or adjusting protection policies, but it stops short of executing them without human approval.
The reasoning is that at exabyte scale, infrastructure decisions are not just technical optimisations; they directly intersect with compliance, data sovereignty, resilience, operational continuity, and cost governance. AI-driven infrastructure must remain aligned with enterprise intent, not merely system efficiency. Fully autonomous execution risks drifting away from that intent, particularly in environments where policies, regulatory constraints, and business priorities are layered and sometimes in tension.
There is also a trust and observability dimension. Many enterprises operate complex, accumulated infrastructure estates where cause-and-effect relationships are already difficult to trace. Introducing full automation without a human checkpoint can reduce interpretability at precisely the moment when auditability and explainability are becoming more important, not less.
The real danger of “faster-than-governance” autonomy is not just incorrect decisions, but compounding errors at scale: a suboptimal optimization executed across petabytes can amplify cost, performance, or resilience issues before they are even detected. In AI-augmented infrastructure, speed is valuable, but unchecked speed can outpace validation, compliance review, and operational rollback capability.
Human approval therefore preserves a clear accountability boundary while still allowing AI systems to continuously improve recommendations based on observed workload behaviour.
Q2. The press release makes the bold claim that “AI has broken the old storage model.” That is a strong statement given that object storage has been evolving continuously for over a decade. What specifically about the combination of training, inference, RAG, multimodal agentic workflows, and KV cache for distributed inference makes the demands fundamentally incompatible with traditional tiered storage architectures, and not just a matter of scaling what already exists?
A: The point isn’t that object storage hasn’t evolved (it clearly has), but that AI workloads have changed the access model faster than tiered architectures can adapt.
Training, inference, RAG, multimodal agent workflows, and KV cache all collapse the old assumption that data is accessed in predictable hot-warm-cold phases. Instead, data is continuously recontextualised across the entire AI lifecycle: high-throughput streaming during training, ultra-low-latency stateful access during inference (including KV cache reuse), and highly dynamic multi-source retrieval for RAG and agentic reasoning, often within the same pipeline, and sometimes even within the same request sequence.
Traditional tiering treats storage as a temporal optimisation issue: move data across media over time based on assumed access decay. AI turns this into a concurrent demand challenge: the same datasets may need bulk throughput, sub-millisecond latency, and high-frequency random retrieval simultaneously, without lifecycle policies being able to “pre-stage” data effectively enough.
In that sense, the incompatibility is not about scale alone but about semantics. Tiering assumes predictability over time. AI systems increasingly operate with adaptive, context-driven access patterns where relevance is computed at runtime rather than pre-classified by age or frequency. That breaks the very core abstraction on which classical storage hierarchies were initially built.
Q3. Scality ADI is delivered as open-code software source code available for inspection and governed contributions, alongside outcome-based SLAs spanning availability, performance, protection posture, power consumption, and operational efficiency. Those are unusual commitments in enterprise infrastructure. What do they signal about who the intended customer is, and what does it actually mean in practice for a government agency or regulated enterprise to hold a vendor accountable to an outcome-based SLA on something as complex as autonomous storage operations?
A: This combination is aimed squarely at highly regulated, sovereign, and large-scale enterprise environments where infrastructure is not only operationally critical but also subject to audit, oversight, and long-term accountability requirements.
Open-code access provides inspectability and governance. In the Scality ADI model, customers are not forced to treat autonomous operations as a black box, Instead, they can inspect behaviour, validate logic, and integrate it into internal compliance and security review processes. That is particularly important in government as well as in regulated industries where software provenance and behavioural transparency are mandatory constraints, not optional features.
Outcome-based SLAs shift the vendor relationship away from “you run the system yourself” or “trust our managed service execution” toward a model where the provider is accountable for measurable system outcomes (availability, performance, data protection posture, energy efficiency, and operational efficiency), regardless of the underlying automation mechanisms used to achieve them.
In practice, for a government agency or regulated enterprise, this means they are not expected to micromanage infrastructure operations, but they retain auditability, contractual enforceability, and policy control over results. The complexity of autonomous behaviour is abstracted away operationally, but not legally or structurally. That distinction is critical: accountability does not disappear with automation, but rather is relocated to measurable outcomes that can be independently verified.
Q4. Power consumption and sustainability appear as first-class design constraints in Scality ADI, with real-time power telemetry at system, node, and workload levels, and near-zero power consumption on tape and cold storage tiers. This is relatively rare in storage infrastructure conversations, which tend to focus on performance and cost. What is driving that priority now, and how significant a factor is energy constraint becoming in how large enterprises and government organizations actually make infrastructure decisions?
A: Energy has become a first-class constraint because it is now a structural limit on digital infrastructure growth, not just an operational expense.
At multi-petabyte and exabyte scale, power availability, cooling capacity, and data centre density increasingly define how much infrastructure can physically be deployed, regardless of budget. This is especially true in AI-heavy environments, where GPU utilisation, data movement efficiency, and storage placement are tightly coupled. Inefficiencies in storage translate directly into wasted compute cycles and therefore wasted energy.
The Scality ADI perspective highlights that storage is no longer separate from compute in energy terms. It is part of an integrated energy system supporting AI workloads. If data is not where it needs to be, when it needs to be there, GPUs sit idle or are underutilised, and the energy cost of those idle cycles becomes significant at scale.
As a result, enterprises and governments are increasingly treating energy efficiency as a procurement and architectural constraint alongside performance, durability, and cost. This is reinforced by sustainability reporting requirements, ESG commitments, and in some cases sovereign infrastructure strategies that explicitly limit or regulate energy consumption growth.
Real-time power telemetry and tiered approaches such as ultra-low-energy tape or “ice cold” storage make energy consumption observable and governable at the workload level, rather than an aggregate facility metric. That shift enables infrastructure decisions to be made with explicit awareness of energy impact across the entire data lifecycle.
Q5. Your previous conversation with ODBMS.org (*) established that production AI is fundamentally a data pipeline problem and that storage is the bottleneck nobody plans for. ADI represents Scality’s architectural answer to that problem. Looking at where enterprise AI deployments are headed over the next three to five years, particularly as agentic AI matures and inference workloads diversify, what aspects of data infrastructure do you believe are still being systematically underestimated, and what decisions should infrastructure leaders be making today that most are not yet making?
A: What’s still underestimated is how decisively AI reshapes infrastructure into a continuous data-flow system, where the binding constraint is not compute availability but sustained, low-latency, high-throughput data movement across the entire AI lifecycle. Three things are consistently undervalued.
First, data movement is still treated as subordinate to compute scaling, yet training, RAG, and agentic workflows depend on constant retrieval, recombination, and re-ingestion, which makes data flow the dominant operational constraint.
Second, inference is becoming stateful and storage-integrated: as models grow more agentic, inference maintains context, accesses KV caches, retrieves external knowledge, and coordinates multi-step reasoning, embedding storage directly into the execution path rather than the backend.
Third, lifecycle and governance complexity is underestimated and data is simultaneously active for training, fine-tuning, retrieval, audit, and compliance, often under different policy and sovereignty regimes, which breaks the linear “hot to cold” model. For infrastructure leaders, the implication is to prioritise continuous data mobility, unified access across workloads, and minimal manual orchestration, rather than optimising isolated tiers or individual systems. Storage should be treated as an active component of AI systems: a dynamic data plane that supports inference, training, and agentic execution, not a passive repository of records.
Resources
(*) ODBMS.org MAY 4, 2026 Production AI is a Data Pipeline Problem: Paul Speciale on Why Storage is the Bottleneck Nobody Plans For.
……………………………………………………..
Paul Speciale, CMO Scality
Over 20 years of experience in Technology Marketing & Product Management. Key member of team at four high-profile startup companies and two fortune 500 companies.
Sponsored by Scality