Lior Koriat, CEO of Quali – Interview Series

Lior Koriat, CEO of Quali, is a seasoned technology executive and entrepreneur with nearly two decades at Quali, where he has helped shape the company from its early engineering and operations foundations into a global provider of cloud-agnostic Environment as a Service solutions for DevOps, IT, and platform teams. Since becoming CEO and Board Member in 2011, he has led Quali’s expansion across the U.S., Europe, Asia, and Israel, building on earlier roles as VP R&D, COO, and CEO, where he established engineering, sales, customer support, legal, business development, and international operations. His background also includes founding and leading Intellitech Engineering, an acquired systems and mechanical engineering company serving defense and civilian customers, as well as mentorship roles with Google Launchpad Accelerator and UC Berkeley’s Sutardja Center for Entrepreneurship & Technology, where he advises startups on product strategy, go-to-market execution, DevOps, and scaling.

Quali develops infrastructure automation and platform engineering software designed to help organizations deliver governed, self-service cloud environments without slowing down development teams. Its Torque platform provides a catalog-based self-service layer for platform teams, enabling approved blueprints, role-based access controls, lifecycle management, cost attribution, policy enforcement, and real-time visibility across environments. Quali positions Torque as a way to standardize infrastructure delivery across multi-cloud and hybrid-cloud setups, with support for use cases including software development, testing, demos, training, proof-of-concept deployments, MLOps, agentic AI, CI/CD pipeline automation, and GPU-as-a-Service. Its broader product portfolio also includes CloudShell, which turns infrastructure into reusable low-code or no-code building blocks for complex environments across on-premises and cloud infrastructure.

You founded and scaled multiple technology companies before leading Quali, and your career spans military systems, robotics, infrastructure engineering, and cloud automation. How have those experiences influenced your vision for building AI infrastructure platforms, and what convinced you that this was the right moment to focus on AI-native infrastructure management?

My career has consistently centered on complex systems that have to operate reliably under demanding conditions. I began in robotics and aerospace, developing automation and simulation systems for defense applications, where repeatability and governance were operational requirements rather than engineering preferences. Those environments taught me that the hardest problems emerge where multiple systems and processes have to work together consistently over time.

That perspective carried into Quali. Over the past decade, enterprise infrastructure has become increasingly distributed across different layers. Every new layer solved a specific technical problem, but it also introduced another management interface, another operational workflow, and another source of complexity. Organizations became very good at creating infrastructure, yet much less effective at governing it once it existed.

AI accelerated that imbalance, and operational challenges increasingly determine whether an AI initiative reaches production or remains stuck in the pilot stage.  That is why I believe AI-native infrastructure management has become necessary. The challenge is operating AI infrastructure continuously across heterogeneous environments while maintaining governance, security, and cost discipline while optimizing utilization.

Sovereign AI has rapidly become a priority for governments and enterprises alike. What do most organizations misunderstand about sovereignty, and why do you believe infrastructure governance is becoming just as important as model ownership?

Many Sovereign AI discussions focus on data residency and infrastructure ownership. Those are necessary components, but they address only part of the operational reality.

Once organizations begin deploying autonomous agents, the primary question shifts from where infrastructure resides to how AI systems operate inside that infrastructure. Modern agents provision resources, retrieve information, initiate workflows, and interact with production systems. Every one of those actions requires policy enforcement.

That changes how sovereignty should be evaluated. Owning the infrastructure does not automatically provide operational control if AI systems can execute actions without deterministic governance. Enterprises need to demonstrate who initiated an action, what policies governed it, what resources were accessed, and how those decisions can be audited. The challenge is operating AI infrastructure continuously across heterogeneous environments while maintaining governance, security, and cost discipline while optimizing utilization.

I expect infrastructure governance to become as strategically important as model development because it provides the operational framework that allows organizations to scale AI responsibly. As enterprises adopt multiple models, multiple clouds, and increasingly autonomous workflows, governance becomes the layer that keeps those environments secure and accountable across the entire infrastructure estate.

AI agents are increasingly being given the authority to provision resources, deploy workloads, and make operational decisions. What guardrails must be in place before enterprises can safely trust autonomous systems with critical infrastructure?

Many organizations are already deploying AI agents in production. The question is how those agents operate within clearly defined boundaries.

The first requirement is deterministic policy enforcement at the execution layer. Agents should only be able to perform actions that comply with predefined infrastructure policies. Those policies cannot rely on the model deciding whether an action is appropriate. They need to be enforced independently of the model itself.

Access control is equally important. Persistent permissions create unnecessary operational risk. Access should be provisioned for the specific task being performed, limited to the resources required, and revoked automatically once the work is complete. That model allows organizations to expand automation without expanding their attack surface.

Every action also needs a complete audit trail. As AI systems become more autonomous, enterprises need to understand which agent performed an action, when it occurred, what policy authorized it, and what changes were made. That level of traceability supports operational troubleshooting as well as regulatory compliance.

I also believe autonomous infrastructure should operate within governed environments rather than across unrestricted infrastructure. Well-defined environments provide explicit resource boundaries, lifecycle controls, and operational policies before an agent begins executing workloads. That allows organizations to increase automation while maintaining predictable operational behavior as AI deployments continue to scale.

You often discuss the shift from traditional automation to what you call an intelligent control plane. What fundamentally changes when infrastructure begins operating around intent rather than static rules and workflows?

Traditional automation executes predefined steps. That works when infrastructure is predictable, workloads are stable, and changes happen through human-managed workflows. AI infrastructure does not operate that way.

AI workloads are dynamic. GPU demand changes quickly, environments are created and torn down frequently, and agents may need access to resources for specific tasks. An intelligent control plane understands the purpose of an environment, who owns it, what it costs, what policies apply, and whether the live state still matches the intended state.

That context changes infrastructure management. The system can detect drift, recommend optimization, enforce policy, and manage lifecycle decisions against the intended outcome rather than a static script.

GPU infrastructure has become one of the biggest bottlenecks in AI deployment. Do you believe the industry is approaching GPU utilization incorrectly, and what operational changes could dramatically improve efficiency without simply adding more hardware?

The industry often treats the GPU problem as a procurement problem. In many enterprises, the more immediate issue is utilization.

We see organizations with GPU clusters that are pre-allocated for peak demand, sit idle between training runs, or remain attached to environments after the work has ended. That creates high cost without matching business value.

Improving GPU efficiency starts with lifecycle control. Environments should be provisioned on demand, matched to the workload, attributed to a team or project, and torn down automatically when the task is complete. Enterprises also need visibility into utilization, cost per workload, and the business purpose behind each environment.

Adding hardware may still be necessary, but unmanaged GPU capacity only expands the cost problem.

Many enterprises now operate across public cloud, private cloud, on-premises environments, and edge infrastructure. How do you see hybrid AI environments evolving over the next five years, and what challenges remain unsolved today?

Hybrid AI will become the default enterprise model. Organizations will use public cloud for flexibility, on-premises infrastructure for control and cost management, and edge environments where latency, data locality, or operational requirements demand it.

The unresolved challenge is consistency. Most enterprises already operate across multiple vendors, clouds, automation tools, and infrastructure types. AI adds GPU clusters, model-serving environments, fine-tuning pipelines, and agentic workloads to that mix.

Over the next five years, the leading organizations will be the ones that create one operational standard across those environments. They will need consistent provisioning, policy enforcement, cost attribution, auditability, and lifecycle management, regardless of where the workload runs.

DevOps transformed software delivery over the past decade. Do you think AI infrastructure requires an entirely new operational model, or is it the natural evolution of DevOps and platform engineering?

AI infrastructure extends DevOps, but it also introduces requirements that traditional DevOps practices were not designed to handle.

DevOps improved software delivery by standardizing pipelines, automating repeatable work, and giving teams faster access to the resources they needed. AI infrastructure requires the same discipline, but the workloads are more resource-intensive, less predictable, and more dependent on specialized infrastructure.

Platform engineering becomes central in this environment. Teams need governed self-service access to validated environments, rather than ticket-based provisioning or one-off manual configurations. The operating model still values speed and automation, but it also requires stronger governance, cost control, and infrastructure context from the beginning.

Zero-touch operations have long been a goal in infrastructure management. How close are we to truly autonomous infrastructure, and what are the biggest technical and organizational barriers that still stand in the way?

We are much closer to autonomous infrastructure than most people realize, but the limiting factor is governance rather than automation. AI agents can already provision environments, deploy workloads, investigate anomalies, and recommend remediation. The challenge is ensuring those actions happen within clearly defined policy boundaries.

Organizations also need confidence that autonomous systems can explain what they did, why they did it, and under which policy they acted. That level of traceability is still missing in many enterprise environments. Autonomous infrastructure will become mainstream when governance, auditability, and policy enforcement are built into the operational layer instead of being added afterward.

As AI workloads become more dynamic and resource-intensive, cost overruns are becoming a major concern. What are the most common mistakes enterprises make when attempting to scale AI infrastructure, and how can they avoid them?

Many organizations focus on acquiring more compute before understanding how efficiently they are using the infrastructure they already have.

We regularly see environments that remain active after projects finish, GPU resources reserved for workloads that never run, and little visibility into which teams or applications are driving infrastructure costs. AI workloads amplify those inefficiencies because demand changes rapidly and GPU infrastructure is expensive.

The solution starts with governance. Every environment should have a defined owner, business purpose, lifecycle, and cost profile from the moment it is provisioned. When organizations combine policy-driven provisioning with automated teardown and continuous visibility into utilization, infrastructure spending becomes much easier to predict and justify.

Looking ahead, what will separate the organizations that successfully operationalize AI at scale from those that struggle, and what infrastructure trends should technology leaders be paying attention to right now that are still flying under the radar?

The organizations that succeed will treat AI infrastructure as an operational capability rather than a collection of technologies. Models will continue to improve, and hardware will continue to become more powerful. Operating those environments consistently across multiple teams and platforms will remain the harder challenge.

One trend that deserves more attention is the shift toward infrastructure intelligence. Infrastructure platforms are beginning to understand the context of what they manage, who owns an environment, what workload it supports, how much it costs, whether it complies with policy, and when it should be optimized or retired. That operational context will become increasingly valuable as enterprises deploy more AI agents and more heterogeneous infrastructure.

I also expect governance to become a competitive advantage. The organizations that can provision infrastructure quickly while maintaining visibility, policy enforcement, cost control, and auditability will move AI projects from pilot to production much faster than those still relying on fragmented operational models.

Thank you for the great interview, readers who wish to learn more should visit Quali.