Jayashree Arunkumar is a Distinguished Member of Technical Staff, driving enterprise-scale AI adoption as part of the Technology Advisory.
AI demand is pushing data center operators into a new kind of bottleneck. As rack densities rise and critical infrastructure lead times stretch, delivery increasingly depends on how quickly operators can make coordinated decisions across construction, procurement, finance and operations. A data center rack used to draw about 15 kW. GPU racks now draw 100–120 kW, and designs already on the board plan for 1 MW liquid-cooled halls. The hard part is no longer the design. It is getting the equipment. High-voltage transformers have long lead times, so schedules now depend on how carefully operators split owner-furnished equipment (OFE) from contractor-furnished equipment (CFE).
The constraint has moved. It is no longer capacity alone; it is how quickly an operator can make coordinated decisions. Building and running a data center requires construction, equipment readiness, transport, commissioning, operations, asset utilization and finance to move in step, yet most of the delays occur in the gaps between them. This matters most in high-commitment data center programs, where a delay in a single critical component ripples straight into customer commitments, revenue realization and capital efficiency. In my conversations with hyperscalers, one concern surfaces repeatedly: traditional planning methods have not kept pace with the dramatic growth in product variants, demand volatility and compressed technology cycles.
With GPU generations turning in less than a year, the window to reorder is shrinking—planning decisions must be grounded in connected data intelligence. Most data center articles treat the AI buildout as a supply problem: not enough power, chips, land or capital. What they often miss is the decision problem: hyperscalers must commit to build quantities before demand is clear, then coordinate purchasing, planning and execution fast enough to keep pace with the technology cycle. That’s why the following autonomous supply chain workflow resonated with client teams. It addresses a constraint they are already experiencing.
The Infrastructure Challenge: When A Delayed Asset Becomes A Strategic Constraint
Consider an illustrative high-density data center program in the construction phase at Red Oak, Texas, that depends on a set of high-voltage transformers. The facility carries significant capital investment and is tied to customer capacity commitments. If supply chain volatility delays those transformers, the impact extends beyond construction sequencing. It cascades into commissioning timelines, connected-load readiness, revenue realization, customer confidence and the program’s overall return profile. In a traditional operating model, this becomes an escalation. Resolving it requires manual coordination across construction, procurement, inventory owners, finance, logistics and project governance. Each team checks its own systems, validates availability, assesses cost impact, secures approvals and manually triggers updates across downstream systems. Because every step waits on the one before it, the resolution moves at the speed of the slowest handoff rather than the urgency of the constraint.
The Agentic Shift: From Manual Escalation To Governed Decision Orchestration
An AI-enabled, agentic operating model changes that pattern. Rather than treating the transformer shortage as a linear escalation, multiple agents work in parallel to sense the issue, assess options and recommend a governed resolution. One agent scans global inventory for owner-furnished equipment across other data center programs. A second evaluates project-level priority, timing and customer impact. A third assesses cost, budget, transfer implications and policy constraints. In the Red Oak case, the agents surface a suitable high-voltage transformer sitting in Arizona, where permitting delays mean the local project has no immediate need for it. The system then evaluates whether rerouting that transformer to Red Oak delivers higher business value without creating disproportionate risk to the Arizona program.
The important shift here is not automation alone; it is governed decision intelligence. The hardest decision is whether to redirect a scarce transformer to Red Oak without creating a more damaging delay elsewhere. It requires weighing Red Oak’s customer, revenue and schedule impact against the priorities and risks of the project giving up the asset. Fragmented applications make that enterprise-wide trade-off slow and difficult for human teams to evaluate consistently. The workflow weighs business impact, construction criticality, timing urgency, cost benefit, customer commitment and inter-project budget implications. Because the transformer is a capital asset, the workflow also carries governance. A cloud financial controller receives a notification, reviews the recommendation and approves the capital transfer. Once approved, the system creates the required journal entries so the asset movement is reflected correctly in the ledger and the receiving project can consume the asset without downstream financial or compliance issues.
The deeper point is that AI closes one of the biggest gaps in data center delivery: the fragmentation between physical execution and enterprise systems. Construction schedules, supplier readiness, inventory availability, transport planning, capital budgets, ERP records and compliance approvals typically sit in different applications, owned by different teams. Agents connect those fragments into a single, outcome-driven flow, so the system doesn’t merely report a delay but actively resolves it within approved guardrails. Deterministic guardrails such as cost ceilings, approval thresholds, budget limits and policy rules let an agent move fast without bypassing governance. An agent may be permitted to recommend or initiate actions only within a defined financial threshold, while a human approver handles higher-impact decisions.
What would make me skeptical is an agentic solution that scales globally without enforcing regional data boundaries, localized access controls and clear accountability for high-impact decisions. If agents can access data outside their authorized region or reason from incomplete, outdated or fragmented records, their recommendations cannot be trusted, regardless of how sophisticated the model is.
The Operating Model Becomes The Differentiator
What will separate data center operators is not access to chips, liquid cooling or higher rack density alone, but an operating model that can anticipate constraints, evaluate trade-offs, trigger approvals and optimize capital across the asset lifecycle. AI can become the orchestration fabric connecting engineering, procurement, construction, logistics, finance, operations and governance, thus improving asset utilization and accelerating capacity delivery. But that advantage depends on trusted asset data, clear ownership of capital transfers, enforceable regional boundaries and approval thresholds that define what an agent may decide.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?