OpenAI is embedding energy planning directly within its data center development arm, signaling a strategic shift toward treating power technologies and flexible computing as a single, integrated infrastructure function. The reorganization reflects the company's broader effort to build repeatable pathways for incorporating clean power, grid services, and compute scheduling across its rapidly expanding portfolio of AI data centers.
The company is now hiring for a Clean Energy and New Technology Lead position within its Industrial Compute organization. The role will evaluate clean power generation, advanced energy storage, grid flexibility mechanisms, low-carbon backup power, and efficiency technologies. According to the job description, the hire will own the emerging energy strategy and its execution, assessing options against reliability, cost, carbon impact, schedule, and resilience, while also identifying pilot projects and scalable deployment approaches.
OpenAI's emphasis on energy flexibility has public precedents. At Project Camellia, a planned 3.2 GW data center campus in Effingham County, Georgia, the company has committed to providing Georgia Power with up to 1,000 MW of flexible demand response under a 25-year agreement. OpenAI has said the facility will be designed to proactively reduce power consumption before residential customers are affected during periods of high demand. The 1 GW commitment illustrates the scale of flexibility this new role may evaluate, though OpenAI has not confirmed whether the position is directly tied to Project Camellia or any specific project.
Industry observers say the placement of the role inside Industrial Compute is a telling organizational decision. Neil Osnato, founder of Persistence Analytics Group, noted that the key shift is structural: the role sits within data center development and is accountable for execution. He said the responsibilities suggest OpenAI is building an internal capability to determine which power architectures can actually be deployed across its data center portfolio, rather than treating energy as a project-by-project procurement decision.
The hiring comes amid broader changes to OpenAI's data center operations. Chris Malone, who joined OpenAI in March 2025 as head of data centers, has left the company, the Wall Street Journal reported. OpenAI has said it reorganized its infrastructure organization to support the scale and pace of its work, distributing responsibilities among several executives, including Uday Ruddarraju, who leads the data center team; Brent Mayo, who leads data center build and delivery; and Spas Lazarov, who leads data center engineering, according to TechCrunch. There is no indication that the Clean Energy and New Technology Lead role is a replacement for Malone.
OpenAI continues to expand its Industrial Compute organization with roles spanning powered land, electrical infrastructure, data center development, and other elements of physical infrastructure. The company's focus on flexible load could have significant implications for grid capacity. Chris Dunlap, a University of Chicago researcher whose preprint on AI data center flexibility is undergoing peer review, estimates that large AI data centers could provide flexibility equivalent to roughly 25% to 40% of nameplate capacity, depending on the facility. Essentially, none of that potential currently counts toward resource adequacy, Dunlap said.
His analysis considers several mechanisms, including moving workloads between regions, delaying workloads, and slowing or modulating computing activity. For a modeled 500 MW inference-dominant facility, his baseline yielded a mean commitment depth of about 39.8%. Dunlap emphasizes that resource adequacy – whether enough capacity is available at peak – is primarily a peak-capacity problem, not an energy-consumption problem. In his modeling, a 10 GW fleet operating at the flexibility levels described could provide capacity in the low thousands of megawatts, he said. That would not eliminate a regional capacity shortfall, but the potential capacity would not require a new generation project, an interconnection queue position, or a construction timeline. Dunlap's estimate remains provisional because the manuscript is still under peer review.
Coordination remains the central challenge, Dunlap said. The issue is not whether data centers can change their electricity use but how multiple sites coordinate that behavior. Data centers have already demonstrated the ability to shed large amounts of load quickly. But if multiple facilities independently respond to a price signal or emergency notice, the aggregate result could create a new problem: load disappears in one region while unscheduled load appears in another. Dunlap likened moving compute between regions to an interregional transfer of demand without the scheduling and transmission mechanisms that govern physical electricity transfers. "Power cannot flow from Illinois to Texas in meaningful quantity," he said. "But data center load can move there, via fiber rather than transmission lines." That creates a need for an entity above individual data center operators – such as an RTO, aggregator, or third party – to coordinate transfers.
The problem becomes more acute as participation grows. Dunlap said 4 GW of nominal commitments would not necessarily amount to 4 GW of dependable grid capacity if shifted workloads land in the same constrained places. Location also plays a critical role in determining the value of flexible AI facilities to the grid. Abhijit Das, whose research examines flexible computing and power-system constraints, noted that renewable curtailment tends to occur where wind and solar resources are concentrated, while data centers have historically followed other siting priorities. A flexible training facility sited where surplus generation is regularly available could absorb some of that energy by shifting compute to periods of abundant supply.
Das described the mechanism as functioning like transmission without new wires: shifting energy through time by moving compute, rather than moving electricity through space. But it does not eliminate transmission constraints. "Virtual transmission only works if the flexible load is physically in the curtailment zone," Das said. "A cluster in Virginia isn't absorbing curtailed Nebraska wind. You still need the wires for that." Das also noted that access to detailed grid information needed to identify viable sites remains an unresolved issue. While curtailment data is generally public, determining whether a particular substation or transmission bus can host a large new load often requires detailed power-flow analysis and restricted planning information.
For Project Camellia, the 1 GW flexible-load commitment is public, but key terms – how long a reduction could be sustained, what would trigger it, and how the response would be dispatched and measured – have not been disclosed. Those details will determine how much value the commitment provides as a grid resource. Dunlap said utilities would need an enforceable load ceiling, performance testing, and penalties for nonperformance, with duration included in accreditation. He suggested a Firm Service Level approach – committing to stay below a specified load during a reliability event – may fit data centers better than baseline methods. Commitments that rely on shifting workloads between facilities would also need to account for overlapping capacity at their destination sites.
OpenAI has acknowledged receipt of questions related to this story but has not provided comments as of publication. Data Center Knowledge will update this article with any responses.