In Texas, ERCOT, the regional grid operator, records a stark contrast in its figures: large-load interconnection requests have queued up to approximately 474 GW, while only about 9 GW have been approved for energization, and only about 4 GW have actually begun drawing power. Roughly 90% of these requests come from data centers. This approximately 120-fold disparity represents the most expensive wait in the artificial intelligence industry today.
Chips continue to be delivered on schedule, and data center facilities can be built within one to two years. Yet as long as the high-voltage transmission lines connecting them to the grid cannot supply electricity on time, every chip in the server racks remains unpowered. Securing a firm energization date has become the hard bottleneck that dictates when computing capacity can actually go live.
On September 17, 2026, AWS announced in Houston a software product targeting this very bottleneck: Agentic Grid Planning on AWS (press release). The cloud giant, which sells compute to developers worldwide, introduced automated agents into the complex engineering study workflows inside utility companies. Rather than directly controlling physical grid equipment, this software aims to accelerate the turnaround efficiency of preliminary grid evaluations.
The push for turnaround efficiency stems from a severe mismatch in pacing between grid expansion and data center development. A modern data center facility can be constructed in just 18 to 24 months. On the grid side, however, manufacturing a large power transformer currently requires an average lead time of about 128 weeks, with generator step-up transformers taking as long as 144 weeks. Building new high-voltage transmission lines takes even longer: in data-center-dense regions like Northern Virginia, building a new backbone transmission line takes more than five years; nationwide, transmission corridors typically require 7 to 15 years, and in 2024, only 322 miles of new high-voltage transmission lines were actually completed across the United States (When Data Centers Turned From Darlings to Hot Potatoes).
While equipment manufacturing and transmission construction crawl forward, the queue of applications continues to lengthen. According to statistics from Lawrence Berkeley National Laboratory (LBNL) in its Queued Up 2025 report, by the end of 2024, project requests waiting in interconnection queues across the United States had reached 2,290 GW. The median time from initial request submission to commercial operation was under two years in earlier years, but has stretched to more than four years for recently built projects; projects coming online in 2024 experienced a typical timeline of 55 months.
Confronting heavily clogged queues, regulators have begun tightening schedule requirements. FERC issued Order No. 2023, directing grid operators to transition away from the inefficient first-come, first-served model to a “first-ready, first-served” cluster study approach, and establishing a statutory cluster study deadline of 150 calendar days. Planning engineers at utilities must now execute technical computations of unprecedented scale within this tight regulatory window.
Whether feeding power from a new generation plant or requesting power for a massive data center, connecting to the high-voltage grid requires rigorous technical calculations. Generation and storage projects inject power into the grid, while data centers draw power from it; though their power flows run in opposite directions, the underlying laws of physics governing grid stability evaluations are identical. Engineers call this gatekeeping procedure an interconnection study.
Interconnection studies directly determine whether a project receives an approval for energization. At this threshold, data center developers wait for power supply agreements, utility engineers race against regulatory deadlines, and hundreds of millions of dollars in infrastructure investments hinge on the delivery of this technical report. What this study actually calculates determines where subsequent technological optimization can genuinely deliver an impact.
The core work of an interconnection study begins with engineers building mathematical models of the power grid inside specialized simulation software. Engineers open the simulation program and input the substation location and power ratings of the proposed project. Based on the physical laws of electrical circuits, the software calculates whether voltages across the entire grid remain within nominal limits and how much load each surrounding high-voltage transmission line carries. The power industry calls this assessment a power flow study. If power flow on any line exceeds its permissible rating, or if voltage at a substation bus dips or spikes outside its allowable band, the software triggers an alert. Power engineers refer to these abnormal operational conditions collectively as violations.
Maintaining everyday equilibrium is not enough; the grid must also withstand unexpected disruptions. Engineers simulate contingencies in the software, such as a lightning strike severing a critical transmission corridor or a baseload generator tripping offline. The software recalculates the grid state following load redistribution, evaluating whether remaining lines can handle the surge in current without triggering cascading blackouts. Engineers call this stress test contingency analysis. In practical engineering, a faster DC power flow model is typically used for broad screening, followed by a more rigorous AC power flow model to thoroughly re-examine high-risk scenarios.
Once simulations are complete, the attribution of responsibility for every grid weakness must be clearly demarcated. The engineering team first models the baseline grid without the project, then runs the model with the project integrated. Comparing the two sets of simulation data reveals precisely which violations are new issues directly caused by the project. Grid upgrade costs follow the industry’s “cost-causation” principle: every newly identified violation directly determines which transmission facilities the applicant must fund to expand or upgrade.
The software service launched by AWS targets precisely this complex engineering workflow woven together by multiple algorithms. On its product page, AWS explicitly scopes the service to steady-state transmission studies for generation and large-load interconnections.
The entire agent workflow consists of three steps.
The first step is study drafting. The agent automatically drafts proposed points of interconnection and network upgrade options based on existing grid topology. Power engineers review each option item by item, retaining full authority to approve or manually modify them. Only after an engineer confirms the proposal does the agent use the validated parameters to automatically generate batch simulation cases across multiple planning scenarios.
The second step is simulation execution and report drafting. The agent invokes specialized simulation software to run power flow and contingency analyses, automatically categorizes the resulting violations, proposes corresponding mitigation measures for engineering approval, and drafts the initial study report. If external planning parameters change, the system can rerun the batch workflow directly, eliminating the redundant manual work of rebuilding cases. According to an AWS team member in a public professional exchange (post), the agent assists in identifying the root causes of violations and drafting remediation proposals, while qualified engineers strictly control every decision gate.
The third step is workflow codification. Once engineers complete an initial interconnection study through interactive guidance, the agent distills the logged steps to automatically draft a standardized workflow guide. After the engineering team reviews and publishes this guide, it serves as a reusable template across the organization for subsequent, similar studies.
Only when these three steps are complete is an interconnection study finished. The press release encapsulates this division of labor in a single sentence: engineers guide the study and make final engineering decisions, physics simulation software handles the underlying modeling and numerical analysis, and AI agents orchestrate the study workflow in between, seamlessly stitching the pieces together (press release).
Throughout the process, the agent maintains version control over all artifacts, creating a comprehensive audit trail of every operational step. Reviewers can clearly differentiate between simulation evidence, AI reasoning processes, and human engineering decisions, with every output figure traceable to its specific execution batch. This strict separation of simulation data, AI reasoning, and engineering decisions serves as an indispensable safeguard in a highly regulated grid planning environment. The conclusions of interconnection studies involve the allocation of massive upgrade costs and frequently face formal disputes or even litigation from applicants. Years later, when undergoing regulatory audits or judicial scrutiny, grid operators must be capable of faithfully reproducing the technical evidence and chain of reasoning behind their decisions.
This versioned traceability is purely a software engineering capability. The product itself carries no formal certification from regulatory authorities, nor does it confer any statutory liability exemption.
The software suite deploys inside the customer’s dedicated cloud account, with access control governed by the customer’s own identity systems. Model inference calls are powered by Amazon Bedrock, and transmitted prompts and generated content are not used for model training or shared with model providers. In the power industry, using Python scripts to automate batch runs in legacy tools like PSS/E has long been an industry standard; Duke Energy also explicitly requires interconnection applicants to submit PSS/E models in its public procurement filings. Script automation is nothing new; the real shift here is introducing agents to assist in analyzing violations, drafting mitigation schemes, and stringing together previously disjointed stages into a continuous pipeline.
In the announcement of this new product, the most notable benchmark came from Duke Energy. According to AWS, Duke Energy reduced the data preparation phase from two weeks of manual work to hours in pilot testing (press release). Industry trade publication TDWorld, in reporting on the development, was similarly cautious in prefacing the claim with an “according to AWS” note.
Placing this reduction from two weeks to hours back into the context of real-world grid engineering reveals sharp boundaries. The first boundary is scope: the speedup applies solely to the data preparation phase—cleaning raw grid data and formatting it into input files readable by simulation software. AWS disclosed nothing about whether this refers to calendar duration or labor hours, the size of the grid modeled, error and rework rates, or overall costs. The second boundary is potential confusion with earlier news: in early 2024, reports highlighted that Duke Energy leveraged AWS cloud computing to cut power flow calculation times from six weeks to hours (Latitude Media). However, that earlier partnership belonged to AWS’s simulation infrastructure line, aimed at accelerating solver compute speed; the 2026 agent product, by contrast, accelerates preliminary data preparation. The two initiatives address distinct phases of interconnection studies and represent entirely separate product offerings.
The true timeline lies outside these boundaries. A yawning chasm still separates time savings in an isolated stage from the multi-year wait across the entire pipeline. In data compiled by Lawrence Berkeley National Laboratory, the median time a project spends in the interconnection queue exceeds four years, with typical projects taking as long as 55 months. Within this protracted timeline, data preparation for interconnection studies is merely a sliver; a local acceleration cannot automatically translate into a breakthrough for the overall timeline.
Returning to the large-load application reality logged by Texas grid operators, energization dates have undeniably throttled the expansion pace of the artificial intelligence industry. Chip procurement and data center construction can advance on monthly schedules, but waiting for transmission network approvals often takes years. This is the first hard constraint the tech industry has collided with as it expands into physical reality.
Throughout this long and convoluted energization approval pipeline, the interconnection study happens to be a predominantly software-driven, clearly computable segment. MISO previously used an automated analytics platform to replicate in just 10 days a large cluster study that had previously taken nearly two years of manual labor, yielding largely consistent conclusions. This demonstrates that at this software-mediated technical gateway, there is substantial engineering headroom to compress computational and workflow orchestration cycles.
Yet the critical causal transmission chain—from accelerating study phases to shortening overall queue durations—remains unproven. To date, the industry possesses no publicly available causal empirical data showing whether speeding up an individual study or data preparation step smoothly translates into reduced queue times and earlier power delivery for data centers. Between completing a draft study report and getting utilities, power generators, and data center operators to agree on who pays for network upgrades—and onward through regulatory sign-offs and scheduling transmission construction—the process remains bottlenecked by multiple rounds of human review and stakeholder negotiations.
Even without empirical proof of full technological transmission, commercial entities have already deployed massive capital on both sides of the power equation. AWS plays a dual role: it sells large-scale computing infrastructure to developers on one hand, while selling agent software to utilities bottlenecked by interconnection queues on the other. This positioning echoes capital flows from other computing giants: Nvidia plans to invest $2 billion in power infrastructure developer Lancium for a 20% stake, retaining the right to invest up to an additional $1 billion (After GPUs, AI Companies Are Scrambling for Energization Dates); Amazon has built on-site gas-fired generation in West Texas to directly power its data centers. Driven by massive capital allocation and clear commercial imperatives, this ecosystem has not yet formed a self-improving closed loop at the level of software code.
To draw a historical analogy: when the Newcomen atmospheric steam engine was invented in the early eighteenth century, its premier practical deployment was at the pitheads of coal mines, using steam power to pump water from deep underground and enabling miners to extract more coal. The machine lowered the extraction cost of the very fuel required to power itself. In economic history, this self-reinforcing feedback loop is known as increasing returns. Today, tech giants attempting to leverage artificial intelligence tools to accelerate grid interconnection studies stand at the very beginning of a structurally similar dynamic in business logic.
In exploring the evolution of intelligent systems, a key criterion is that the exam hall comes before the test-taker (analysis article on self-improving AI). The reason many agents struggle to achieve reliable capability gains in open domains is the absence of objective, rigorous evaluation benchmarks. When writing an essay or drafting a marketing proposal, where scoring depends entirely on self-evaluation by the model or human subjective preference, the system easily falls into a self-referential, subjective loop.
In grid interconnection studies, however, the underlying specialized physical simulation software serves as an objective physics exam. Any grid topology modification or mitigation scheme proposed by an agent that violates the laws of electromagnetism is immediately exposed during simulation runs—manifesting as power flow divergence, equipment voltage violations, or transmission line thermal overloads. This makes interconnection studies one of the cells with the strongest verification signal as agents enter the physical world. Much like OpenAI’s collaboration with MIT to explore a night shift operator for quantum experiments, automated systems consistently find their firmest initial footing in technical domains where acceptance criteria are crisp and objective.
Scoring high on an objective exam does not mean the interconnection queue is suddenly unlocked. Between accelerating a single study or data prep workflow and actually clearing the overall grid queue stand multiple real-world walls that software cannot breach. Beyond the currently missing end-to-end causal measurements, each hurdle stands firmly outside the software world: protracted administrative reviews by public utility commissions, the physical manufacturing of large transformers with lead times of two to three years, the multi-year construction of field transmission corridors, and contentious negotiations with local communities and environmental organizations along prospective routes.
AWS’s product rollout offers a textbook example of agents orchestrating specialized tools without touching underlying numerical computation, targeted directly at the energy feedpipe constraining AI expansion. Whether this initiative remains merely an isolated software workflow optimization or evolves into genuine industrial-scale self-improvement hinges on whether empirical verification can ultimately confirm transmission through that third link. The next time Duke Energy or other utilities publish fresh metrics, there is no need to look at how many multiples an internal technical step has accelerated; simply observe whether the duration of the interconnection queue has experienced a tangible reduction.