Artificial intelligence has created an electricity challenge unlike anything the U.S. power industry has faced. Hyperscale developers are proposing AI campuses that need hundreds of megawatts — and increasingly multiple gigawatts — while utilities confront transmission constraints, generation shortages, multi-year interconnection queues, and unprecedented requests for new load. Texas has become ground zero for this collision between the AI economy and the electric grid, and a massive project being developed for Amazon in West Texas may be a preview of where the whole industry is headed.
Amazon is planning an AI data center campus at the GW Ranch development in Pecos County that would initially operate largely outside the Texas grid. The project would buy electricity from a dedicated plant being developed by Pacifico Energy — permitted for 35 natural-gas turbines with a combined nameplate capacity of up to 7.65 gigawatts, one of the largest gas plants ever proposed in the United States. (The permit describes roughly 5 GW of nominal delivered output, with Pacifico expecting a first 1 GW phase to come online in the first half of 2027.) To put 7.65 GW in perspective: ERCOT's historical system peak was about 85.5 GW as of its April 2026 forecast — so the capacity contemplated for this single development equals nearly 9% of the entire state's historical peak demand.
The strategy targets the single biggest obstacle facing AI developers: waiting for the grid. Amazon has said the generation would initially run without drawing from the Texas grid, transitioning toward grid-connected service as interconnection becomes available. The logic is straightforward — rather than dropping an enormous new load onto an already strained utility system, you build dedicated generation specifically to serve the data center. Amazon has also said it's evaluating solar and battery storage for the site (Pacifico has outlined up to 750 MW of on-site solar and 1.8 GW of storage), and that the plant will use non-potable, brackish groundwater rather than drinking or irrigation water — pointing toward a possible hybrid system combining dispatchable generation with renewables.
From an infrastructure-development standpoint, there's a real argument for this model. If an AI developer needs several gigawatts, building dedicated generation instead of immediately loading the grid can reduce pressure on utility infrastructure while dramatically accelerating time to power. For an industry where billions in computing equipment may sit idle waiting for electricity, bypassing a multi-year interconnection process has enormous economic value. It also shifts at least some responsibility for new generation onto the company creating the demand, rather than automatically onto utilities and their existing customers — a point Amazon leans on directly, framing the approach as "powered by new on-site generation that won't raise electricity costs for Texas families."
But this project also exposes the other side. The gas generation at GW Ranch has been permitted for up to roughly 33 million metric tons of carbon dioxide per year. That figure is a permitted maximum, not a prediction of actual emissions — an important distinction, and plants rarely run at their ceiling. Even so, the scale has generated real controversy, because a facility operating anywhere near that level would be an enormous new source of carbon emissions created largely to support digital infrastructure. For context, the dirtiest power plant currently operating in the U.S. emits around 16 million tons — this project is permitted for roughly double that.
Which raises an uncomfortable question for both industries: if AI companies solve the grid-capacity problem by building enormous fossil-fueled plants behind the meter, are we actually solving the energy problem — or just moving it somewhere else?
Why behind-the-meter power is becoming so attractive
The Amazon project highlights a shift already underway: access to power is becoming more important than almost every other site-selection factor. Developers used to weigh land cost, fiber, taxes, workforce, permitting, water, and proximity to population. Those still matter — but for large AI campuses, available megawatts increasingly determine whether a site is viable at all. A developer can lock up cheap land and world-class fiber, but if the local utility can't deliver 500 MW or 1 GW in the required timeframe, the project may never move.
That's why behind-the-meter generation has become central to the conversation. Instead of relying entirely on a utility to build new generation and transmission, a developer builds dedicated power at or near the data center and delivers it straight to the campus. Natural gas is especially attractive here: it's dispatchable and available around the clock, can often be built faster than major transmission, and doesn't depend on weather. In regions with existing pipelines and abundant gas — like West Texas — the economics get even more compelling.
The biggest advantage may simply be speed to power. A hyperscale campus is billions of dollars of servers, GPUs, cooling, networking, and buildings. Every year that facility waits for interconnection is potentially enormous lost value. If dedicated generation lets it start operating years earlier, developers may happily accept higher electricity costs or extra infrastructure investment, because bringing compute online sooner outweighs those costs. That's fundamentally changing the economics of energy procurement for the tech industry — and it's why Amazon is following a path pioneered at scale by xAI's gas-turbine buildout in Memphis, and joined this year by Microsoft, Google, and Meta.
Behind-the-meter generation can also ease some immediate burden on existing utility customers. One of the most contentious questions of the AI boom is who pays for the transmission, substations, generation, and grid upgrades needed to serve enormous new loads. If utilities spend billions on infrastructure for data centers and those costs enter the rate base, residential and commercial customers may reasonably ask whether they're subsidizing infrastructure built primarily for some of the world's largest tech companies. Requiring — or incentivizing — large developers to provide more of their own generation could help address that, though the details depend heavily on market structure, regulation, and how each project ultimately interacts with the grid.
But "bring your own power" creates a new set of problems
The catch is that solving the grid problem with dedicated generation doesn't eliminate the broader energy impacts — it changes where they occur and who manages them. A multi-gigawatt gas facility still needs pipelines, turbines, substations, transmission equipment, fuel supply, air permits, water, and major capital. It still produces emissions. And when that generation exists primarily to power AI, those emissions become part of a much larger debate over AI's environmental footprint.
The harder question isn't whether gas should power a data center — gas already plays a major role in the U.S. system and is especially important in Texas. It's whether the extraordinary growth of AI could drive construction of an entirely new generation fleet that otherwise wouldn't have been built. If dozens of hyperscale campuses adopt the same strategy, the cumulative effect could be substantial — potentially adding tens of gigawatts of dedicated fossil generation, even as the same technology companies pursue aggressive carbon-reduction commitments. (Amazon, for its part, reported nearly 81 million tons of CO2-equivalent in 2025 while maintaining a net-zero-by-2040 pledge and pointing to roughly 10 GW of carbon-free projects it has backed in Texas.)
There's also a genuine reliability question. A behind-the-meter plant reduces a data center's immediate dependence on the grid, but hyperscale facilities still need extraordinary redundancy. AI workloads can't just stop when a turbine trips or a pipeline is disrupted. So developers need multiple layers of generation, backup, storage, redundant electrical systems, and often eventual grid connectivity. What looks at first like a simple fix — build a plant next to the data center — quickly becomes a sophisticated microgrid requiring the same reliability engineering as traditional utility infrastructure.
Which is why the future is unlikely to be defined by gas alone. The more realistic model is a hybrid energy campus: gas providing firm generation, solar and wind supplying lower-carbon power when available, Battery Energy Storage Systems delivering fast-response capacity and short-duration backup, and the grid eventually adding another reliability layer. Depending on location and timeline, future campuses might also fold in fuel cells, geothermal, long-duration storage, hydrogen, or advanced nuclear. The goal isn't necessarily complete energy independence — it's enough local generation and flexibility that the data center doesn't put its entire load on the grid from day one.
The bigger question: who should pay for the AI power buildout?
This may become the most politically important question around AI data centers. The U.S. clearly has an economic incentive to build AI infrastructure — AI is increasingly viewed as strategic to growth, competitiveness, research, manufacturing, defense, and technological leadership, and supporting it will require enormous investment in generation and transmission.
But supporting AI doesn't automatically mean existing customers should absorb the financial risk. If a developer requests 1 GW of new capacity, the utility may need new substations, transmission, and generation to serve it — investments that stay in service for decades, while the tech industry changes far faster. Regulators have to weigh what happens if projected AI demand doesn't materialize, a campus is never fully built, or computing becomes dramatically more efficient before utilities recover their investment.
So developers should expect rising scrutiny over how much of the infrastructure serving their facilities they finance themselves. That doesn't mean every campus should build its own power plant. It does mean cost allocation, minimum-load commitments, long-term contracts, collateral requirements, and infrastructure contributions are likely to become central to large-load utility agreements. The objective is straightforward: enable economic development while protecting existing customers from disproportionate financial risk.
There's a flip side worth remembering, too. Large data centers can be extremely valuable utility customers — substantial, predictable demand that can help support investment in new generation and transmission. Properly structured, these projects could contribute significant revenue toward grid modernization rather than simply adding cost. The challenge for regulators is designing tariffs and commercial structures that capture those benefits while appropriately allocating the costs the new loads create.
My perspective: the answer isn't "no data centers" or "build at any cost"
The debate is polarizing fast. On one side, technology companies argue the U.S. must rapidly expand computing infrastructure to stay competitive in AI. On the other, communities and environmental advocates worry the rush could raise electricity prices, accelerate fossil-fuel development, consume water, and strain already-constrained grids. Both sides raise legitimate points — but framing it as progress versus the environment oversimplifies a much more complicated energy challenge.
AI data centers are going to be built somewhere; global demand is growing too fast for the industry to simply stop. And the U.S. has strong economic and competitiveness reasons to attract that investment. The more important questions are how those facilities should be powered and who should bear the infrastructure cost — because those decisions determine whether the AI expansion strengthens the grid or creates new financial and reliability problems.
In my view, one principle should be clear: the companies creating extraordinary new demand should assume a meaningful share of the responsibility for the infrastructure to serve it. If a company wants a multi-gigawatt campus, existing residential and commercial customers shouldn't automatically finance billions in dedicated infrastructure through higher rates. Utilities, regulators, and developers need structures that allocate those costs fairly while still encouraging investment.
Where it gets harder is what kind of generation should provide the power. Gas is attractive because it's dispatchable, proven, scalable, and capable of 24/7 operation — and in markets like Texas, abundant supply and existing pipelines make it especially practical. Ignoring those advantages because gas produces emissions doesn't make the reliability problem vanish. But building enormous new fossil fleets exclusively to serve AI could lock in decades of additional emissions precisely as the industry invests heavily in lower-carbon generation.
The better path is likely between the extremes. Dedicated gas may be necessary for certain projects, particularly where utilities can't provide capacity in reasonable timeframes — but it should increasingly be evaluated as part of a hybrid architecture rather than the whole solution. Solar, wind, batteries, demand flexibility, fuel cells, geothermal, long-duration storage, and eventually advanced nuclear can all contribute depending on location, economics, and schedule. The goal: minimize both grid impact and emissions while maintaining the reliability AI demands.
Battery storage deserves particular attention, because it can fundamentally change how large data centers interact with the grid. A well-designed BESS can cut peak demand, provide ride-through capability, support backup generation, improve power quality, absorb excess renewable output, and let a campus operate more flexibly during grid stress. Batteries can't yet economically provide days of continuous gigawatt-scale power, so they don't eliminate the need for firm generation — but pairing them with dispatchable generation reduces generator cycling, improves efficiency, and creates a far more flexible system than gas turbines alone.
Data centers themselves may also need to become more flexible consumers. Hyperscale facilities have traditionally been treated as near-constant loads needing uninterrupted power. AI workloads could challenge that: some computing tasks might eventually be shifted geographically or scheduled for when electricity is abundant, letting operators reduce consumption during emergencies without touching mission-critical work. Even small amounts of flexibility across gigawatts of AI demand could be enormously valuable to grid operators.
Texas may be showing us what comes next
Texas matters because it combines nearly every characteristic likely to define the next phase of AI infrastructure: enormous renewable resources, abundant gas, cheap land, a competitive electricity market, major transmission, and a historically development-friendly regulatory environment. Those advantages have made it a magnet for data centers — but the sheer volume of proposed projects is testing the system's limits. ERCOT's preliminary large-load forecasts show just how dramatically these requests could reshape demand, with data centers representing a substantial share of proposed new load — so large that traditional utility-planning assumptions are being challenged. Instead of forecasting gradual annual increases, planners now evaluate individual projects that can add hundreds or thousands of megawatts within short development windows.
The state's experience could shape data center policy nationwide. Utilities and regulators in Virginia, Georgia, Arizona, Ohio, Pennsylvania, the Carolinas, and other fast-growing markets face the same questions about generation, transmission, interconnection timelines, water, and cost allocation. If Texas develops workable frameworks for integrating extremely large loads while protecting existing customers and maintaining reliability, other markets will study them closely.
One real possibility: the traditional line between electricity consumer and producer begins to blur. The AI campus of the future may resemble a small utility system of its own — dedicated generation, renewables, batteries, substations, sophisticated energy management, and eventually multiple grid connections. Instead of simply requesting 1 GW from the local utility, developers may arrive with a comprehensive energy plan: how much they'll generate themselves, how much they need from the grid, how they'll respond during emergencies, and how costs will be allocated. A 50 MW facility can behave like a very large commercial customer; a multi-gigawatt campus begins to resemble an industrial energy system, where electricity strategy becomes one of the project's fundamental design considerations.
The real opportunity: make AI an asset to the grid
The most interesting possibility is that AI data centers could become more than a burden on the system. With the right generation, storage, controls, and commercial agreements, they could become valuable grid participants. Large batteries at AI campuses could provide ancillary services; dedicated generation could potentially support the surrounding system during emergencies where market rules permit; and flexible computing loads could cut consumption during extreme stress.
Picture a large AI campus with dedicated gas generation, several hundred megawatts of solar, a utility-scale BESS, and a grid connection. Under normal conditions it optimizes among those resources based on prices, renewable availability, operational needs, and emissions. During grid stress, it reduces utility draw, discharges storage, or leans on its own generation — and under the right agreements, its excess generation or storage might even support the surrounding system. That's a very different model from simply connecting a massive load and expecting the utility to supply every megawatt around the clock.
Getting there requires cooperation among technology companies, utilities, regulators, developers, manufacturers, and communities — plus more sophisticated tariffs and interconnection agreements that recognize the unique characteristics of enormous AI loads. The traditional utility model was never designed for customers requesting several gigawatts at a single development, and forcing these projects into conventional commercial-rate structures may no longer make sense.
The Amazon project may therefore be an early version of something we'll see far more often. Behind-the-meter generation won't fit everywhere, and heavy reliance on gas raises legitimate environmental concerns. But the broader concept — that extremely large electricity consumers should participate directly in solving their own power requirements — could become a defining principle of the AI infrastructure era. The challenge is making sure that solving the AI industry's power problem doesn't create a larger problem for the grid, for electricity customers, or for the communities hosting these projects.
The bottom line: AI needs more power — but it also needs a better energy strategy
The controversy around Amazon's Texas project is a preview of a much larger debate that will unfold across the country over the next decade. AI is creating enormous economic opportunity — and an equally enormous appetite for electricity. Utilities accustomed to gradual demand growth are now asked to serve individual campuses needing hundreds of megawatts or several gigawatts of continuous power. The grid was never designed to absorb this level of load growth at the speed the tech industry wants to build.
Behind-the-meter generation offers one answer. If technology companies finance dedicated generation, substations, storage, and the rest of the infrastructure their facilities require, they can bring compute online faster while easing the immediate burden on the grid. From that angle, Amazon's approach raises an idea worth taking seriously: companies creating unprecedented demand should participate directly in developing the energy infrastructure to support it.
The controversy is about how that power gets produced. Building gigawatts of gas may solve the reliability and interconnection problem, but it raises real questions about emissions and whether the AI boom triggers a new wave of fossil infrastructure — concerns that shouldn't be dismissed as mere opposition to progress. At the same time, expecting renewables and batteries alone to reliably support every multi-gigawatt campus today ignores the technical and economic realities of running facilities that need power 24/7.
The better solution is almost certainly a diversified strategy: gas where firm generation is genuinely necessary, solar and wind reducing fuel use and emissions when available, batteries providing fast response and peak reduction, and — over time — geothermal, fuel cells, hydrogen, long-duration storage, and advanced nuclear adding options. The specific mix will vary by location, but the objective stays constant: reliable electricity without unnecessarily transferring the financial or environmental costs of AI onto surrounding communities.
Utilities and regulators have a crucial role too. Large-load tariffs, minimum-demand commitments, infrastructure contributions, and demand-response programs can ensure developers pay an appropriate share of the costs they create — while regulators should also reward campuses that provide grid benefits through flexible demand, storage, or dedicated generation. The goal isn't to prevent AI development; it's to structure it so it strengthens rather than weakens the electricity system.
The most important question isn't whether Amazon should build a gas plant in Texas. It's what energy model we establish as hundreds of billions of dollars flow into AI infrastructure nationwide. If every campus simply connects to the grid and expects utilities to build whatever's needed, existing customers face real financial and reliability risk. If every developer instead builds massive standalone fossil plants, we solve one infrastructure challenge while creating another. Neither extreme is the best long-term answer.
The opportunity is to build something better — AI campuses as sophisticated energy ecosystems combining generation, storage, grid connectivity, advanced controls, and flexible demand. Rather than passive consumers of enormous power, these facilities could increasingly participate in the energy system itself. Done right, the tremendous investment flowing into AI could help finance the generation, storage, substations, and transmission that ultimately strengthen the broader grid.
Amazon's Texas project deserves attention for reasons far beyond Amazon. It's an early example of what happens when the speed of technological development collides with the much slower process of building energy infrastructure — and similar decisions will soon confront technology companies, utilities, regulators, developers, and communities across the country. AI unquestionably needs more power. The challenge is making sure we build that power responsibly, allocate its costs fairly, and use the moment to modernize the grid along the way.
The question isn't whether we should power the AI revolution. It's whether we can use the AI revolution to build a better power system.
Run the numbers behind projects like this — The scale in this story is something you can size yourself. Our Data Center Load Calculator estimates the electrical demand of a GPU campus, the Utility-Scale BESS Sizing Calculator scopes the storage a hybrid campus would need, and the Utility-Scale Solar Array Calculator sizes the solar layer developers like Pacifico pair with gas generation.
Written by Chris Kalowes, founder of WattThe?! — 15+ years in utility-scale battery energy storage (BESS), renewable energy, and AI infrastructure, across utilities, IPPs, EPCs, developers, and technology providers. Energy Intelligence. Simplified.