Texas Becomes the Laboratory

Texas did not become the center of the data-center power debate by accident. The state offers abundant land, major fiber routes, large natural-gas production, enormous wind and solar resources, a business-friendly tax structure, relatively fast development pathways, large metropolitan markets and an electrical system that is unusually independent from the rest of the country.

That final feature is both advantage and vulnerability. ERCOT operates an electrical island covering roughly 90 percent of Texas load. Because most of the system is not synchronously connected to the major eastern and western U.S. interconnections, Texas has historically had greater freedom over its market design. It has also had fewer opportunities to lean on neighboring regions during extreme system stress.

For a data-center developer, Texas can look almost ideal: build near gas, solar, wind, fiber and inexpensive land; connect to one of the country’s most competitive wholesale power markets; and construct at industrial scale. For a grid planner, the same proposition can look terrifying when many developers arrive simultaneously.

The Texas Tribune’s statewide project inventory found major concentrations in North Texas, Central Texas and West Texas, along with unusually large campuses in places such as Abilene. The scale is no longer measured only in buildings. It is measured in gigawatts.

One gigawatt is one thousand megawatts. That is enough power to make an individual technology campus relevant to regional generation and transmission planning.

OpenAI’s Stargate program makes the transformation visible. OpenAI says its U.S. infrastructure plan is targeting multi-gigawatt compute expansion, while its Texas projects include the operating Abilene campus and a separate 1.2-gigawatt site in Milam County being developed with SB Energy. Oracle’s project documentation shows the physical buildout in Abilene: substations, data halls and turbine infrastructure arriving alongside the compute.

The word cloud has become almost comically inadequate. This is heavy industry.

The Queue Is Not the Grid

Power-system forecasts traditionally move slowly because large industrial projects move slowly. A refinery, steel mill or chemical complex has a site, permits, engineering plans, financing and a visible construction schedule. Utilities can evaluate whether the project is credible and then decide which wires and generators need to be built.

AI disrupted that rhythm.

Data-center developers can option multiple sites, ask utilities about capacity in multiple regions, negotiate with several cloud tenants, and abandon one location when a faster path to power appears elsewhere. A single end customer can therefore appear indirectly in several interconnection requests.

Some requests are serious.

Some are placeholders.

Some are land speculation.

Some are competing versions of the same future project. Some are backed by companies that can write billion-dollar checks.

Others are power-point slides with acreage. That makes a raw queue number dangerous.

ERCOT’s preliminary 2032 load forecast of 367,790 MW is itself a planning construct, not a prediction carved into stone. The 438,595 MW associated with the larger request pipeline is even less like a forecast. But both numbers tell planners something important: the system has entered a period in which the option value of electricity access has become enormous.

A data-center company may be willing to spend money merely to preserve a place in the queue because power availability determines whether billions of dollars of servers can earn revenue. Texas regulators are therefore learning a new forecasting discipline: before planning for demand, prove the demand exists.

Batch Zero: Separating Real Projects From Paper Projects

ERCOT’s response is called Batch Zero. The Public Utility Commission approved the process in June. Instead of evaluating each major load in isolation, ERCOT now groups qualifying large-load projects—75 MW and above—into a system-wide study.

The logic is straightforward.

If three one-gigawatt campuses all want service in the same region, studying them one at a time can produce a fiction. Each individual study may assume transmission capacity that disappears once the next project connects.

A batch study asks the harder question: What happens if the credible projects arrive together? ERCOT requires projects to satisfy maturity and commitment criteria, submit technical models, demonstrate site control or other evidence, and provide information necessary for system-wide stability analysis. The process also includes pathways for projects that intend to limit how much they withdraw from the public grid by pairing the data center with on-site generation—a model ERCOT describes in part through Withdrawal-Limited Private Use Networks and “Bring Your Own Generation.”

The batch concept is important far beyond Texas. FERC has opened a national effort around large-load interconnection because regional grid operators across the United States face the same basic problem: giant loads want electricity on software-industry timelines, while the power system still builds on utility-industry timelines.

FERC commissioners have noted that today’s large loads can be orders of magnitude larger than ordinary commercial growth and may also change consumption very quickly. That combination—size plus speed—creates new reliability questions.

The first generation of internet infrastructure asked: Where is the fiber? The AI generation asks: Where is the firm megawatt—and how quickly can I reserve it?

Why AI Data Centers Are Different

Not all data centers are the same load. An enterprise server room might consume a few megawatts. A conventional colocation building may consume tens. A hyperscale cloud campus can move into the hundreds. Frontier AI campuses are pushing toward gigawatt scale.

The reason is density.

AI accelerators such as GPUs and specialized chips concentrate enormous computing capability—and heat—inside relatively small physical spaces. Training large models requires thousands or tens of thousands of accelerators communicating at extremely high speed. Inference at massive scale turns those models into continuously operating services.

Power enters through the utility or on-site generation, is transformed through high-voltage electrical systems, passes through UPS and power-distribution equipment, feeds accelerators, memory, storage and networking, and then largely leaves the computing process as heat. Cooling is therefore part of the electrical load.

So are pumps.

So are fans.

So are networking switches.

So are storage systems. So is the redundancy required to keep the campus alive when one part of the power chain fails.

AI also changes the economics of downtime. If a billion-dollar accelerator fleet sits idle because power is unavailable, the lost opportunity cost can be enormous. That pushes developers toward generation portfolios that are not merely cheap on average but available when needed.

This is why the phrase renewable energy versus fossil energy is too simple for the infrastructure problem. The data-center operator is buying several different products simultaneously:

energy, capacity, reliability, ramping, redundancy, geographic proximity and speed to commercial operation. Different technologies provide different combinations.

The National Electricity Reversal

The United States spent much of the twenty-first century becoming accustomed to flat electricity demand. Efficiency improved. Heavy industry changed. Appliances became better. Economic growth increasingly came from sectors that did not require proportional growth in electricity consumption.

That era has ended.

The U.S. Energy Information Administration says electricity use began growing more rapidly after 2020. Data centers are one important driver alongside manufacturing, electrification, electric vehicles and broader economic growth.

The Department of Energy’s Lawrence Berkeley National Laboratory estimated that U.S. data centers used about 176 terawatt-hours of electricity in 2023—roughly 4.4 percent of U.S. consumption—and could rise into a range of roughly 325 to 580 TWh by 2028. DOE’s later resource hub cites an updated central estimate approaching 11.8 percent of total U.S. electricity use by the end of the decade, with a wide uncertainty range.

The International Energy Agency reaches the same broad conclusion from another direction. Its base case projects global data-center electricity demand rising to around 945–950 TWh by 2030, with the United States responsible for the largest share of the increase. The IEA expects data centers to account for roughly half of U.S. electricity-demand growth to 2030.

That sounds manageable when expressed as a national percentage. The grid problem is local.

Ten percent of U.S. electricity spread evenly across fifty states is one thing.

A one-gigawatt campus trying to connect to one constrained transmission pocket is another. This is why AI’s electricity problem is better understood as a geography problem than as a national energy-total problem.

The United States may have enough energy in aggregate. It may not have enough deliverable power at the right substation on the right date.

Where the Power Will Come From

There is no single “AI power source.” The IEA’s U.S. outlook is instructive. Natural gas is currently the largest source of electricity serving U.S. data centers. Renewables are second, followed by nuclear and coal. Through 2030, the IEA expects natural gas and renewables to provide much of the incremental supply, while nuclear becomes increasingly important over longer horizons.

Texas makes that mixed portfolio especially visible because the state has abundant gas, the country’s largest wind fleet, rapidly growing solar and battery installations, two operating nuclear stations, and developers willing to experiment with private generation. The right question is therefore not:

Which resource wins? It is: Which combination can be financed, permitted, connected and delivered before the servers become obsolete?

Natural Gas: The Speed Advantage

Natural gas has an advantage the AI industry understands instinctively: it can often be turned into dependable megawatts faster than other firm generation. Texas sits on or near enormous gas-producing regions and has a dense pipeline network. Developers can pair gas turbines or reciprocating engines with data-center campuses, potentially avoiding years of waiting for major public-grid upgrades.

That does not make gas simple.

Large turbine supply is constrained. Pipeline capacity may need expansion. Air permits matter. Fuel-price exposure matters. The winter performance of gas infrastructure matters in Texas because the memory of Winter Storm Uri has made fuel security a reliability issue rather than an abstract commodity-market question.

But gas fits the near-term AI problem unusually well: large, controllable, dispatchable output with an existing industrial ecosystem. The IEA expects natural gas to be the largest source of additional U.S. data-center electricity supply through 2030 in its base case.

This is one reason the AI boom may complicate corporate decarbonization pledges. Technology companies can sign renewable-energy contracts on an annual accounting basis. Operating a physical AI campus every second of every day is a different problem. The gap between annual clean-energy matching and hourly firm electricity is where natural gas, batteries, nuclear and grid interconnection become central.

Solar and Batteries: Texas’ Fast-Build Advantage

If natural gas owns the dispatchability advantage, solar and batteries own the construction-speed advantage. Texas has added enormous amounts of solar capacity because modules can be installed relatively quickly, the state has excellent solar resources, and daytime output often aligns with hot-weather demand.

Batteries have changed the value of that solar. A battery can absorb energy when supply is abundant and release it during the evening ramp, during short scarcity events, or when the grid needs fast frequency response. At data-center scale, batteries can also support uninterruptible-power functions, smooth campus demand and potentially participate in grid markets.

But batteries are not magic generators. A four-hour battery can move four hours of energy. It cannot create energy during a prolonged multi-day shortage. Very large AI campuses therefore still need access to generation or the broader grid even if they install massive storage systems.

The interesting future is not solar or gas. It is hybrid architecture:

In that architecture, renewable power lowers fuel consumption and cost when available, batteries absorb short-duration volatility, and firm generation or the grid carries the system through longer mismatches. The economic optimum will vary by location.

So will the political optimum.

Nuclear: Firm Power With a Timing Problem

No energy source has benefited more rhetorically from the AI boom than nuclear power. The logic is obvious.

Nuclear plants produce large amounts of electricity continuously with low direct carbon emissions. Data centers want large amounts of electricity continuously. Technology companies also want to preserve climate commitments.

The alignment is compelling.

The timing is harder.

Existing nuclear plants can sign long-term contracts, pursue uprates and in some cases restart retired units. Those options can matter sooner than building new reactors. Small modular reactors and advanced reactor projects may eventually allow data centers or industrial campuses to pair dedicated nuclear generation with compute, but first-of-a-kind licensing, financing and construction timelines remain much longer than the software industry’s planning cycles.

The IEA expects the first wave of SMRs around the end of the decade rather than as the principal answer to the immediate 2026–2028 electricity surge. Nuclear is therefore strategically important but easy to mischaracterize. It is more likely to shape the next decade of AI infrastructure than solve the next eighteen months of interconnection requests.

The Shadow Grid: Bring Your Own Generation

The most consequential response to grid delay may be occurring outside the grid. Developers are increasingly pairing large data centers with dedicated generation.

ERCOT has formally incorporated the concept into its large-load process through categories that allow a campus to limit how much power it withdraws from the public system. Texas project developers are pursuing on-site gas plants, batteries and renewable portfolios. Oracle’s Abilene project page visibly tracks turbine and substation construction alongside the data halls.

The logic is commercial.

If a developer can spend several billion dollars on power infrastructure and bring a ten-billion-dollar compute campus online two years sooner, the electricity system is no longer merely an operating expense. It is part of the product.

This is producing something close to a shadow power grid: private generation, private substations, private batteries and private power-management systems built primarily for computing campuses. The public grid may remain connected, but its role changes.

Instead of being the sole source of energy, it can become backup, balancing market, export path or supplemental supply. That architecture can reduce stress on public infrastructure if designed well.

It can also create new complications. Private generators still need fuel. They still emit pollutants if they burn gas or diesel. They still interact electrically with the regional system. Their operators may want to import power when market prices are low and generate privately when prices rise. Poorly coordinated private and public systems can create new planning uncertainty. The question is no longer merely whether a data center is “on grid” or “off grid.” It is: How much of its demand is the public system truly obligated to serve under stress?

Transmission, Transformers and the Equipment Bottleneck

Electricity is not useful merely because it exists. It must reach the campus.

This sounds obvious until a developer discovers that a region has abundant wind, solar or gas generation but insufficient transmission capacity to move another gigawatt into one particular county. Transmission lines can take years to route, permit and construct. Substations require land and specialized equipment. Large power transformers are custom industrial machines with manufacturing constraints. Switchgear, breakers, turbines and high-voltage components all have supply chains that do not scale at semiconductor speed.

The Department of Energy’s transmission-needs work now explicitly identifies large load growth from data centers and manufacturing as a driver of new grid requirements. FERC is confronting the same issue nationally by forcing regional grid operators to explain whether their tariffs fairly allocate the cost of connecting very large customers.

That cost question is unavoidable.

Suppose a 1,000-MW campus cannot connect without a $2 billion transmission expansion. Who pays?

The data-center developer?

All customers in the utility territory? Customers across the entire regional grid?

Future users of the transmission system? The answer can determine whether the project appears economically attractive.

Texas Governor Greg Abbott has now directed state regulators toward a clearer principle: data centers should fully fund electric infrastructure needed specifically to serve them rather than shifting those costs onto residential customers. That principle is politically simple. Applying it is technically difficult because new infrastructure often benefits more than one customer over decades.

The Water Constraint

Electricity is only the first utility consumed by advanced computing. The second is water.

Data centers use water directly for certain cooling systems and indirectly through the water consumed by power generation. The amount varies enormously by cooling design, climate, workload, operating temperature and electricity source.

That variability is why simplistic claims such as “one AI query uses X bottles of water” are poor infrastructure analysis. The important question for a host community is peak capacity:

How much water can this facility require on the hottest day, when residents and the grid may also be under the most stress? A University of Texas at Austin research effort estimates that data centers could account for approximately 3 to 9 percent of Texas water use by 2040 under different growth assumptions when both direct and power-sector water are considered. The researchers emphasize uncertainty and the need for standardized reporting.

That is the right frame.

Texas is too large and hydrologically diverse for a statewide percentage to settle the question. A data center built near a water-rich metropolitan system is different from one built in a drought-prone rural county.

Water impact is local.

So is political resistance.

Governor Abbott has already called for water-efficient cooling requirements and better data-center water reporting. Those demands reflect a broader shift: communities increasingly want infrastructure disclosure before tax incentives and zoning decisions are finalized, not after the servers arrive.

Who Pays?

This may become the central political question of the AI infrastructure boom. A data center can be economically transformative without employing many permanent workers relative to its capital cost.

Texas’ tax code illustrates the bargain. Qualifying data centers can receive substantial sales-tax exemptions if they meet investment and job thresholds. Large projects can qualify for longer exemption periods with minimum investment, employment and transmission-capacity commitments.

The logic is familiar economic development: forego some tax revenue to attract capital, construction, equipment purchases, local tax base and strategic industry. The criticism is equally familiar:

why subsidize an industry that also requires enormous public infrastructure? The question becomes sharper if residents believe their electricity bills are rising because a hyperscale customer forced the utility to build new power plants or transmission.

That fear is not theoretical nationally. In Virginia and other mature data-center markets, regulators and utilities are already fighting over whether infrastructure and fuel costs are being allocated fairly.

Texas is trying to confront the problem earlier. Abbott has directed regulators to ensure that data-center growth reduces rather than increases household burden, and to make large-load customers fund infrastructure attributable to their service. The policy challenge is separating three categories:

That is why slogans fail.

“Data centers raise everyone’s bills” is too simple. “Data centers pay their own way” is too simple. The real answer depends on tariff design, timing, utilization, location and whether new generation arrives with the new load.

Can Data Centers Help the Grid?

The most interesting possibility is that AI campuses eventually become unusually flexible industrial loads. Some computation must happen immediately.

Some does not.

Model training jobs, batch inference, data processing, software compilation, synthetic-data generation and other workloads may be moved in time or across geography within service-level constraints. A company operating data centers in Texas, Arizona and Ohio could theoretically shift some non-urgent computation toward the region where electricity is cheapest or cleanest.

That creates an opportunity for the grid. During scarcity, a data center could reduce public-grid withdrawal by delaying flexible workloads, discharging batteries or turning on private generation.

During periods of excess solar or wind, it could increase compute. Research published this year shows the promise and the danger. Grid-modeling studies find that coordinated spatial and temporal workload shifting can reduce congestion and renewable curtailment. Other research shows that purely price-driven workload shifting can create new localized voltage and congestion problems if many centers respond to the same signal simultaneously.

The lesson is classic systems engineering: flexibility is valuable only when coordinated. A million smart loads all making the same “smart” decision can become one very large dumb event.

The Host-Community Bargain

Every major industrial transition eventually becomes local politics. A national strategy may call AI infrastructure essential to American competitiveness.

The person living beside the project experiences something else: construction traffic, housing pressure, industrial noise, land conversion, transmission lines, water demand, tax agreements and a changed horizon. The economic benefits are real too.

Data centers can create thousands of construction jobs, high-value electrical and mechanical work, new utility infrastructure, land payments, local tax revenue and a larger base for schools or county government. But the permanent operating workforce is usually far smaller than the construction workforce.

That mismatch changes the social bargain. A community may host billions of dollars of capital equipment while seeing relatively few permanent jobs.

Texas’ current political backlash reflects that asymmetry. Rural residents are increasingly asking whether tax incentives, power infrastructure, water use and quality-of-life impacts were evaluated under rules designed for a much smaller generation of data centers.

The policy answer is likely to be disclosure. How much power?

How much water?

How much public infrastructure?

How many permanent jobs?

How much tax revenue after incentives? What happens if the tenant leaves?

Who pays to decommission the infrastructure? What reliability service does the campus provide during grid emergencies?

Those are not anti-technology questions. They are the questions mature industrial policy should ask.

Critical Infrastructure and National Security

There is another reason the electricity story matters. Advanced compute is becoming strategic infrastructure.

AI models increasingly support defense, intelligence analysis, logistics, cybersecurity, scientific research, advanced manufacturing, financial systems and communications. A country that cannot supply reliable electricity to its compute infrastructure becomes dependent on countries that can.

The power stack therefore joins the semiconductor stack as part of technological sovereignty. NIST has begun treating AI data centers as a distinct security architecture involving not just cyber controls but facility construction, hardware, storage, supply chain, operational technology, power, sustainability, personnel and physical security.

That is the correct model.

A frontier AI data center can fail because of malware. It can also fail because of a transformer.

Or a cooling loop.

Or a gas-supply interruption.

Or a substation fire.

Or a transmission fault.

Or a water restriction.

Cybersecurity and physical infrastructure security are converging because the same facility depends on both. The cloud has become critical infrastructure with a very large electric bill.

The SURVXCOM AI Power Stack

To understand whether an AI infrastructure project is credible, it is useful to stop looking at the data-center rendering and evaluate the system beneath it. This point matters because the technology should be evaluated as part of the surrounding system rather than as an isolated claim or capability.

What to Watch Next

The Grid Becomes the Technology Story

Artificial intelligence has spent the last several years appearing to escape the limits of the physical world. Models became software services.

Software became cloud infrastructure. The cloud became something users assumed was everywhere.

Now the abstraction is reversing.

The model needs a GPU.

The GPU needs a rack.

The rack needs a cooling loop.

The cooling loop needs pumps, heat rejection and sometimes water. The building needs a substation.

The substation needs transmission.

The transmission system needs generation. The generation needs fuel, sunlight, wind, uranium or stored energy.

Every layer needs permits, equipment, land, capital and public acceptance. Texas is discovering this faster than most places because Texas invited both sides of the equation at once:

the technology industry and the energy industry. The result is not simply a data-center boom.

It is a test of whether the power system can evolve at digital speed without asking ordinary households to absorb risks they never agreed to take. If Texas gets that bargain right, it may show the rest of the country how to build the physical foundation of the AI age.

If it gets it wrong, the constraint on artificial intelligence may not be intelligence at all. It may be the wire.