In the era of AI reasoning, the computing power competition has entered a “power grab mode”! From hoarding GPUs to hoarding megawatts, Tesla (TSLA.US) locks down Arizona optical storage

The Zhitong Finance App learned that ContourGlobal, a well-known industrial project developer, said that Tesla (TSLA.US), a leader in electric vehicles, AI, autonomous driving, and robotics, has reached a major long-term agreement and will spend huge sums of money to purchase electricity produced by a large-scale solar and battery project in Arizona, USA supported by private equity giant KKR.

This latest power supply agreement is rare for Tesla with battery energy storage and solar asset networks, so it also highlights the need for Tesla to sign more power purchase agreements as the electricity market in the US market tightens due to the growth of data centers driven by sky-level artificial intelligence reasoning demand.

According to a statement, the company will sell 90% of the Sterling project's power generation to Tesla. The facility is scheduled to be put into operation in 2028 and will include 509 megawatts of peak solar power generation capacity and 360 megawatts of battery energy storage system with sustainable discharge for four hours. The parties did not disclose financial terms.

Tesla locks down the Arizona optical storage project, and the computing power race escalates into an energy battle

This agreement is uncommon for Tesla, and highlights the increasing demand for companies to sign so-called power purchase agreements as the US electricity market tightens due to the growth in data center scale driven by demand for artificial intelligence inference computing power. In addition to Tesla, other US tech giants such as Microsoft, Google, and Amazon have been signing such contracts for years, and have been the main driving force behind the boom in clean energy such as solar and wind energy over the past decade. The combination of renewable energy projects such as solar energy with large-scale battery energy storage systems has become one of the fastest ways for tech giants to add large-scale power capacity, especially in regions where it is difficult for US utilities to meet growing electricity demand.

The core reason for choosing four hours was that the main solution of the project was not “completely disconnected from the grid 24 hours a day,” but instead transferred large amounts of low-cost solar energy in Arizona at noon to peak electricity consumption of about 4 to 5 hours in the evening. When solar energy declines rapidly in the afternoon and residential and commercial loads are still high, centralized battery discharge can ease the typical “duck curve” and nighttime electricity shortages. ContourGlobal clearly stated that the energy storage section is aimed at supplying clean electricity during peak periods in the evening; CAISO also indicated that most of the large energy storage systems in California are four-hour lithium-ion batteries, which mainly support systems on hot evenings after solar power has declined.

Four hours is also a common balance between current lithium-ion energy storage economy and electricity market value. The arbitrage value of electricity prices is usually concentrated in the first few hours of the highest price per day: the highest price difference can be captured in the first hour, and the new revenue usually decreases with each additional hour; at the same time, extending the energy storage period requires an almost proportional increase in battery cells, cabinets, fire protection systems, and capital investment. According to ARPA-E under the US Department of Energy, short-term lithium batteries are suitable for handling intraday energy transfer from noon to evening, while continuous low wind, low sunshine, or lack of electricity across the day require longer energy storage.

Tesla generally buys power resources from local utility companies and grid systems. The company did not respond to requests for comment outside of normal office hours. Notably, Tesla also has some power generation type hard core assets.

The large power plant will be connected to a grid system managed by the Western Regional Power Administration and will have access to California. With the boom in AI data center construction, power grids across the US are facing an unprecedented trend of demand growth, as power grid areas with large-scale oversupply are under severe pressure from short supply.

According to long-term statistics from the US Energy Information Administration dating back to the late 1990s, the average electricity price for ordinary American consumers is expected to rise 4.3% this year, reaching an unprecedented record 14.22 cents per kilowatt-hour.

This is the first agreement between Tesla and ContourGlobal. Such agreements usually last 10 to 15 years, and can not only provide buyer forces with long-term monitoring and forecasting of electricity resource costs, but also provide developers with the certainty of revenue needed to finance new projects.

The Sterling project will be the largest renewable energy asset in ContourGlobal's portfolio. The company acquired the project in late 2024 and will trade the remaining 10% of the electricity generation on its own in the market.

From buying GPUs to buying power plants: tech giants bring their own power supplies, electricity becomes the ultimate constraint on AI capital expenditure

The Trump administration is speeding up approval of advanced nuclear reactors, pushing technology companies to pay for new power generation and grid upgrades in data centers, and Tesla's early lockdown of large-scale optical storage projects in Arizona all essentially point to the same structural change: AI competition has been upgraded from “whether it can obtain enough chips” to “whether it can obtain enough stable, predictable, and electricity that does not drive up residents' electricity prices over a long period of time.”

The Berkeley Laboratory under the US Department of Energy estimates that by 2030, data centers may account for about 11.8% of the total electricity consumption in the US, with a range of 9.5% to 15.3% in different scenarios; the International Energy Agency predicts that global data center electricity consumption will increase from about 485 terawatt-hours in 2025 to about 950 terawatt-hours in 2030, which means data center electricity demand will nearly double.

The so-called AI inference era means that electricity demand is shifting from phased model training loads to high-frequency online computation continuously triggered by applications such as search, AI agents, video generation, enterprise Copilot, and autonomous driving. The energy efficiency of single inference will certainly decline due to improvements in chips and algorithms, but the expansion of model calls, context length, inference chain depth, and number of concurrent users may eat up efficiency dividends more quickly; the IEA (International Energy Agency) points out that AI training and model use can also cause large and rapidly changing power loads, making energy storage and power supply flexibility an important condition for reliable operation. Therefore, what will limit AI expansion in the future will not only be the number of GPUs, but “effective megawatts” composed of usable electricity, grid-connected capacity, transformers, transmission lines, cooling systems, and backup power sources.

The policy side has begun to reallocate costs around this bottleneck. The Trump administration's “Electricity Bill User Protection Pledge” requires large technology companies to take responsibility for their own additional load by building or expanding power generation facilities, bearing transmission and distribution upgrades, and signing special electricity price agreements; however, the promise is still mainly voluntary, and the actual binding and cost isolation effects are still disputed.

At the same time, nuclear regulatory reforms require a re-examination and compression of the advanced reactor approval process, and the NRC has also introduced a new advanced reactor licensing path. The meaning of the policy mix is very clear: to rapidly increase capacity in the short term through natural gas, solar energy, and energy storage, and to rely on stable power sources such as nuclear power for a long time to support high-utilization AI infrastructure.

Tesla's purchase of 90% of the Sterling project's power generation is a microcosm of this trend on the corporate procurement side. The project is scheduled to be put into operation in 2028. It is equipped with 509 megawatts of peak photovoltaics and 360 megawatts of four-hour energy storage of about 1.4 gigawatt-hours, and is expected to generate more than 1 terawatt-hour per year. It can shift daytime photovoltaics to nighttime peaks and lock in costs and supply through long-term power purchase agreements, but four-hour energy storage still cannot independently bear the all-weather load of the data center, and ultimately still requires the collaboration of power grids, nuclear power, natural gas, or other stable power sources. This shows that Tesla's move is not necessarily directly equivalent to supplying power to an AI data center, but it clearly shows that in the context of tightening electricity, large technology and manufacturing companies are elevating long-term energy procurement to the same important strategic level as chip procurement.