Get ONMINE

Stay Ahead, Stay ONMINE

Everything you need to know about estimating AI’s energy and emissions burden

When we set out to write a story on the best available estimates for AI’s energy and emissions burden, we knew there would be caveats and uncertainties to these numbers. But, we quickly discovered, the caveats are the story too.  Measuring the energy used by an AI model is not like evaluating a car’s fuel economy or an appliance’s energy rating. There’s no agreed-upon method or public database of values. There are no regulators who enforce standards, and consumers don’t get the chance to evaluate one model against another.  Despite the fact that billions of dollars are being poured into reshaping energy infrastructure around the needs of AI, no one has settled on a way to quantify AI’s energy usage. Worse, companies are generally unwilling to disclose their own piece of the puzzle. There are also limitations to estimating the emissions associated with that energy demand, because the grid hosts a complicated, ever-changing mix of energy sources.  It’s a big mess, basically. So, that said, here are the many variables, assumptions, and caveats that we used to calculate the consequences of an AI query. (You can see the full results of our investigation here.) Measuring the energy a model uses Companies like OpenAI, dealing in “closed-source” models, generally offer access to their  systems through an interface where you input a question and receive an answer. What happens in between—which data center in the world processes your request, the energy it takes to do so, and the carbon intensity of the energy sources used—remains a secret, knowable only to the companies. There are few incentives for them to release this information, and so far, most have not. That’s why, for our analysis, we looked at open-source models. They serve as a very imperfect proxy but the best one we have. (OpenAI, Microsoft, and Google declined to share specifics on how much energy their closed-source models use.)  The best resources for measuring the energy consumption of open-source AI models are AI Energy Score, ML.Energy, and MLPerf Power. The team behind ML.Energy assisted us with our text and image model calculations, and the team behind AI Energy Score helped with our video model calculations. Text models AI models use up energy in two phases: when they initially learn from vast amounts of data, called training, and when they respond to queries, called inference. When ChatGPT was launched a few years ago, training was the focus, as tech companies raced to keep up and build ever-bigger models. But now, inference is where the most energy is used. The most accurate way to understand how much energy an AI model uses in the inference stage is to directly measure the amount of electricity used by the server handling the request. Servers contain all sorts of components—powerful chips called GPUs that do the bulk of the computing, other chips called CPUs, fans to keep everything cool, and more. Researchers typically measure the amount of power the GPU draws and estimate the rest (more on this shortly).  To do this, we turned to PhD candidate Jae-Won Chung and associate professor Mosharaf Chowdhury at the University of Michigan, who lead the ML.Energy project. Once we collected figures for different models’ GPU energy use from their team, we had to estimate how much energy is used for other processes, like cooling. We examined research literature, including a 2024 paper from Microsoft, to understand how much of a server’s total energy demand GPUs are responsible for. It turns out to be about half. So we took the team’s GPU energy estimate and doubled it to get a sense of total energy demands.  The ML.Energy team uses a batch of 500 prompts from a larger dataset to test models. The hardware is kept the same throughout; the GPU is a popular Nvidia chip called the H100. We decided to focus on models of three sizes from the Meta Llama family: small (8 billion parameters), medium (70 billion), and large (405 billion). We also identified a selection of prompts to test. We compared these with the averages for the entire batch of 500 prompts.  Image models Stable Diffusion 3 from Stability AI is one of the most commonly used open-source image-generating models, so we made it our focus. Though we tested multiple sizes of the text-based Meta Llama model, we focused on one of the most popular sizes of Stable Diffusion 3, with 2 billion parameters.  The team uses a dataset of example prompts to test a model’s energy requirements. Though the energy used by large language models is determined partially by the prompt, this isn’t true for diffusion models. Diffusion models can be programmed to go through a prescribed number of “denoising steps” when they generate an image or video, with each step being an iteration of the algorithm that adds more detail to the image. For a given step count and model, all images generated have the same energy footprint. The more steps, the higher quality the end result—but the more energy used. Numbers of steps vary by model and application, but 25 is pretty common, and that’s what we used for our standard quality. For higher quality, we used 50 steps.  We mentioned that GPUs are usually responsible for about half of the energy demands of large language model requests. There is not sufficient research to know how this changes for diffusion models that generate images and videos. In the absence of a better estimate, and after consulting with researchers, we opted to stick with this 50% rule of thumb for images and videos too. Video models Chung and Chowdhury do test video models, but only ones that generate short, low-quality GIFs. We don’t think the videos these models produce mirror the fidelity of the AI-generated video that many people are used to seeing.  Instead, we turned to Sasha Luccioni, the AI and climate lead at Hugging Face, who directs the AI Energy Score project. She measures the energy used by the GPU during AI requests. We chose two versions of the CogVideoX model to test: an older, lower-quality version and a newer, higher-quality one.  We asked Luccioni to use her tool, called Code Carbon, to test both and measure the results of a batch of video prompts we selected, using the same hardware as our text and image tests to keep as many variables as possible the same. She reported the GPU energy demands, which we again doubled to estimate total energy demands.  Tracing where that energy comes from After we understand how much energy it takes to respond to a query, we can translate that into the total emissions impact. Doing so requires looking at the power grid from which data centers draw their electricity.  Nailing down the climate impact of the grid can be complicated, because it’s both interconnected and incredibly local. Imagine the grid as a system of connected canals and pools of water. Power plants add water to the canals, and electricity users, or loads, siphon it out. In the US, grid interconnections stretch all the way across the country. So, in a way, we’re all connected, but we can also break the grid up into its component pieces to get a sense for how energy sources vary across the country.  Understanding carbon intensity The key metric to understand here is called carbon intensity, which is basically a measure of how many grams of carbon dioxide pollution are released for every kilowatt-hour of electricity that’s produced.  To get carbon intensity figures, we reached out to Electricity Maps, a Danish startup company that gathers data on grids around the world. The team collects information from sources including governments and utilities and uses them to publish historical and real-time estimates of the carbon intensity of the grid. You can find more about their methodology here.  The company shared with us historical data from 2024, both for the entire US and for a few key balancing authorities (more on this in a moment). After discussions with Electricity Maps founder Olivier Corradi and other experts, we made a few decisions about which figures we would use in our calculations.  One way to measure carbon intensity is to simply look at all the power plants that are operating on the grid, add up the pollution they’re producing at the moment, and divide that total by the electricity they’re producing. But that doesn’t account for the emissions that are associated with building and tearing down power plants, which can be significant. So we chose to use carbon intensity figures that account for the whole life cycle of a power plant.  We also chose to use the consumption-based carbon intensity of energy rather than production-based. This figure accounts for imports and exports moving between different parts of the grid and best represents the electricity that’s being used, in real time, within a given region.  For most of the calculations you see in the story, we used the average carbon intensity for the US for 2024, according to Electricity Maps, which is 402.49 grams of carbon dioxide equivalent per kilowatt-hour.  Understanding balancing authorities While understanding the picture across the entire US can be helpful, the grid can look incredibly different in different locations.  One way we can break things up is by looking at balancing authorities. These are independent bodies responsible for grid balancing in a specific region. They operate mostly independently, though there’s a constant movement of electricity between them as well. There are 66 balancing authorities in the US, and we can calculate a carbon intensity for the part of the grid encompassed by a specific balancing authority. Electricity Maps provided carbon intensity figures for a few key balancing authorities, and we focused on several that play the largest roles in data center operations. ERCOT (which covers most of Texas) and PJM (a cluster of states on the East Coast, including Virginia, Pennsylvania, and New Jersey) are two of the regions with the largest burden of data centers, according to research from the Harvard School of Public Health.  We added CAISO (in California) because it covers the most populated state in the US. CAISO also manages a grid with a significant number of renewable energy sources, making it a good example of how carbon intensity can change drastically depending on the time of day. (In the middle of the day, solar tends to dominate, while natural gas plays a larger role overnight, for example.) One key caveat here is that we’re not entirely sure where companies tend to send individual AI inference requests. There are clusters of data centers in the regions we chose as examples, but when you use a tech giant’s AI model, your request could be handled by any number of data centers owned or contracted by the company. One reasonable approximation is location: It’s likely that the data center servicing a request is close to where it’s being made, so a request on the West Coast might be most likely to be routed to a data center on that side of the country.  Explaining what we found To better contextualize our calculations, we introduced a few comparisons people might be more familiar with than kilowatt-hours and grams of carbon dioxide. In a few places, we took the amount of electricity estimated to be used by a model and calculated how long that electricity would be able to power a standard microwave, as well as how far it might take someone on an e-bike.  In the case of the e-bike, we assumed an efficiency of 25 watt-hours per mile, which falls in the range of frequently cited efficiencies for a pedal-assisted bike. For the microwave, we assumed an 800-watt model, which falls within the average range in the US.  We also introduced a comparison to contextualize greenhouse gas emissions: miles driven in a gas-powered car. For this, we used data from the US Environmental Protection Agency, which puts the weighted average fuel economy of vehicles in the US in 2022 at 393 grams of carbon dioxide equivalent per mile.  Predicting how much energy AI will use in the future After measuring the energy demand of an individual query and the emissions it generated, it was time to estimate how all of this added up to national demand.  There are two ways to do this. In a bottom-up analysis, you estimate how many individual queries there are, calculate the energy demands of each, and add them up to determine the total. For a top-down look, you estimate how much energy all data centers are using by looking at larger trends.  Bottom-up is particularly difficult, because, once again, closed-source companies do not share such information and declined to talk specifics with us. While we can make some educated guesses to give us a picture of what might be happening right now, looking into the future is perhaps better served by taking a top-down approach. This data is scarce as well. The most important report was published in December by the Lawrence Berkeley National Laboratory, which is funded by the Department of Energy, and the report authors noted that it’s only the third such report released in the last 20 years. Academic climate and energy researchers we spoke with said it’s a major problem that AI is not considered its own economic sector for emissions measurements, and there aren’t rigorous reporting requirements. As a result, it’s difficult to track AI’s climate toll.  Still, we examined the report’s results, compared them with other findings and estimates, and consulted independent experts about the data. While much of the report was about data centers more broadly, we drew out data points that were specific to the future of AI.  Company goals We wanted to contrast these figures with the amounts of energy that AI companies themselves say they need. To do so, we collected reports by leading tech and AI companies about their plans for energy and data center expansions, as well as the dollar amounts they promised to invest. Where possible, we fact-checked the promises made in these claims. (Meta and Microsoft’s pledges to use more nuclear power, for example, would indeed reduce the carbon emissions of the companies, but it will take years, if not decades, for these additional nuclear plants to come online.)  Requests to companies We submitted requests to Microsoft, Google, and OpenAI to have data-driven conversations about their models’ energy demands for AI inference. None of the companies made executives or leadership available for on-the-record interviews about their energy usage. This story was supported by a grant from the Tarbell Center for AI Journalism.

When we set out to write a story on the best available estimates for AI’s energy and emissions burden, we knew there would be caveats and uncertainties to these numbers. But, we quickly discovered, the caveats are the story too. 

Measuring the energy used by an AI model is not like evaluating a car’s fuel economy or an appliance’s energy rating. There’s no agreed-upon method or public database of values. There are no regulators who enforce standards, and consumers don’t get the chance to evaluate one model against another. 

Despite the fact that billions of dollars are being poured into reshaping energy infrastructure around the needs of AI, no one has settled on a way to quantify AI’s energy usage. Worse, companies are generally unwilling to disclose their own piece of the puzzle. There are also limitations to estimating the emissions associated with that energy demand, because the grid hosts a complicated, ever-changing mix of energy sources. 

It’s a big mess, basically. So, that said, here are the many variables, assumptions, and caveats that we used to calculate the consequences of an AI query. (You can see the full results of our investigation here.)

Measuring the energy a model uses

Companies like OpenAI, dealing in “closed-source” models, generally offer access to their  systems through an interface where you input a question and receive an answer. What happens in between—which data center in the world processes your request, the energy it takes to do so, and the carbon intensity of the energy sources used—remains a secret, knowable only to the companies. There are few incentives for them to release this information, and so far, most have not.

That’s why, for our analysis, we looked at open-source models. They serve as a very imperfect proxy but the best one we have. (OpenAI, Microsoft, and Google declined to share specifics on how much energy their closed-source models use.) 

The best resources for measuring the energy consumption of open-source AI models are AI Energy Score, ML.Energy, and MLPerf Power. The team behind ML.Energy assisted us with our text and image model calculations, and the team behind AI Energy Score helped with our video model calculations.

Text models

AI models use up energy in two phases: when they initially learn from vast amounts of data, called training, and when they respond to queries, called inference. When ChatGPT was launched a few years ago, training was the focus, as tech companies raced to keep up and build ever-bigger models. But now, inference is where the most energy is used.

The most accurate way to understand how much energy an AI model uses in the inference stage is to directly measure the amount of electricity used by the server handling the request. Servers contain all sorts of components—powerful chips called GPUs that do the bulk of the computing, other chips called CPUs, fans to keep everything cool, and more. Researchers typically measure the amount of power the GPU draws and estimate the rest (more on this shortly). 

To do this, we turned to PhD candidate Jae-Won Chung and associate professor Mosharaf Chowdhury at the University of Michigan, who lead the ML.Energy project. Once we collected figures for different models’ GPU energy use from their team, we had to estimate how much energy is used for other processes, like cooling. We examined research literature, including a 2024 paper from Microsoft, to understand how much of a server’s total energy demand GPUs are responsible for. It turns out to be about half. So we took the team’s GPU energy estimate and doubled it to get a sense of total energy demands. 

The ML.Energy team uses a batch of 500 prompts from a larger dataset to test models. The hardware is kept the same throughout; the GPU is a popular Nvidia chip called the H100. We decided to focus on models of three sizes from the Meta Llama family: small (8 billion parameters), medium (70 billion), and large (405 billion). We also identified a selection of prompts to test. We compared these with the averages for the entire batch of 500 prompts. 

Image models

Stable Diffusion 3 from Stability AI is one of the most commonly used open-source image-generating models, so we made it our focus. Though we tested multiple sizes of the text-based Meta Llama model, we focused on one of the most popular sizes of Stable Diffusion 3, with 2 billion parameters. 

The team uses a dataset of example prompts to test a model’s energy requirements. Though the energy used by large language models is determined partially by the prompt, this isn’t true for diffusion models. Diffusion models can be programmed to go through a prescribed number of “denoising steps” when they generate an image or video, with each step being an iteration of the algorithm that adds more detail to the image. For a given step count and model, all images generated have the same energy footprint.

The more steps, the higher quality the end result—but the more energy used. Numbers of steps vary by model and application, but 25 is pretty common, and that’s what we used for our standard quality. For higher quality, we used 50 steps. 

We mentioned that GPUs are usually responsible for about half of the energy demands of large language model requests. There is not sufficient research to know how this changes for diffusion models that generate images and videos. In the absence of a better estimate, and after consulting with researchers, we opted to stick with this 50% rule of thumb for images and videos too.

Video models

Chung and Chowdhury do test video models, but only ones that generate short, low-quality GIFs. We don’t think the videos these models produce mirror the fidelity of the AI-generated video that many people are used to seeing. 

Instead, we turned to Sasha Luccioni, the AI and climate lead at Hugging Face, who directs the AI Energy Score project. She measures the energy used by the GPU during AI requests. We chose two versions of the CogVideoX model to test: an older, lower-quality version and a newer, higher-quality one. 

We asked Luccioni to use her tool, called Code Carbon, to test both and measure the results of a batch of video prompts we selected, using the same hardware as our text and image tests to keep as many variables as possible the same. She reported the GPU energy demands, which we again doubled to estimate total energy demands. 

Tracing where that energy comes from

After we understand how much energy it takes to respond to a query, we can translate that into the total emissions impact. Doing so requires looking at the power grid from which data centers draw their electricity. 

Nailing down the climate impact of the grid can be complicated, because it’s both interconnected and incredibly local. Imagine the grid as a system of connected canals and pools of water. Power plants add water to the canals, and electricity users, or loads, siphon it out. In the US, grid interconnections stretch all the way across the country. So, in a way, we’re all connected, but we can also break the grid up into its component pieces to get a sense for how energy sources vary across the country. 

Understanding carbon intensity

The key metric to understand here is called carbon intensity, which is basically a measure of how many grams of carbon dioxide pollution are released for every kilowatt-hour of electricity that’s produced. 

To get carbon intensity figures, we reached out to Electricity Maps, a Danish startup company that gathers data on grids around the world. The team collects information from sources including governments and utilities and uses them to publish historical and real-time estimates of the carbon intensity of the grid. You can find more about their methodology here. 

The company shared with us historical data from 2024, both for the entire US and for a few key balancing authorities (more on this in a moment). After discussions with Electricity Maps founder Olivier Corradi and other experts, we made a few decisions about which figures we would use in our calculations. 

One way to measure carbon intensity is to simply look at all the power plants that are operating on the grid, add up the pollution they’re producing at the moment, and divide that total by the electricity they’re producing. But that doesn’t account for the emissions that are associated with building and tearing down power plants, which can be significant. So we chose to use carbon intensity figures that account for the whole life cycle of a power plant. 

We also chose to use the consumption-based carbon intensity of energy rather than production-based. This figure accounts for imports and exports moving between different parts of the grid and best represents the electricity that’s being used, in real time, within a given region. 

For most of the calculations you see in the story, we used the average carbon intensity for the US for 2024, according to Electricity Maps, which is 402.49 grams of carbon dioxide equivalent per kilowatt-hour. 

Understanding balancing authorities

While understanding the picture across the entire US can be helpful, the grid can look incredibly different in different locations. 

One way we can break things up is by looking at balancing authorities. These are independent bodies responsible for grid balancing in a specific region. They operate mostly independently, though there’s a constant movement of electricity between them as well. There are 66 balancing authorities in the US, and we can calculate a carbon intensity for the part of the grid encompassed by a specific balancing authority.

Electricity Maps provided carbon intensity figures for a few key balancing authorities, and we focused on several that play the largest roles in data center operations. ERCOT (which covers most of Texas) and PJM (a cluster of states on the East Coast, including Virginia, Pennsylvania, and New Jersey) are two of the regions with the largest burden of data centers, according to research from the Harvard School of Public Health. 

We added CAISO (in California) because it covers the most populated state in the US. CAISO also manages a grid with a significant number of renewable energy sources, making it a good example of how carbon intensity can change drastically depending on the time of day. (In the middle of the day, solar tends to dominate, while natural gas plays a larger role overnight, for example.)

One key caveat here is that we’re not entirely sure where companies tend to send individual AI inference requests. There are clusters of data centers in the regions we chose as examples, but when you use a tech giant’s AI model, your request could be handled by any number of data centers owned or contracted by the company. One reasonable approximation is location: It’s likely that the data center servicing a request is close to where it’s being made, so a request on the West Coast might be most likely to be routed to a data center on that side of the country. 

Explaining what we found

To better contextualize our calculations, we introduced a few comparisons people might be more familiar with than kilowatt-hours and grams of carbon dioxide. In a few places, we took the amount of electricity estimated to be used by a model and calculated how long that electricity would be able to power a standard microwave, as well as how far it might take someone on an e-bike. 

In the case of the e-bike, we assumed an efficiency of 25 watt-hours per mile, which falls in the range of frequently cited efficiencies for a pedal-assisted bike. For the microwave, we assumed an 800-watt model, which falls within the average range in the US. 

We also introduced a comparison to contextualize greenhouse gas emissions: miles driven in a gas-powered car. For this, we used data from the US Environmental Protection Agency, which puts the weighted average fuel economy of vehicles in the US in 2022 at 393 grams of carbon dioxide equivalent per mile. 

Predicting how much energy AI will use in the future

After measuring the energy demand of an individual query and the emissions it generated, it was time to estimate how all of this added up to national demand. 

There are two ways to do this. In a bottom-up analysis, you estimate how many individual queries there are, calculate the energy demands of each, and add them up to determine the total. For a top-down look, you estimate how much energy all data centers are using by looking at larger trends. 

Bottom-up is particularly difficult, because, once again, closed-source companies do not share such information and declined to talk specifics with us. While we can make some educated guesses to give us a picture of what might be happening right now, looking into the future is perhaps better served by taking a top-down approach.

This data is scarce as well. The most important report was published in December by the Lawrence Berkeley National Laboratory, which is funded by the Department of Energy, and the report authors noted that it’s only the third such report released in the last 20 years. Academic climate and energy researchers we spoke with said it’s a major problem that AI is not considered its own economic sector for emissions measurements, and there aren’t rigorous reporting requirements. As a result, it’s difficult to track AI’s climate toll. 

Still, we examined the report’s results, compared them with other findings and estimates, and consulted independent experts about the data. While much of the report was about data centers more broadly, we drew out data points that were specific to the future of AI. 

Company goals

We wanted to contrast these figures with the amounts of energy that AI companies themselves say they need. To do so, we collected reports by leading tech and AI companies about their plans for energy and data center expansions, as well as the dollar amounts they promised to invest. Where possible, we fact-checked the promises made in these claims. (Meta and Microsoft’s pledges to use more nuclear power, for example, would indeed reduce the carbon emissions of the companies, but it will take years, if not decades, for these additional nuclear plants to come online.) 

Requests to companies

We submitted requests to Microsoft, Google, and OpenAI to have data-driven conversations about their models’ energy demands for AI inference. None of the companies made executives or leadership available for on-the-record interviews about their energy usage.

This story was supported by a grant from the Tarbell Center for AI Journalism.

Shape
Shape
Stay Ahead

Explore More Insights

More Articles

Stay ahead with more perspectives on cutting-edge power, infrastructure, energy,  bitcoin and AI solutions. Explore these articles to uncover strategies and insights shaping the future of industries.

Shape

Survey: AMD continues to take server share from Intel

Dean McCarron, president of Mercury, said it’s not AMD stealing Intel business but mostly a case of AMD growing faster than Intel. “AMD’s growth rate in the quarter was multiples of Intel’s, resulting in significant server share gains,” he said in a research note. “Server processor shipments were definitively the

Read More »

Government asked to remove carbon capture from climate delivery plan

Opposition to the Net Zero Teesside (NZT) gas-fired power station with carbon capture has ramped up as developers await a court decision next month on the project’s climate impact. Environmental consultant Andrew Boswell has called for ministers to remove carbon capture, usage and storage (CCUS) from the new carbon budget delivery plan. “Ed Miliband is doing some very good work with his ambitions for a clean power system by 2030,” said Boswell. “However, he fails to see that the plans for fossil fuel-based carbon capture which he inherited from the previous government are hugely damaging to the climate. The government is conveniently ignoring the true climate costs and impacts when they fall outside UK territory.” NZT would be the first gas-fired plant in the UK to capture carbon emissions. A final judgment on the appeal case against its impact on the climate is expected by mid-June, Boswell told Energy Voice. He has partnered with Zero Hour, the campaign for the Climate and Nature Bill, asking for ex-territorial carbon emissions to be “properly accounted for by the UK”, and adequate policies. They have teamed with climate groups and MP Watch to initiate a dialogue with government and mobilise local constituents to write to ministers. A DESNZ spokesperson said in a response to a request for comment: “Carbon capture, usage and storage is vital to boost our energy security, and the Climate Change Committee describes it as a ‘necessity not an option’ for reaching our climate goals. “There is no route to protecting jobs in our industrial heartlands and securing the future of heavy industry in the UK without it.” The government’s net zero strategy was previously thrown out of the High Court in 2022 as it initially failed to show how the plan could be feasibly delivered, including the necessary risk

Read More »

Zonal pricing feud escalates as Octopus accuses pollster

London-based opinion polling firm Survation has been forced to withdraw a survey on zonal pricing after Octopus accused the company of breaching an industry code of conduct. The firm, which pulled the results on 16 May before publication, in a written response to a request for comment, admitted that questions in the survey “were not subject to our usual internal review procedures”. Government is expected to make a decision on reforming the UK electricity market by the end of next month, with discussions on regional power pricing forming a key part of the possible measures. Regional pricing would carve up the UK power market into regional zones with local power prices based on supply and demand. Energy utilities and renewable energy companies have pushed back against plans for zonal pricing, warning that it would jeopardise billions of pounds of investment in technologies such as offshore wind. Aberdeen-based communications agency True North commissioned and distributed the results from the survey on zonal pricing on 16 May under an embargo. The communications agency then contacted the media with an urgent message to rescind the results of the poll on the same day they were issued , following a complaint from Octopus. Energy supplier Octopus claimed that the survey asked people “highly leading questions”, known as ‘push polling’. A spokesman for True North said: “True North has partnered with Survation for several years, publicising numerous polls that gauge public opinion on politics and issues of importance to the Scottish business community. “Survation made us aware this particular data set had not been processed through its normal filters and apologised to True North for the error.” Survation said in an emailed response to a request to comment: “It is common practice for clients to propose draft polling questions in their own words. Survation’s role

Read More »

UK may need to build equivalent of 10 Rough facilities to meet hydrogen storage demand

The UK may need to build 10 times the capacity of the existing Rough offshore gas storage facility to meet its future hydrogen needs, a Holyrood committee has heard. Appearing before the Net Zero, Energy and Transport Committee, University of Edinburgh professor Stuart Hazeldine said the UK could effectively need to build “another 20 salt caverns” onshore to provide sufficient hydrogen storage. It comes after Rough owner Centrica earlier this week warned it may shut down the UK’s largest gas storage facility without additional government support. Hazeldine made the comments as the Holyrood committee considered the future of the Grangemouth refinery and industrial complex. Grangemouth refinery owner Petroineos ceased operations at the site earlier this year, leading to the loss of 400 jobs. A joint UK and Scottish government project, known as Project Willow, has identified up to £13 billion of future investment opportunities at Grangemouth. These include the production of synthetic fuels, including sustainable aviation fuel (SAF) and e-methanol, as well as opportunities for e-ammonia and fuel switching. Many of these require hydrogen as part of the production process, which is part of why the Scottish government has said hydrogen is one of Scotland’s “greatest industrial opportunities since oil and gas”. Grangemouth hydrogen future But for Grangemouth to produce these fuels, and jobs, the UK and Scotland will need to significantly scale up production of blue and green hydrogen production. The UK sector is also calling for clarity on the development of hydrogen transport and storage networks. But Scotland’s hydrogen sector has warned a lack of government support is leading to “disappointing results” from efforts to build up the country’s hydrogen economy. Asked by the Holyrood committee how the UK can bring down the costs of hydrogen storage, Hazeldine said the country will need to invest in more underground

Read More »

Vistra to Acquire Lotus Gas Power Plants for $1.9B

Vistra Corp. has signed a definitive agreement to buy seven natural gas-fired generation facilities spread across different states from Lotus Infrastructure Partners for $1.9 billion. The acquisition consists of five combined-cycle gas turbine plants and two combustion turbine plants in California, Delaware, New York, Pennsylvania and Rhode Island, “further geographically diversifying Vistra’s natural gas fleet”, Vistra said in an online statement. The plants have a combined capacity of 2,557 megawatts (MW). The biggest of them is the Fairless plant in Pennsylvania, which produces up to 1,320 MW. The transaction price translates to $743 per kilowatt, subject to net working capital adjustments, according to the Irving, Texas-based power utility. “Vistra expects to fund the transaction with the assumption of an existing term loan from Lotus and cash on hand”, it said. “Although the principal amount of the term loan to be assumed is subject to change, it is currently expected to be approximately 50 percent of the consideration at closing. “The purchase price implies a multiple of approximately 7x 2026 Adjusted EBITDA [earnings before interest, taxes, depreciation and amortization), excluding any potential synergies”. The parties expect to complete the transaction late this year or early 2026, subject to regulatory approvals including anti-trust clearance. Vistra said the acquisition will not change its previously set capital allocation plans including $300 million in annual dividends and at least $1 billion in share buybacks each year. “We are excited to announce another opportunistic expansion of our generation footprint in some of our key competitive markets”, said Vistra president and chief executive Jim Burke. “We believe natural gas-fired generation will continue to play an ever-increasing role in the reliability, affordability, and flexibility of U.S. power grids for years to come. “The addition of this attractive portfolio of combined cycle and peaking assets allows Vistra to serve growing power demand while

Read More »

There Were 3 Drivers to Monday’s USA Natural Gas Price Pullback

There were three drivers to yesterday’s U.S. natural gas price pullback, Art Hogan, Chief Market Strategist at B. Riley Wealth, told Rigzone in an exclusive interview on Tuesday. “Weather forecasts maintained mild conditions. The threat of re-escalating tariffs on U.S. trading partners stood as a further demand-side risk, while production remained steady,” Hogan said. Rigzone contacted the White House for comment on Hogan’s statement. At the time of writing, the White House has not responded to Rigzone with a comment. In a separate exclusive interview today, Phil Flynn, a senior market analyst at the PRICE Futures Group, said “shoulder season is on full display, with producers coming out of maintenance raising output while demand is weak”. Flynn also told Rigzone that the weather is in a “goldilocks” state, “not too hot and not too cold”. In an EBW Analytics Group report sent to Rigzone by the EBW team today, Eli Rubin, an analyst at the company, said the June natural gas contract “shed another 22.1¢ yesterday, testing support as low as $3.098, as indelibly weak shoulder season fundamentals continue to pummel the market”. “The front-month has now lost 68.2¢ in six sessions, and technicals remain bearish,” Rubin warned. In the report, Rubin went on to state that gas production at a monthly high over the weekend appears to have spooked longs, although the analyst added that gains were small in absolute terms. Rubin also noted in the report that “early-season heating demand for Weeks 2 and 3 may total 10 cooling degree days below 30-year norms”. Rubin said in the report that near-term pricing is likely to remain dominated by trader positioning ahead of the Memorial Day holiday and June contract options expiration and final settlement next week. “Over the next 30-45 days, however, confidence is increasing for a rebound if and when

Read More »

America’s climate brain drain is real — and it’s just getting started

Joe Curtatone is president of the Alliance for Climate Transition. While China now leads the U.S. in the number of patents, the United States has long been the global leader in scientific innovation and technological breakthroughs. From launching the internet to decoding the human genome, our universities, national labs and startup ecosystems have powered progress not just at home, but across the globe. But that leadership is now at serious risk — especially in climate and tough tech sectors — thanks to a combination of short-sighted federal policies, tariffs, attacks on academic institutions and an increasingly hostile posture toward international students and researchers. At the heart of this crisis is a growing brain drain — a phenomenon where top talent, including scientists, engineers and entrepreneurs, increasingly look outside the U.S. to study, work and build the technologies of the future. The warning signs are everywhere. In recent months, we’ve seen sweeping attempts by the current administration to undermine the independence and integrity of university research, especially in areas tied to climate science, advanced energy and environmental policy. Key research grants have been frozen or redirected based on political ideology. Experts have been sidelined or dismissed. Federal agencies have been pressured to strip climate and environmental language from their reports. The message is clear: If your work doesn’t align with the administration’s narrative, you may be punished for it. Simultaneously, international scholars and students — the lifeblood of many of our top science and engineering programs — are being pushed away. Visas are being revoked without due process, delays are mounting and the overall tone from federal leadership is one of suspicion and exclusion. This isn’t just bad optics. It directly threatens our innovation economy and the recovery could take decades. Engineering, environmental science and computer science international students don’t just

Read More »

Tariff uncertainty weighs on networking vendors

“Our guide assumes current tariffs and exemptions remain in place through the quarter. These include the following: China at 30%, partially offset by an exemption for semiconductors and certain electronic components; Mexico and Canada at 25% for the components and products that are not eligible for the current exemptions,” Cisco CFO Scott Herron told Wall Street analysts in the company’s quarterly earnings report on May 14. At this time, Cisco expects little impact from tariffs on steel and aluminum and retaliatory tariffs, Herron said. “We’ll continue to leverage our world-class supply chain team to help mitigate the impact,” he said, adding that “the flexibility and agility we have built into our operations over the last few years, the size and scale of our supply chain, provides us some unique advantages as we support our customers globally.” “Once the tariff scenario stabilizes, there [are] steps that we can take to mitigate it, as you’ve seen us do with China from the first Trump administration. And only after that would we consider price [increases],” Herron said. Similarly, Extreme Networks noted the changing tariff conditions during its earnings call on April 30. “The tariff situation is very dynamic, I think, as everybody knows and can appreciate, and it’s kind of hard to call. Yes, there was concern initially given the magnitude of tariffs,” said Extreme Networks CEO Ed Meyercord on the earnings call. “The larger question is, will all of the changes globally in trade and tariff policy have an impact on demand? And that’s hard to call at this point. And we’re going to hold as far as providing guidance or judgment on that until we have finality come July.” Financial news Meanwhile, AI is fueling high expectations and influencing investments in enterprise campus and data center environments.

Read More »

Liquid cooling becoming essential as AI servers proliferate

“Facility water loops sometimes have good water quality, sometimes bad,” says My Troung, CTO at ZutaCore, a liquid cooling company. “Sometimes you have organics you don’t want to have inside the technical loop.” So there’s one set of pipes that goes around the data center, collecting the heat from the server racks, and another set of smaller pipes that lives inside individual racks or servers. “That inner loop is some sort of technical fluid, and the two loops exchange heat across a heat exchanger,” says Troung. The most common approach today, he says, is to use a single-phase liquid — one that stays in liquid form and never evaporates into a gas — such as water or propylene glycol. But it’s not the most efficient option. Evaporation is a great way to dissipate heat. That’s what our bodies do when we sweat. When water goes from a liquid to a gas it’s called a phase change, and it uses up energy and makes everything around it slightly cooler. Of course, few servers run hot enough to boil water — but they can boil other liquids. “Two phase is the most efficient cooling technology,” says Xianming (Simon) Dai, a professor at University of Texas at Dallas. And it might be here sooner than you think. In a keynote address in March at Nvidia GTC, Nvidia CEO Jensen Huang unveiled the Rubin Ultra NVL576, due in the second half of 2027 — with 600 kilowatts per rack. “With the 600 kilowatt racks that Nvidia is announcing, the industry will have to shift very soon from single-phase approaches to two-phase,” says ZutaCore’s Troung. Another highly-efficient cooling approach is immersion cooling. According to a Castrol survey released in March, 90% of 600 data center industry leaders say that they are considering switching to immersion

Read More »

Cisco taps OpenAI’s Codex for AI-driven network coding

“If you want to ask Codex a question about your codebase, click “Ask”. Each task is processed independently in a separate, isolated environment preloaded with your codebase. Codex can read and edit files, as well as run commands including test harnesses, linters, and type checkers. Task completion typically takes between 1 and 30 minutes, depending on complexity, and you can monitor Codex’s progress in real time,” according to OpenAI. “Once Codex completes a task, it commits its changes in its environment. Codex provides verifiable evidence of its actions through citations of terminal logs and test outputs, allowing you to trace each step taken during task completion,” OpenAI wrote. “You can then review the results, request further revisions, open a GitHub pull request, or directly integrate the changes into your local environment. In the product, you can configure the Codex environment to match your real development environment as closely as possible.” OpenAI is releasing Codex as a research preview: “We prioritized security and transparency when designing Codex so users can verify its outputs – a safeguard that grows increasingly more important as AI models handle more complex coding tasks independently and safety considerations evolve. Users can check Codex’s work through citations, terminal logs and test results,” OpenAI wrote.  Internally, technical teams at OpenAI have started using Codex. “It is most often used by OpenAI engineers to offload repetitive, well-scoped tasks, like refactoring, renaming, and writing tests, that would otherwise break focus. It’s equally useful for scaffolding new features, wiring components, fixing bugs, and drafting documentation,” OpenAI stated. Cisco’s view of agentic AI Patel stated that Codex is part of the developing AI agent world, where Cisco envisions billions of AI agents will work together to transform and redefine the architectural assumptions the industry has relied on. Agents will communicate within and

Read More »

US companies are helping Saudi Arabia to build an AI powerhouse

AMD announced a five-year, $10 billion collaboration with Humain to deploy up to 500 megawatts of AI compute in Saudi Arabia and the US, aiming to deploy “multi-exaflop capacity by early 2026.” AWS, too, is expanding its data centers in Saudi Arabia to bolster Humain’s cloud infrastructure. Saudi Arabia has abundant oil and gas to power those data centers, and is growing its renewable energy resources with the goal of supplying 50% of the country’s power by 2030. “Commercial electricity rates, nearly 50% lower than in the US, offer potential cost savings for AI model training, though high local hosting costs due to land, talent, and infrastructure limit total savings,” said Eric Samuel, Associate Director at IDC. Located near Middle Eastern population centers and fiber optic cables to Asia, these data centers will offer enterprises low-latency cloud computing for real-time AI applications. Late is great There’s an advantage to being a relative latecomer to the technology industry, said Eric Samuel, associate director, research at IDC. “Saudi Arabia’s greenfield tech landscape offers a unique opportunity for rapid, ground-up AI integration, unburdened by legacy systems,” he said.

Read More »

AMD, Nvidia partner with Saudi startup to build multi-billion dollar AI service centers

Humain will deploy the Nvidia Omniverse platform as a multi-tenant system to drive acceleration of the new era of physical AI and robotics through simulation, optimization and operation of physical environments by new human-AI-led solutions. The AMD deal did not discuss the number of chips involved in the deal, but it is valued at $10 billion. AMD and Humain plan to develop a comprehensive AI infrastructure through a network of AMD-based AI data centers that will extend from Saudi Arabia to the US and support a wide range of AI workloads across corporate, start-up, and government markets. Think of it as AWS but only offering AI as a service. AMD will provide its AI compute portfolio – Epyc, Instinct, and FPGA networking — and the AMD ROCm open software ecosystem, while Humain will manage the delivery of the hyperscale data center, sustainable power systems, and global fiber interconnects. The partners expect to activate a multi-exaflop network by early 2026, supported by next-generation AI silicon, modular data center zones, and a software platform stack focused on developer enablement, open standards, and interoperability. Amazon Web Services also got a piece of the action, announcing a more than $5 billion investment to build an “AI zone” in the Kingdom. The zone is the first of its kind and will bring together multiple capabilities, including dedicated AWS AI infrastructure and servers, UltraCluster networks for faster AI training and inference, AWS services like SageMaker and Bedrock, and AI application services such as Amazon Q. Like the AMD project, the zone will be available in 2026. Humain only emerged this month, so little is known about it. But given that it is backed by Crown Prince Salman and has the full weight of the Kingdom’s Public Investment Fund (PIF), which ranks among the world’s largest and

Read More »

Check Point CISO: Network segregation can prevent blackouts, disruptions

Fischbein agrees 100% with his colleague’s analysis and adds that education and training can help prevent such incidents from occurring. “Simulating such a blackout is impossible, it has never been done,” he acknowledges, but he is committed to strengthening personal and team training and risk awareness. Increased defense and cybersecurity budgets In 2025, industry watchers expect there will be an increase in the public budget allocated to defense. In Spain, one-third of the budget will be allocated to increasing cybersecurity. But for Fischbein, training teams is much more important than the budget. “The challenge is to distribute the budget in a way that can be managed,” he notes, and to leverage intuitive and easy-to-use platforms, so that organizations don’t have to invest all the money in training. “When you have information, management, users, devices, mobiles, data centers, clouds, cameras, printers… the security challenge is very complex. You have to look for a security platform that makes things easier, faster, and simpler,” he says. ” Today there are excellent tools that can stop all kinds of attacks.” “Since 2010, there have been cybersecurity systems, also from Check Point, that help prevent this type of incident from happening, but I’m not sure that [Spain’s electricity blackout] was a cyberattack.” Leading the way in email security According to Gartner’s Magic Quadrant, Check Point is the leader in email security platforms. Today email is still responsible for 88% of all malicious file distributions. Attacks that, as Fischbein explains, enter through phishing, spam, SMS, or QR codes. “There are two challenges: to stop the threats and not to disturb, because if the security tool is a nuisance it causes more harm than good. It is very important that the solution does not annoy [users],” he stresses. “As almost all attacks enter via e-mail, it is

Read More »

Microsoft will invest $80B in AI data centers in fiscal 2025

And Microsoft isn’t the only one that is ramping up its investments into AI-enabled data centers. Rival cloud service providers are all investing in either upgrading or opening new data centers to capture a larger chunk of business from developers and users of large language models (LLMs).  In a report published in October 2024, Bloomberg Intelligence estimated that demand for generative AI would push Microsoft, AWS, Google, Oracle, Meta, and Apple would between them devote $200 billion to capex in 2025, up from $110 billion in 2023. Microsoft is one of the biggest spenders, followed closely by Google and AWS, Bloomberg Intelligence said. Its estimate of Microsoft’s capital spending on AI, at $62.4 billion for calendar 2025, is lower than Smith’s claim that the company will invest $80 billion in the fiscal year to June 30, 2025. Both figures, though, are way higher than Microsoft’s 2020 capital expenditure of “just” $17.6 billion. The majority of the increased spending is tied to cloud services and the expansion of AI infrastructure needed to provide compute capacity for OpenAI workloads. Separately, last October Amazon CEO Andy Jassy said his company planned total capex spend of $75 billion in 2024 and even more in 2025, with much of it going to AWS, its cloud computing division.

Read More »

John Deere unveils more autonomous farm machines to address skill labor shortage

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Self-driving tractors might be the path to self-driving cars. John Deere has revealed a new line of autonomous machines and tech across agriculture, construction and commercial landscaping. The Moline, Illinois-based John Deere has been in business for 187 years, yet it’s been a regular as a non-tech company showing off technology at the big tech trade show in Las Vegas and is back at CES 2025 with more autonomous tractors and other vehicles. This is not something we usually cover, but John Deere has a lot of data that is interesting in the big picture of tech. The message from the company is that there aren’t enough skilled farm laborers to do the work that its customers need. It’s been a challenge for most of the last two decades, said Jahmy Hindman, CTO at John Deere, in a briefing. Much of the tech will come this fall and after that. He noted that the average farmer in the U.S. is over 58 and works 12 to 18 hours a day to grow food for us. And he said the American Farm Bureau Federation estimates there are roughly 2.4 million farm jobs that need to be filled annually; and the agricultural work force continues to shrink. (This is my hint to the anti-immigration crowd). John Deere’s autonomous 9RX Tractor. Farmers can oversee it using an app. While each of these industries experiences their own set of challenges, a commonality across all is skilled labor availability. In construction, about 80% percent of contractors struggle to find skilled labor. And in commercial landscaping, 86% of landscaping business owners can’t find labor to fill open positions, he said. “They have to figure out how to do

Read More »

2025 playbook for enterprise AI success, from agents to evals

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More 2025 is poised to be a pivotal year for enterprise AI. The past year has seen rapid innovation, and this year will see the same. This has made it more critical than ever to revisit your AI strategy to stay competitive and create value for your customers. From scaling AI agents to optimizing costs, here are the five critical areas enterprises should prioritize for their AI strategy this year. 1. Agents: the next generation of automation AI agents are no longer theoretical. In 2025, they’re indispensable tools for enterprises looking to streamline operations and enhance customer interactions. Unlike traditional software, agents powered by large language models (LLMs) can make nuanced decisions, navigate complex multi-step tasks, and integrate seamlessly with tools and APIs. At the start of 2024, agents were not ready for prime time, making frustrating mistakes like hallucinating URLs. They started getting better as frontier large language models themselves improved. “Let me put it this way,” said Sam Witteveen, cofounder of Red Dragon, a company that develops agents for companies, and that recently reviewed the 48 agents it built last year. “Interestingly, the ones that we built at the start of the year, a lot of those worked way better at the end of the year just because the models got better.” Witteveen shared this in the video podcast we filmed to discuss these five big trends in detail. Models are getting better and hallucinating less, and they’re also being trained to do agentic tasks. Another feature that the model providers are researching is a way to use the LLM as a judge, and as models get cheaper (something we’ll cover below), companies can use three or more models to

Read More »

OpenAI’s red teaming innovations define new essentials for security leaders in the AI era

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More OpenAI has taken a more aggressive approach to red teaming than its AI competitors, demonstrating its security teams’ advanced capabilities in two areas: multi-step reinforcement and external red teaming. OpenAI recently released two papers that set a new competitive standard for improving the quality, reliability and safety of AI models in these two techniques and more. The first paper, “OpenAI’s Approach to External Red Teaming for AI Models and Systems,” reports that specialized teams outside the company have proven effective in uncovering vulnerabilities that might otherwise have made it into a released model because in-house testing techniques may have missed them. In the second paper, “Diverse and Effective Red Teaming with Auto-Generated Rewards and Multi-Step Reinforcement Learning,” OpenAI introduces an automated framework that relies on iterative reinforcement learning to generate a broad spectrum of novel, wide-ranging attacks. Going all-in on red teaming pays practical, competitive dividends It’s encouraging to see competitive intensity in red teaming growing among AI companies. When Anthropic released its AI red team guidelines in June of last year, it joined AI providers including Google, Microsoft, Nvidia, OpenAI, and even the U.S.’s National Institute of Standards and Technology (NIST), which all had released red teaming frameworks. Investing heavily in red teaming yields tangible benefits for security leaders in any organization. OpenAI’s paper on external red teaming provides a detailed analysis of how the company strives to create specialized external teams that include cybersecurity and subject matter experts. The goal is to see if knowledgeable external teams can defeat models’ security perimeters and find gaps in their security, biases and controls that prompt-based testing couldn’t find. What makes OpenAI’s recent papers noteworthy is how well they define using human-in-the-middle

Read More »
More Articles