By one estimate, Nvidia provides the hardware for more than 70% of the market for AI chips, and estimates from 2023/2024 suggested it was previously closer to 95%. This is because Nvidia chips are exceptionally efficient, performing the heavy computational tasks required for machine learning at a much lower energy threshold than common computer chips.

“This year we created a definition for sustainability in AI hardware based on energy processing per unit,” explains Michael Yow, director of rankings at Corporate Knights. The sustainability threshold used by Corporate Knights is measured in gigaFLOPS (floating-point operations per second) per watt, the number of computations performed per watt of energy. The threshold was first established by the Green500, a ranking in which Nvidia’s Grace Hopper chips power seven of the top 10 most energy-efficient supercomputer systems.

As a result, $75 billion of Nvidia’s 2025 revenues – about 57% of the total – was deemed sustainable according to the Corporate Knights methodology. That sustainable revenue was not only one of the largest gross totals in the ranking; it was also among the fastest growing, registering 123.5% growth since 2022. (All figures in U.S. dollars.)

The environmental impact of data centres populated with millions of these chips is increasingly well understood. One study in December reported that current AI infrastructure has roughly the carbon footprint of the city of New York (about 80 billion tons) and that its collective water use is in the range of the global bottled water industry (765 billion litres).

But these figures pale in comparison to the potential future impact. Last September, OpenAI’s Sam Altman issued an internal memo saying the company’s “audacious long-term goal is to build 250 gigawatts of capacity by 2033.” An analysis from Truthdig found that this would put ChatGPT’s energy use on par with India’s 1.5 billion people, which, depending how that energy is sourced, could produce carbon emissions twice that of ExxonMobil, currently one of the largest non-state emitters in the world.

Nvidia is the primary supplier of the chips populating OpenAI’s data centres, and the two companies recently announced a partnership to continue to build out this infrastructure.

The 2026 Global 100 list puts speed in the spotlight

The power sources

There is an irony to all this: the scale of these operations and their environmental impact is almost exclusively a result of the resource efficiency of Nvidia’s chips. Data centre operations for AI would be impossible if they relied on general-purpose servers. It’s the efficiency of Nvidia’s chips that has enabled this “insatiable demand” for computing power. This phenomenon, known as Jevons paradox, has a historical precedent: coal. In 1865, the English economist William Stanley Jevons observed that the more efficiently the industrial revolution’s primary fuel source could be used, the greater the demand grew.

In the United States, it’s estimated that AI data centres could account for as much as 12% of all energy consumption by 2028. Where is that power coming from? A 2025 report from the International Energy Agency found that coal currently accounts for about 30% of electricity generation for AI data centres globally, mostly in China and the United States. Renewables – wind, solar and hydro – account for 27%, and natural gas accounts for 26% (40% in the United States).

Corporate Knights’ Yow points out that Nvidia as a company cannot be held responsible for the power sources of the AI industry, even if its chips do make up much of its infrastructure. “Would we be having this conversation if all the electricity was 100% carbon-free?” he asks. “The problem is not with the product but rather with the lack of planning and faster deployment of renewable energy.”

Because of the outsized electrical needs, AI companies are exploring options for generating their own on-site power. Altman has, for example, personally invested in a start-up developing nuclear fusion, the holy grail of renewable power generation, and his company has joined the likes of Google, Microsoft and Amazon to invest in small modular nuclear reactor technology. Nevertheless, the most common on-site power sources at present are fossil fuels. At OpenAI’s Stargate facility under construction in West Texas, one of the largest in the world, they are currently deploying dozens of turbines adapted from aircraft engines.

That data centre is populated with Nvidia chips and operated by OpenAI and Oracle. Though the majority of Nvidia chips are not in Nvidia-owned or -operated data centres, the company reported that all offices and data centres under its control are powered exclusively by renewable energy and that it purchases additional carbon-free electricity to cover 100% of the footprint of its leased data centres.

The optimistic outlook

Nvidia has taken measures to address the water consumption at its own data centres, which use it to cool overheating servers. In its 2025 annual sustainability report, the company said it conducts annual water-risk assessments near all its facilities. Its sites in Santa Clara, California, and Hyderabad, India, have water treatment facilities so wastewater can be used for landscape irrigation. The company is also introducing closed-loop liquid cooling systems to reduce water use, and the Blackwell computing architecture – the successor to the Grace Hopper chips – is 300 times more water efficient than air-cooled architecture. (This does not address water use during manufacturing, however, which relies on ultra-pure jets of water to etch the silicone wafers, becoming contaminated in the process. Nor does it pertain to the data centres Nvidia does not control or operate.) A Nvidia spokesperson declined to participate in this story beyond referencing existing company communications.

Those bullish on the prospect of an AI-optimized future will argue that the environmental impact of AI itself will be dwarfed by the efficiencies the expansion of these systems will allow. On one blog, Nvidia cited reports looking at projected U.S. energy demand into 2035. If AI applications are “fully realized,” estimates suggest it could save nearly 2,500 petajoules of energy – about 25% of the entire country’s energy use in 2023.

Such savings would be transformative, and there is a much wider range of environmental applications that AI might come to revolutionize: wildfire detection and management, global climate simulations and extreme weather modelling, electrical grid applications to manage fires and outages, wildlife population tracking, materials science for cleantech applications, and other conservation or carbon-reduction efforts.

To date, the economy-altering value of these companies has been largely measured against their future potential, even as the imminent costs of doing business continue to mount. The environmental balance sheet isn’t so different.

Tristan Bronca is a magazine writer and editor based in Newmarket, Ontario.

The post There’s a paradox at the heart of Nvidia’s sustainability performance appeared first on Corporate Knights.

It’s energy efficiency.

Energy efficiency is the cheapest source of new energy because every kilowatt-hour that I save on the grid is one that someone else can use.

Within the world of computers, the price-performance efficiency is even more pronounced. A 2023 iPhone 15 is 60,000 times cheaper than a 1976 Cray-1 supercomputer. It’s also 180,000 times more energy efficient and 5,000 times more powerful.

Kevin Weil, chief product officer at OpenAI, says that OpenAI’s models are improving 10-fold every year in energy efficiency. Deferring compute-heavy tasks such as AI training to times when data-centre energy use loads are light and assigning loads across many data centres – in essence adjusting when and where power is used – could have the same benefit of building 200 gigawatts of new capacity, argues Amory Lovins, one of the cofounders of the Rocky Mountain Institute, which is on a mission the transform global energy systems. That’s far more than enough to power all projected new data centres from existing utility assets.

AI driving efficiency in buildings

Computers aren’t the only arena where efficiency is making giant leaps. Buildings represent roughly 40% of global energy consumption and 75% of U.S. electricity use, making them an enormous target for efficiency improvements. Research published in Nature Communications found that AI applications could reduce building energy consumption by 8% to 19% by 2050, and up to 40% when combined with other policies such as retrofits and low-carbon power generation. Building efficiency can free up more electricity use than AI will ever require.

This creates a fascinating dynamic: while AI consumes energy, it also enables efficiency gains across the broader economy. The key is ensuring that AI deployment is strategic and coupled with robust measurement and verification protocols. AI for buildings isn’t theoretical; companies like BrainBox AI and others are already deploying systems that optimize HVAC (heating, ventilation, air conditioning), lighting, and energy storage in real time based on occupancy, weather and grid conditions.

“Artificial intelligence is the latest development in a long-standing megatrend in which information, analysis and innovation have been replacing the waste of energy and materials that characterize the overbuilt technologies of the 20th-century fossil economy,” notes Ralph Torrie, director of research at Corporate Knights. “Of course it uses electricity, but not nearly as much as it displaces.”

There are many examples of rapid, unexpected gains in energy efficiency. In 2021, all the blockchain-based cryptocurrencies combined used somewhere between 190 and 250 terawatt-hours of electricity, which is just about 1% of global electricity demand that year, or roughly the same as Taiwan’s consumption. Critics called blockchain technology fundamentally unsustainable.

Then in September 2022, Ethereum underwent a major transformation it called “The Merge,” shifting its consensus mechanism from “proof of work” to “proof of stake.” This cut the network’s energy consumption by 99.9%. The network’s annual energy use dropped from 80 terawatt-hours – the same amount of electricity that Austria or Finland uses in a year – to just 0.01 terawatt-hours.

Proof of stake achieved these gains by eliminating wasteful competition. Instead of miners racing to solve puzzles, Ethereum now relies on validators who are chosen to create new blocks and confirm transactions based on the amount of Ether (Ethereum’s native currency) they have staked as collateral. This method secures the network through financial commitment rather than raw computational power.

When real limits appear, innovation often accelerates in unexpected ways. Constraints can become the spark for new possibilities.

— Anthony Di Iorio, Ethereum co-founder

The shift to proof of stake also enhanced its security, scalability and environmental sustainability. This landmark move positioned Ethereum as a leader in sustainable blockchain innovation and demonstrated that large-scale decentralized systems can evolve to meet global energy and climate goals without compromising performance or decentralization.

Ethereum co-founder Anthony Di Iorio says that Ethereum’s massive efficiency gain is part of a broader pattern of disruptive innovation across technologies: “When real limits appear, innovation often accelerates in unexpected ways. Constraints can become the spark for new possibilities.”

For Di Iorio, the key lesson is that “fundamental architectural redesign beats incremental optimization. When you eliminate structural inefficiency rather than just making a system slightly less inefficient, you unlock orders-of-magnitude improvements.”

DeepSeek, a Chinese AI company, also showed that algorithmic innovation can cut costs – in its case by 97% even under severe hardware constraints. Meanwhile, SETI@home coordinated millions of personal computers to create a huge virtual supercomputer to power its search for extra terrestrial intelligence.

Better orchestration, multi-tenant GPU sharing (in which multiple users share computational resources) and carbon-aware scheduling could push AI infrastructure use from today’s 25% to 40% to 55% to 60%, effectively doubling capacity without building any new facilities.

“The question isn’t whether AI will consume catastrophic amounts of energy,” Di Iorio notes. “The question is whether we will apply what we’ve already learned about radical efficiency gains.” That means implementing the measurement and transparency frameworks that enable market discipline and establishing the policy guardrails that ensure that efficiency serves sustainability rather than just enabling endless expansion.

The path forward: Three essential actions

Three near-term actions could save 15 to 70 terawatt-hours by 2028:

Default to efficiency. Major cloud platforms and AI frameworks should make lean, right-sized models the default choice rather than requiring developers to opt in. When efficiency becomes the path of least resistance rather than an extra step, adoption accelerates dramatically.

Focus on hardware use. Better workload-management software can increase effective use from today’s 40% to 55 to 60%.

Mandate transparency. Energy consumption per task should be as visible as price and latency. When enterprises and governments demand kilowatt-hours-per-million-tokens disclosure in their AI procurement, providers will compete on efficiency. Market mechanisms work, but they require information.

The projections of AI’s looming energy crisis aren’t wrong if we assume nothing changes. But stasis isn’t how technology works when talented people face hard constraints with clear incentives to solve them.

Jim Harris is a #1 international bestselling author writing on AI, disruption and innovation. He speaks at 50+ events a year.

The post The single greatest tool for reducing greenhouse gas emissions is energy efficiency appeared first on Corporate Knights.

But instead of using energy to cool their servers, some companies are trying to recover heat and put it to use. “What we try to do,” says Sacha van Geffen, a lead engineer at Leafcloud, “is actually see if we can place some of those compute resources that generate a lot of heat really close to a place where heat is needed.”

Leafcloud is an Amsterdam-based cloud services company that turns waste heat into warm showers through a distributed network of servers, or “leaf sites.” To do this, Leafcloud uses off-the-shelf components to build heat-waste-recovery systems inside the technical rooms of facilities like large apartment complexes, swimming pools or retirement homes, van Geffen says. The heat captured from the servers is used to produce hot water for the building.

Leafcloud isn’t the only company putting heat waste to work. U.K. company Heata offers a similar service to individual homeowners. The 2025 Paris Olympics pool was partially heated from data-centre servers run by the company Equinix. In Brooklyn, Bathhouse uses waste heat from Bitcoin mining rigs to warm its spa pools. Nexalus in Ireland harnesses heat waste to improve efficiency and lower data-centre energy consumption.

Heat waste from computing workloads is considered low-grade heat, meaning anything less than 100°C. “With low temperature grades,” says Amin Mohammadi, a PhD candidate at Simon Fraser University’s Laboratory for Alternative Energy Conversion, “converting that energy to thermal energy or cooling energy would be the best option that you have.” In other words, it’s too gentle to be converted to electricity such as with steam from a boiler, but with the help of heat pumps, it’s perfect for making a hot shower or keeping interiors toasty on a cold day. 

There are data centres and centralized district heating systems working together, too, such as Stockholm Exergi or Denmark’s Fjernvarme Fyn. This application could play an important role in heat-waste recovery from the growing demand for computing capacity. A recent Leafcloud white paper outlined that different thermal outputs could be optimized for different applications: high-performance computing for higher heat-waste temperatures, such as 60°C to 80°C, could be sent to district heating, whereas standard servers that produce temperatures in the 40°C to 60°C range could be used for building heat. 

The pieces are already in place to start making a difference. Leafcloud and other companies can capture heat waste now and without complex infrastructure that would otherwise take years to build out, such as constructing a data centre or retrofitting pipe systems for larger-scale municipal heating. By placing the heat source close to where it will be used, overall efficiency is improved, because it avoids the energy loss that occurs when heat has to move through a large system, Mohammadi says. 

In addition to heat-waste recovery, van Geffen says that Leafcloud’s distributed server system has other benefits. Using residual heat onsite not only reduces the reliance on other energy sources needed to heat water, such as natural gas; it also reduces costs. In a world where our data is so often controlled by secretive corporations, installing local servers in basements in Europe offers security and control, van Geffen says: “What we are offering is sovereign data storage, where the heat is also put to a good use.”

The post Companies are finding creative ways to use heat from data centres appeared first on Corporate Knights.

Planned in an area battling drought, the project’s water needs would also be vast, rivalling the annual use of hundreds of thousands of Alberta households.

Called “Wonder Valley,” the project is still nowhere near getting off the ground. But if it ever gets built as planned, it could pump out 25.7 to 30.5 megatonnes of emissions a year, depending on what turbines and gas are used, and whether or not it includes carbon capture, found Pembina Institute senior analyst Jason Wang, who crunched the numbers for The Mix.

“It would be the equivalent of a return to the era of mostly coal-fired electricity,” said Wang.

The reversal would set the province and Canada back about 20 years, matching the 27 megatonnes of coal emissions Alberta phased out between 2005 and 2023.

The data centre complex, which O’Leary claims will be the “largest on Earth,” would require the equivalent of about 10% of all gas supply in Alberta once fully operational.

It’s still just a concept, but in the Municipal District of Greenview where it’s planned, local officials are confident it will be built. “We’re about to pull off the largest project in Canadian history in this sector, and I think Greenview should be really proud of that,” chief administrative officer Stacey Wabick told council members at a budget review meeting in November.

Wonder Valley was announced with great fanfare last December, a few days after the Alberta government unveiled a new data centre strategy designed to attract $100 billion in investments.

At the time, Innovation Minister Nate Glubish said data centres would play a “significant role” in Premier Danielle Smith’s plan to double oil and gas production by increasing domestic demand for gas.

The land designated for Wonder Valley is south of Grande Prairie, about 460 kilometres northwest of Edmonton. It’s located on the massive Montney Formation, one of North America’s, and perhaps one of the world’s, largest gas reservoirs.

Wang said the first phase of the project, requiring 1,400 megawatts (MW) of power, would generate about 4.7 megatonnes of carbon dioxide per year using shale gas to run a combined-cycle gas turbine, the most efficient type of gas power. That would make it one of the largest industrial facilities in the province, he said.

OpenAI selects Argentina for its first data centre in Latin America

Carbon capture and storage has been promised for Wonder Valley, but not necessarily at start-up. Wang said a carbon pipeline would need to be built between the facility and an injection site. “With carbon-capture efficacy at about 80%, it would still mean the first phase of Wonder Valley would be 1.3 megatonnes per year of greenhouse gas emissions,” Wang said.

Another requirement for data centres is the large amount of water they use for cooling. But water is also needed to run gas power plants. Even though the Grande Prairie area is known for longer, colder winters, Wang estimated that the water needed for the fully completed data centre, including both cooling and for the power plant, would be between 112 and 195 billion litres per year. That is roughly one-third to two-thirds of the total annual water consumption of all the households in Alberta.

Multiple communities in the region are struggling with drought and water supply issues. Greenview itself declared an agricultural disaster due to drought this past summer, and that same day, its council approved adding more land to the purchase agreement being negotiated with O’Leary.

This story is part of the Hidden Wonder Valley investigative series by The Energy Mix. It has been edited to conform with Corporate Knights style.

Jody MacPherson is a freelance journalist based in Calgary. 

The post The Wonder Valley data centre in Alberta would erase two decades of emission reductions appeared first on Corporate Knights.

Days after OpenAI became the most valuable private company on the planet, a letter of intent was signed with a little-known Argentine firm called Sur Energy to develop all aspects of the $25-billion project.

“Sur Energy plans to ensure that the data-center ecosystem is powered by secure, efficient, and sustainable sources,” OpenAI said in a statement this week. The company said it “welcomes the potential opportunity to be an offtaker.” An offtake agreement – or a commitment to buy a portion of the energy produced – with OpenAI would help Sur Energy secure financing for the large-scale power facility.

“This milestone is about more than just infrastructure, it’s about putting AI into the hands of more people across Argentina,” added Sam Altman, the CEO of OpenAI, who praised the country’s programming prowess and said it was among the top Latin American countries for developers building on OpenAI tools. “We’re excited to work with Argentina as it builds toward becoming an AI hub for all of Latin America.”

The deal falls under the umbrella of Stargate, a company formed in the United States to shepherd $500 billion worth of investments to secure data capacity for OpenAI. Plans are in place for five U.S. sites – two in Texas, one in New Mexico, one in Ohio and another in an as yet undisclosed location in the Midwest – that together have the potential to generate nearly seven gigawatts of data-centre capacity over the next three years.

The five locations are expected to generate 25,000 onsite jobs and “tens of thousands of additional jobs across the U.S.,” OpenAI said. The company has said the jobs will range from short-term construction roles to full-time operational jobs, but the employment heft appears to be in the set-up phase. Bloomberg reported that Stargate’s first data centre, for example, was slated to create 57 permanent jobs.

OpenAI has development agreements for data centres in the United Kingdom, Norway, Japan and South Korea. The specifics of the Argentina data-centre proposal are not yet known, nor is it clear how many jobs will be created in Argentina, which is going through a volatile economic period as its currency falters. President Javier Milei’s government has implemented deep spending cuts to rein in public spending and offered new incentives for foreign investment, under which the new data centre is said to have landed.

OpenAI is operating full speed ahead in growth and energy usage. It launched in November 2022, and by July 2025, some 18 billion messages were being sent each week by 700 million users, according to a study that was co-authored by researchers at Harvard, Duke University and OpenAI. That represents about 10% of the global adult population. “For a new technology, this speed of global diffusion has no precedent,” the usage report’s authors noted. The study found that non-work messages have grown faster in the last year of study and now represent more than 70% of all messages sent.

Last month, OpenAI unveiled Sora 2, its latest video-generation model that is “more physically accurate, realistic and more controllable than prior systems.” Its test videos included a figure skater performing jumps with a cat on her head and a man doing backflips on a boogie board in a lake.

Sophisticated computing carries a high energy cost

A study this year found that AI use already represents 20% of data-centre consumption and is on track to double by the end of 2025. The International Energy Agency said data centres represented up to 1.5% of global electricity consumption in 2024. Various studies have shown that the energy required to answer even the simplest of queries uses a significant amount of water, which is used to cool the data centres working hard to create all the AI slop now filling the Internet.

The consumption levels are fast changing as the technology changes, and there are competing assessments, but the cumulative impact seems clear. Google, for example, reported that it consumed more than six billion gallons of water to cool all its data centres in 2023.

An investigation by The Washington Post published last year found that the amount of water required to cool the data-centre processing for a 100-word AI-composed email via ChatGPT4 was more than a 500ml bottle of water. Meanwhile, in a blog post earlier this year, Altman said that the average ChatGPT query uses about 0.000085 gallons of water, or “roughly one fifteenth of a teaspoon.”

With scant details available about what shape the data centre will take in Argentina, the discussion about what such a piece of infrastructure could mean has yet to coalesce.

Natalie Alcoba is a Buenos Aires–based journalist and senior editor at Corporate Knights.

The post OpenAI selects Argentina for its first data centre in Latin America appeared first on Corporate Knights.