Why Some AI Models Spew 50 Times More Greenhouse Gas to Answer the Same Question

Like it or not, large language models have quickly become embedded into our lives. And due to their intense energy and water needs, they might also be causing us to spiral even faster into climate chaos. Some LLMs, though, might be releasing more planet-warming pollution than others, a new study finds.
Queries made to some models generate up to 50 times more carbon emissions than others, according to a new study published in Frontiers in Communication. Unfortunately, and perhaps unsurprisingly, models that are more accurate tend to have the biggest energy costs.
It's hard to estimate just how bad LLMs are for the environment, but some studies have suggested that training ChatGPT used up to 30 times more energy than the average American uses in a year. What isn't known is whether some models have steeper energy costs than their peers as they're answering questions.
Researchers from the Hochschule München University of Applied Sciences in Germany evaluated 14 LLMs ranging from 7 to 72 billion parameters—the levers and dials that fine-tune a model's understanding and language generation—on 1,000 benchmark questions about various subjects.
LLMs convert each word or parts of words in a prompt into a string of numbers called a token. Some LLMs, particularly reasoning LLMs, also insert special 'thinking tokens' into the input sequence to allow for additional internal computation and reasoning before generating output. This conversion and the subsequent computations that the LLM performs on the tokens use energy and releases CO2.
The scientists compared the number of tokens generated by each of the models they tested. Reasoning models, on average, created 543.5 thinking tokens per question, whereas concise models required just 37.7 tokens per question, the study found. In the ChatGPT world, for example, GPT-3.5 is a concise model, whereas GPT-4o is a reasoning model.
This reasoning process drives up energy needs, the authors found. 'The environmental impact of questioning trained LLMs is strongly determined by their reasoning approach,' study author Maximilian Dauner, a researcher at Hochschule München University of Applied Sciences, said in a statement. 'We found that reasoning-enabled models produced up to 50 times more CO2 emissions than concise response models.'
The more accurate the models were, the more carbon emissions they produced, the study found. The reasoning model Cogito, which has 70 billion parameters, reached up to 84.9% accuracy—but it also produced three times more CO2 emissions than similarly sized models that generate more concise answers.
'Currently, we see a clear accuracy-sustainability trade-off inherent in LLM technologies,' said Dauner. 'None of the models that kept emissions below 500 grams of CO2 equivalent achieved higher than 80% accuracy on answering the 1,000 questions correctly.' CO2 equivalent is the unit used to measure the climate impact of various greenhouse gases.
Another factor was subject matter. Questions that required detailed or complex reasoning, for example abstract algebra or philosophy, led to up to six times higher emissions than more straightforward subjects, according to the study.
There are some caveats, though. Emissions are very dependent on how local energy grids are structured and the models that you examine, so it's unclear how generalizable these findings are. Still, the study authors said they hope that the work will encourage people to be 'selective and thoughtful' about the LLM use.
'Users can significantly reduce emissions by prompting AI to generate concise answers or limiting the use of high-capacity models to tasks that genuinely require that power,' Dauner said in a statement.

Try Our AI Features

Explore what Daily8 AI can do for you:

Learn with AIFact CheckingText to SpeechAI Summary

Related Articles

Forbes

16 minutes ago

• Forbes

A Tale Of Two Cities In Venture Capital

Like a desert and an oasis, there is a tale of two cities in venture capital 'It was the best of times, it was the worst of times.' On the heels of Super Venture & Super Return Berlin, and a LP/GP summit in Montreal organized by Inovia, I wanted to reflect on what felt like a panoramic view of the industry, across 1) managers of VC funds 2) startups raising capital and 3) incumbents in the public markets. What struck me most wasn't any one deal or datapoint, but the duality of two starkly different realities. The technology industry today, for startups and venture capital firms alike, is best described as a tale of two cities. In venture capital, the divergence could not be more stark. Research by Pitchbook demonstrated that nine firms collected 50% of all venture funding last year. The top 30 took the vast majority, across a few hundred firms that announced raises. At SuperReturn, partners from megafunds spoke about oversubscribed raises. Meanwhile, emerging managers are scraping to survive and define their edge. As Scott Hartley, co-founder and managing partner of Everywhere Ventures told me: 'For emerging managers differentiation is key, as venture returns heavily skew due to a long tail of large outcomes. While it may seem safer to get on-base at bats, median venture returns barely outpace other asset classes. The reason allocators look at VC is for asymmetric exposure, which means providing capital to yet undiscovered drivers of the future economy.' First-time and even second-time funds have faced arguably one of the hardest fundraising markets in history. Similar research from Pitchbook suggests a risk of certain emerging managers dying off. One panelist summed it up succinctly: 'Consolidation is the new diversification.' The gulf between the haves and have-nots is perhaps even more stark in startups, the companies VCs fund. Asaf Horresh of Vintage explained during his Super Venture keynote: 'if you're AI it's 2021'. Capital is flowing freely into foundation model startups, AI infrastructure plays, and workflow automation layers. Several venture firms have reoriented their entire sourcing around AI, and LPs are watching closely. But for nearly everyone else, Asaf put it perfectly: 'you're in the desert.' Even solid companies are seeing down rounds or resorting to insider-led bridges. The bar has never been higher. While I am nervous about nosebleed valuations and non-Camel strategies- the potential for AI justified. But the capital imbalance is striking. At SuperReturn, AI was on nearly every main stage. As Arpan Punyani, Co-Founder and General Partner at Garuda Ventures told me: 'Like the transistor, cloud, or mobile, AI is a foundational enabling technology. If you're not building an AI company at some level, the bar is just incredibly high to raise venture capital now.' Carta noted that seed deals are growing scarcer, demonstrating the growing scarcity. However, average valuations, pulled up by A.I. startups have increased. The other vector in startups was geography. Today, like it or not, capital is re-concentrating in the old capitals partially because of AI: San Francisco, London, Tel Aviv. Startups in non-central ecosystems are finding it that much harder to raise capital. From side conversations, there's a growing sense that global GPs need to 'earn the right' to go abroad again. The irony (and certainly central to my day job at Fluent Ventures) some of the best startups are being built in emerging ecosystems—with less burn, less competition, and lower valuations. The bifurcation is not limited to venture. In the public markets, the 'Magnificent Seven' (Nvidia, Meta, Apple, et al.) account for the lion's share of returns. Small caps and international equities remain underloved. It's the same pattern: consolidation of value, divergence of experience. In the technology industry, the power law is the rule of the game. But the magnitude of the power law is not. As exits increase, releasing DPI, market sentiment calms, and desire for diversification outside the U.S. grows, I expect much of this to normalize. But this will take time and there will certainly be collateral damage. But, as history tells us, sometimes, the revolution begins in the second city.

New York Times

20 minutes ago

• New York Times

Europe's Growing Fear: How Trump Might Use U.S. Tech Dominance Against It

When President Trump issued an executive order in February against the chief prosecutor of the International Criminal Court for investigating Israel for war crimes, Microsoft was suddenly thrust into the middle of a geopolitical fight. For years, Microsoft had supplied the court — which is based in The Hague in the Netherlands and investigates and prosecutes human rights breaches, genocides and other crimes of international concern — with digital services such as email. Mr. Trump's order abruptly threw that relationship into disarray by barring U.S. companies from providing services to the prosecutor, Karim Khan. Soon after, Microsoft, which is based in Redmond, Wash., suspended Mr. Khan's I.C.C. email account, freezing him out of communications with colleagues just a few months after the court had issued an arrest warrant for Prime Minister Benjamin Netanyahu of Israel for his country's actions in Gaza. Microsoft's swift compliance with Mr. Trump's order, reported earlier by The Associated Press, shocked policymakers across Europe. It was a wake-up call for a problem far bigger than just one email account, stoking fears that the Trump administration would leverage America's tech dominance to penalize opponents, even in allied countries like the Netherlands. 'The I.C.C. showed this can happen,' said Bart Groothuis, a former head of cybersecurity for the Dutch Ministry of Defense who is now a member of the European Parliament. 'It's not just fantasy.' Mr. Groothuis once supported U.S. tech firms but has done a '180-degree flip-flop,' he said. 'We have to take steps as Europe to do more for our sovereignty.' Want all of The Times? Subscribe.

Android Authority

25 minutes ago

• Android Authority

Transparent specs at last: Why the new smartphone energy labels are a big deal

Aamir Siddiqui / Android Authority TL;DR The European Union has introduced mandatory energy efficiency labels for all smartphones and tablets sold within its region starting today, June 20, 2025. These labels provide detailed information on energy use, battery longevity, drop resistance, IP rating for dust and water resistance, and repairability scores. Complementing this rollout are ecodesign requirements mandating durable batteries, extended software support, readily available spare parts, and more. It's not easy to discern how efficient, durable, or repairable a smartphone or tablet is, just from its spec sheet or marketing materials. Manufacturers are all too happy to hide the information that shows their product in a bad light, and we can only hope for regulations to come along to bring some much-needed transparency. The European Union is leading the way, thanks to its new energy label and ecodesign requirements that go into force today in the region, forcing smartphone makers to change for the better. EU energy labels: What does the label mean? Starting today, June 20, 2025, smartphones and tablets sold in the EU must feature a new energy efficiency label that displays information on their energy efficiency, battery longevity, protection from dust and water, and against accidental drops. It doesn't end here, as phones and tablets must also showcase a repairability score. In the energy level above, graphic number 1 points to the scale of energy efficiency classes ranging from A (best) to G (worst), but icon number 2 showcases where the product is actually placed (C, in this reference example). Icon 3 denotes the battery endurance per cycle, in hours and minutes per full battery charge, indicating how many hours of use you can expect from a full charge on a new device. This number is determined based on a standardized test of usage, so hopefully, manufacturers can't just make this up. The EU's website isn't very clear on how icon 3 differs from icon 5, though, as the text explanation for number 5 applies better to number 3 (and even has the same icon). Further, icon 4 indicates the device's repeated free-fall reliability class. Think of it as the smartphone or tablet's impact resistance, determined based on the number of falls in a standardized test that the device can withstand without functionality problems. A is the most robust, and E is the least robust. Icon 6 denotes the smartphone or tablet's repairability class, where A is the most repairable and E is the least. Finally, icon 7 represents the device's IP rating for dust and water resistance. These energy labels will help EU consumers make more informed purchase decisions. They will also push companies to be more transparent about their products, hopefully prompting the entire ecosystem to move towards more sustainable choices. How do current phones compare in the EU energy label? Here are the energy labels for the Apple iPhone 16 Pro Max, Samsung Galaxy S25 Ultra, Google Pixel 9 Pro XL, and the OnePlus 13, for reference: Apple iPhone 16 Pro Max Samsung Galaxy S25 Ultra Google Pixel 9 Pro XL OnePlus 13 Comparing the labels, we can see that the OnePlus 13 scores high on energy efficiency and battery life, decently for repairability, but poorly on impact resistance. On the other hand, the Galaxy S25 Ultra is the worst amongst these four for battery life, and it has just okay repairability but excellent impact resistance. Knowing this information beforehand will most definitely impact purchase decisions. Curiously, this energy label also helps us figure out the rated battery capacity of the iPhones, which are as follows for the current lineup: iPhone 15: 3,349mAh iPhone 15 Plus: 4,383mAh iPhone 16: 3,349mAh iPhone 16 Plus: 4,674mAh iPhone 16 Pro: 3,582mAh iPhone 16 Pro Max: 4,685mAh iPhone 16e: 4,005mAh EU's ecodesign requirements also go into effect today Alongside the energy labels, the EU's ecodesign requirements will also apply to smartphones, feature phones, cordless phones, and slate tablets (which include the usual Android tablets). However, they won't apply to tablet computers (tablets that run Windows, for example), products with flexible/roll-up main displays, and high-security communication smartphones. The ecodesign requirements include: Resistance to accidental drops or scratches and protection from dust and water. Sufficiently durable batteries which can withstand at least 800 charge and discharge cycles while retaining at least 80% of their initial capacity. Rules on disassembly and repair, including obligations for producers to make critical spare parts available within 5-10 working days, and for 7 years after the end of sales of the product model on the EU market. Availability of operating system upgrades for longer periods (at least 5 years from the date of the end of placement on the market of the last unit of a product model). Non-discriminatory access for professional repairers to any software or firmware needed for the replacement. These ecodesign requirements should usher in a new era of phones and tablets that are built to last. Coupled with the energy labels, we hope to see more and better smartphones in the market that don't force planned obsolescence upon us. Got a tip? Talk to us! Email our staff at Email our staff at news@ . You can stay anonymous or get credit for the info, it's your choice.