Bloomberg报道复杂AI任务环境足迹显著放大但未形成明确公司级交易催化

As AI use shifts from simple, one-off queries to longer and more complex tasks, the added computing required results in a dramatically larger environmental footprint. The energy needed to build a web app using some models is equivalent to powering a home for 2 1/2 hours.

That’s according to the results of a new analysis from Vals AI, an independent AI benchmarking company that evaluates models on real-world tasks. The firm found that lengthier tasks that require AI models to “think” for longer, like building a software app, can have an environmental impact 10,000 times greater than simple queries, such as asking a basic question that can be answered almost immediately.

That suggests AI’s toll on the environment will grow considerably as usage shifts toward longer, agentic tasks that require more computational steps.

“Usage is transitioning from just using ChatGPT as a Google Search alternative to actually having conversations with these chatbots,” said Omar Almatov, a founding engineer at Vals. “Footprints scale dramatically because these are no longer just single-shot questions — they’re a lot more carbon intensive.”

Read More: AI Giants Are Quiet on Climate in Sign of Post-ESG Wall Street

The firm evaluated 16 models using its proprietary benchmarking index together with EcoLogits, an open-source tool that estimates AI’s environmental impact. It looked at three measures — carbon emissions, water consumption and electricity use — and employed baseline assumptions about factors including the hardware running each model, energy mix and data center efficiency.

Fourteen of the 16 models analyzed were developed by Chinese companies, with one from US-based Thinking Machines Lab and one from France’s Mistral AI. The researchers focused on open-weight models, whose training parameters are publicly available and can therefore be analyzed directly. Most leading US models, including those from OpenAI and Anthropic, are closed-weight, meaning their underlying parameters are private and can’t be evaluated using the same method.

American AI companies have disclosed a handful of estimates for the water and electricity consumed by a typical query — though what is typical is changing rapidly. OpenAI Chief Executive Officer Sam Altman wrote in 2025 that an average ChatGPT query uses about 1/15th of a teaspoon of water, and estimates by Altman and Alphabet Inc.’s Google the same year put the energy toll of the average text query between 0.24 and 0.34 watt hours. Companies have provided less information about the footprint of longer, agentic tasks. However, separate Microsoft Corp. research found that long-reasoning and agentic requests can increase energy consumption by more than an order of magnitude, consistent with Vals’ analysis.

Across the models the firm assessed, Alibaba Group Holding Ltd.’s most powerful AI model, Qwen3.8 Max, scored highest (that is, the worst) on energy, carbon and water intensity across the tasks Vals asked it to perform, while Ling 3.0 Flash 2607, a model from Ant Group designed for simple tasks, scored the lowest.

Another major finding from the Vals analysis: Small performance gains can come with huge environmental costs. Kimi K3, the most accurate open-weight model on the Vals Index, has an outsized footprint compared with the runner-up, DeepSeek V4 Flash, which is only marginally less accurate. The report authors describe the difference as the gap between charging your laptop once and 20 times, or between drinking a glass of water and flushing a toilet.

Read More: AI Is Draining Water From Areas That Need It Most

The companies whose models were analyzed did not immediately respond to requests for comment. Vals counts several leading AI firms as its customers, including some of these companies.

Amazon.com Inc. and Google have said that their greenhouse gas emissions rose overall in 2025, in part because of data center expansion. But more detailed disclosure about just how much carbon AI emits remains unknown, aside from outside estimates such as the one Vals conducted. Businesses that use AI tools lack a clear way to account for that in their own emissions reporting. Some AI firms are under pressure to disclose more because of a new California law on climate disclosure coming into effect.

“Policy conversations should be rooted in real evidence and data,” said Vals co-founder and CEO Rayan Krishnan, but AI companies reveal little about their model architecture, the types of hardware they use and the cooling and electrical infrastructure powering their data centers — information that would make it easier to measure their footprint, and do so more accurately.