Altman recently shared a concrete figure for the energy and water consumption of ChatGPT queries. According to his blog post, each query to ChatGPT consumes about 0.34 Wh of electricity (0.00034 KWh) and about 0.000085 gallons of water. The equivalent to what a high-efficiency lightbulb uses in a couple of minutes and roughly one-fifteenth of a teaspoon.
This is the first time OpenAI has publicly shared such data, and it adds an important data point to ongoing debates about the environmental impact of large AI systems. The announcement sparked widespread discussion – both supportive and skeptical. In this post I analyze the claim and unpack reactions on social media to look at the arguments on both sides.
What Supports the 0.34 Wh Claim?
Let’s look at the arguments that lend credibility to OpenAI’s number.
1. Independent estimates align with OpenAI’s number
A key reason some consider the figure credible is that it aligns closely with previous third-party estimates. In 2025, research institute Epoch.AI estimated that a single query to GPT-4o consumes approximately 0.0003 KWh of energy – closely aligning with OpenAI’s own estimate. This assumes GPT-4o uses a mixture-of-experts architecture with 100 billion active parameters and a typical response length of 500 tokens. However, they do not account for other factors than the energy consumption by the GPU servers and they do not incorporate power usage effectiveness (PUE) as is otherwise customary.
A recent academic study by Jehham et al (2025) estimates that GPT-4.1 nano uses 0.000454 KWh, o3 uses 0.0039 KWh and GPT-4.5 uses 0.030 KWh for long prompts (approximately 7,000 words of input and 1,000 words of output).
The agreement between the estimates and OpenAI’s data point suggests that OpenAI’s figure falls within a reasonable range, at least when focusing only on the stage where the model responds to a prompt (called “inference”).
2. OpenAI’s number might be plausible at the hardware level
It’s been reported that OpenAI servers 1 billion queries per day. Let’s consider the math behind how ChatGPT could serve that amount of queries per day. If this is true, and the energy per query is 0.34 Wh, then the total daily energy could be around 340 megawatt-hours, according to an industry expert. He speculates that this would mean OpenAI could support ChatGPT with about 3,200 servers (assuming Nvidia DGX A100). If 3,200 servers have to handle 1 billion daily queries, then each server would have to handle around 4.5 prompts per second. If we assume one instance of ChatGPT’s underlying LLM is deployed on each server, and that the average prompt results in 500 output tokens (roughly 375 words, according to OpenAI’s rule of thumb), then the servers would need to generate 2,250 tokens per second. Is that realistic?
Stojkovic et al (2024) have been able to achieve a throughput of 6,000 tokens per second from Llama-2–70b on an Nvidia DGX H100 server with 8 H100 GPUs.
However, Jegham et al (2025) have found that three different OpenAI models generated between 75 and 200 tokens per second on average. It is, however, unclear how they arrived at this.
So it seems that we cannot reject the idea that 3,200 servers could be able to handle 1 billion daily queries.
Why some experts are skeptical
Despite the supporting evidence, many remain cautious or critical of the 0.34 Wh figure, raising several key concerns. Let’s take a look at those.
1. OpenAI’s number might leave out major parts of the system
I suspect the number only includes the energy used by the GPU servers themselves, and not the rest of the infrastructure – such as data storage, cooling systems, networking equipment, firewalls, electricity conversion loss, or backup systems. This is a common limitation in energy reporting across tech companies.
For instance, Meta has also reported GPU-only energy numbers in the past. But in real-world data centers, GPU power is only part of the full picture.
2. Server estimates seem low compared to industry reports
Some commentators, such as GreenOps advocate Mark Butcher, argue that 3,200 GPU servers seems far too low to support all of ChatGPT’s users, especially if you consider global usage, high availability, and other applications beyond casual chat (like coding or image analysis).
Other reports suggest that OpenAI uses tens or even hundreds of thousands of GPUs for inference. If that’s true, the total energy use could be much higher than what the 0.34 Wh/query number implies.
3. Lack of detail raises questions
Critics, eg David Mytton, also point out that OpenAI’s statement lacks basic context. For instance:
- What exactly is an “average” query? A single question, or a full conversation?
- Does this figure apply to just one model (e.g., GPT-3.5, GPT-4o) or an average across several?
- Does it include newer, more complex tasks like multimodal input (e.g., analyzing PDFs or generating images)?
- Is the water usage number direct (used for cooling servers) or indirect (from electricity sources like hydro power)?
- What about carbon emissions? That depends heavily on the location and energy mix.
Without answers to these questions, it’s hard to know how much trust to place in the number or how to compare it to other AI systems.
Perspectives
Are big tech finally hearing our prayers?
OpenAI’s disclosure comes in the wake of Nvidia’s release of data about the embodided emissions of the GPU’s, and Google’s blog post about the life cycle emissions of their TPU hardware. This could suggest that the corporations are finally responding to the many calls that have been made for more transparency. Are we witnessing the dawn of a new era? Or is Sam Altman just playing tricks on us because it is in his financial interests to downplay the climate impact of his company? I will leave that question as a thought experiment for the reader.
Inference vs training
Historically, the numbers that we have seen estimated and reported about AI’s energy consumption has related to the energy use of training AI models. And while the training stage can be very energy intensive, over time, serving billions of queries (inference) can actually use more total energy than training the model in the first place. My own estimates suggest that training GPT-4 may have used around 50–60 million KWh of electricity. With 0.34 Wh per query and 1 billion daily queries, the energy used to answer user queries would surpass the energy use of the training stage after 150-200 days. This lends credibility to the idea that inference energy is worth measuring closely.
Conclusion: A welcome first step, but far from the full picture
Just as we thought the debate about OpenAI’s energy use had gotten old, the notoriously closed company stirs it up with their disclosure of this figure. Many are excited about the fact that OpenAI has now entered the debate about the energy and water use of their products and hope that this is the first step towards greater transparency about the ressource draw and climate impact of big tech. On the other hand, many are skeptical of OpenAI’s figure. And for good reason. It was disclosed as a parenthesis in a blog post about an a wholly different topic, and no context was given whatsoever as detailed above.
Even though we might be witnessing a shift towards more transparency, we still need a lot of information from OpenAI in order to be able to critically assess their 0.34 Wh figure. Until then, it should be taken not just with a grain of salt, but with a handful.
That’s it! I hope you enjoyed the story. Let me know what you think!
Follow me for more on AI and sustainability and feel free to follow me on LinkedIn.
