Altman recently shared a specific character for querying the amount of energy and water used. According to his blog post, each query of CHATGPT consumes 0.34 WH of electricity (0.00034 kWh) and about 0.000085 gallons of water. The equivalent of efficient light bulbs to use in a few minutes, about one-tenth of a teaspoon.
This is the first time Openai has shared such data publicly, adding an important data point to the ongoing debate on the environmental impact of large AI systems. The announcement sparked widespread discussion – both supportive and skeptical. In this post, I analyse the claim and unravel the reactions on social media to see the arguments of both sides.
What supports the 0.34 WH claim?
Let’s take a look at the arguments about the credibility of Openai numbers.
1. Independent estimates are consistent with Openai’s number
The key reason some people think that the number is trustworthy is that it closely matches previous third-party estimates. In 2025, the Institute Times estimated that a single query of GPT-4O consumed about 0.0003 kWh of energy – closely matched Openai’s own estimates. Suppose GPT-4O uses an Experts architecture with 100 billion active parameters, with a typical response length of 500 tokens. However, they do not take into account other factors other than energy consumption of GPU servers and do not include power usage efficiency (PUE) as usual.
A recent academic study by Jehham et al. (2025) estimated that GPT-4.1 NANO uses 0.000454 kWh, O3 uses 0.0039 kWh, and GPT-4.5 uses 0.030 kWh in long-term cues (approximately 7,000 input words and 1,000 output words).
The consistency between the estimates and OpenAI data points shows that OpenAI numbers fall within a reasonable range, at least only when focusing on the phase of the model’s response prompt (called “inference”).
2. On the hardware level, OpenAI’s numbers may be reasonable
It is reported that OpenAI servers have 1 billion queries per day. Let’s consider how Chatgpt provides the math behind these queries every day. According to industry experts, if this is true and the energy per query is 0.34 WH, the total daily energy could be about 340 MWh. He speculated that this would mean that OpenAI could use about 3200 servers (assuming NVIDIA DGX A100) to support Chatgpt. If 3200 servers have to process 1 billion queries per day, about 4.5 prompts must be processed per second. If we assume that each server deploys an instance of Chatgpt base LLM on each server, and the average prompt results in 500 output tokens (about 375 words according to Openai’s rule of thumb), the server needs to generate 2,250 tokens per second. Is that realistic?
Stojkovic et al. (2024) were able to obtain throughput of 6,000 tokens per second from Llama-2–70B on an NVIDIA DGX H100 server with an 8 h100 GPU.
However, Jegham et al. (2025) found that on average, three different OpenAI models were generated per second. However, it is not clear how they arrived.
So it seems we can’t reject the idea that 3200 servers can handle 1 billion queries per day.
Why some experts are skeptical
Despite the evidence, many people remain cautious or critical of the 0.34 WH figure, which raises some key issues. Let’s look at those.
1. OpenAI’s number may ignore the main parts of the system
I suspect that the number only includes the energy used by the GPU server itself, and not other infrastructures such as data storage, cooling systems, network equipment, firewalls, power conversion losses or backup systems. This is a general restriction on energy reporting by technology companies.
For example, Meta has also reported on the energy count of GPUs only in the past. But in real-world data centers, GPU power is only part of the complete picture.
2. Server estimates appear to be low compared to industry reports
Some commentators, such as GreenOps Advocate Mark Butcher, think that the 3,200 GPU servers seem too low to support all ChatGpt users, especially when you consider global usage, high availability, and other applications other than casual chats (such as encoding or image analysis).
Other reports suggest OpenAI uses dozens or even hundreds of thousands of GPUs for reasoning. If this is true, then the total energy use may be much higher than what the 0.34 WH/query number suggests.
3. Lack of details can cause problems
Critics, such as David Mytton, also pointed out that Openai’s statement lacks basic background. For example:
- What is an “average” query? A question, or a complete conversation?
- This number applies only to one model (e.g., GPT-3.5, GPT-4O) or several models?
- Does it include updated, more complex tasks such as multi-modal input (e.g., analyzing PDFs or generating images)?
- Is the amount of water used directly (for cooling servers) or indirectly (from power sources such as water and electricity)?
- What about carbon emissions? This depends to a lot on the position and energy mixing.
Without answers to these questions, it is difficult to know how much trust is placed in the numbers or how to compare it to other AI systems.
Viewpoint
Have big technology finally heard our prayers?
NVIDIA released data on GPU embodied, as well as Google’s blog post about the lifecycle emissions of its TPU hardware, and released Openai’s disclosure. This may indicate that the company has finally responded to the call for more transparency. Have we witnessed the dawn of a new era? Or is Sam Altman just playing tricks for us because his economic interests are in his economic interests to downplay the climate impact of his company? I will use this question as a thought experiment for readers.
Reasoning and training
Historically, the numbers we have seen about estimates and reports on AI energy consumption are related to energy use train AI model. Although the training phase may be very energy-conscious, over time, providing billions of queries (inferences) can actually use more total energy than training the model first. My own estimates suggest that training GPT-4 may use about 50-600 million kWh of electricity. Each query is 0.34 WH and 1 billion queries are queried daily, and the energy used to answer user queries will exceed the energy use of the training stage after 150-200 days. This makes the idea that reasoning energy is worth carefully measured with credibility.
Conclusion: A popular first step, but stay away from the full picture
Just as we think the debate on Openai’s energy use has grown old, the infamous shutdowns have also caused a sensation with the disclosure of this number. Many are excited that Openai has now published a debate about the energy and water use of its products and hope this is the first step toward greater transparency, the Lasso draw and climate impact on large technologies. On the other hand, many people are skeptical about Openai’s figure. And there are good reasons. It is disclosed in a blog post as a blog post about a completely different topic and does not give any context in any of the cases detailed above.
Even if we may be witnessing the shift to more transparency, we still need a lot of OpenAI information to be able to critically evaluate its 0.34 WH number. Before this, it should be taken not only with a grain of salt, but also with a small amount.
That’s it! Hope you enjoyed this story. Let me know what you think!
Follow me to learn more about AI and sustainability and stay tuned for me on LinkedIn.
