Our approach

AI and Environmental Impact: What Bells Up Can—and Cannot—Measure

A plain-language review of energy data for the AI models available on Bells Up.

For attorneys and the technology advisors who evaluate tools on their behalf.

Purpose and scope

This paper describes what is and is not knowable about the environmental cost of the AI models Bells Up serves. It summarizes the public information Bells Up reviewed as of August 11, 2026. It identifies what is measured, what is estimated, and what providers do not disclose. It is not a certified life-cycle assessment (LCA) or a regulatory environmental disclosure. Nothing in this paper should be interpreted as a claim that Bells Up AI is greener than another legal AI product.

This paper is a companion to Bells Up’s statement, Why Bells Up AI Chose to Offer Models Some Lawyers Will Refuse to Use.

At a glance

AI has a real environmental cost. Training a model requires electricity, water, and specialized hardware. Serving the model requires more electricity and water each time someone submits a prompt. The impact of one query may be small, but it repeats across every user and every task.

Accurate per-model comparisons remain difficult. The providers do not use a common measurement method, and most do not publish current model-level figures at all. Provider studies use their own definitions of a typical prompt or response. The best-known cross-provider study uses fixed test prompts and estimates the hardware and serving conditions behind commercial APIs. Those numbers answer different questions.

Bells Up can disclose the serving path a query takes. Five of our ten model families—Anthropic’s Claude, Meta’s Llama, Mistral, DeepSeek, and Alibaba’s Qwen—are served through Amazon Bedrock. OpenAI, Google, and xAI are reached through their own APIs. Moonshot’s Kimi and Zhipu’s GLM are served through Fireworks, with their older models through Amazon Bedrock. The serving path affects the electricity used for inference. It does not change the energy already used to train a model, and it does not supply model-level figures that the providers have chosen not to disclose.

Our approach is to state what each source actually supports. We identify provider measurements, outside estimates, and unknowns separately. We do not combine them into a ranked leaderboard.

If environmental impact is your priority

Use the smallest, least reasoning-intensive model that can complete the task reliably. For routine work, start with the smallest suitable tier. Current estimates suggest that Google Gemini, Meta Llama 4, and small Qwen models are lower-compute candidates, but those estimates are not current, tier-specific, or quality-adjusted measurements. If you are using Anthropic Claude, choose Claude Haiku when it can handle the task; move to Sonnet or Opus when the task requires more capability. If you are using OpenAI GPT, choose Luna when it can handle the task, and move to Terra, where offered, or Sol when more capability is warranted.

Reserve high-reasoning or always-reasoning models for work that benefits from their additional reasoning, because high-reasoning models can use substantially more energy per query. Also remember, a model that produces an unusable answer may require additional queries, so a lower per-query estimate does not necessarily produce a lower footprint for the work as a whole.

This is a practical selection rule, not a precise cross-family ranking. Smaller models generally require less computation within the same family, but architecture, active parameters, input and output tokens, reasoning mode, serving hardware, batching, and grid conditions all affect the result. A lower-tier model in one family may use more energy than a higher-tier model in another. The trade-off may involve capability, reasoning depth, speed, reliability, or answer quality for a particular task.

Before comparing the numbers

The figures in this paper use different units of work:

  • One provider study reports a median prompt without a fixed token count.
  • Another reports a 400-token response.
  • Jegham et al.’s modeled research estimates use fixed workloads. The “medium” workload contains 1,000 input tokens and 1,000 output tokens.

Energy and carbon are also different measures. Watt-hours describe electricity use. Carbon depends on the amount of electricity, the grid that supplied it, and whether the calculation uses location-based or market-based accounting. A low-energy model can run on a relatively carbon-intensive grid. A market-based carbon figure can reflect renewable-energy purchases that do not change the physical grid at the time of use.

Where both forms are available, this paper uses location-based figures to describe the physical grid where electricity is consumed. That choice can produce a worse-looking result than market-based accounting, which credits contractual purchases such as renewable-energy certificates. Those purchases may support clean generation, but they do not establish what powered a server at the time and place of use.

These differences explain why two published figures may look inconsistent even when the arithmetic is sound.

1. Training and inference

Training is the cost of building a model. It involves specialized computer chips operating continuously, often for months, on very large datasets. This cost occurs before any customers use a model.

Inference is the cost of using a trained model. Each prompt and response requires a pass through the model on serving hardware. The energy cost is smaller than the total energy used for training, but it recurs with every query and grows with product usage.

Training is an important absolute cost. It may be allocated across all queries served during a model’s useful life, but providers generally do not disclose the inputs needed to make that allocation: training energy, the number of lifetime queries, and the effect of retraining or replacement. Any per-query allocation would therefore depend on assumptions Bells Up cannot verify.

This paper gives more attention to inference because it is the recurring cost associated with a customer’s use and because model selection, reasoning depth, and serving path can change it. That focus does not remove training from the model’s total footprint.

Hardware adds another layer. Manufacturing accelerators and servers requires materials, electricity, and water. A life-cycle assessment may allocate some of that upstream cost to each response. A life-cycle figure therefore cannot be treated as though it measured serving electricity alone.

2. What providers disclose

Google and Mistral have published the most detailed provider disclosures in the current Bells Up provider roster, although neither disclosure directly corresponds to the models available on Bells Up.

Google’s August 2025 study reports a production measurement for the median Gemini Apps text prompt in May 2025. Its full-stack method includes active accelerators, host CPU and memory, idle capacity, and datacenter overhead. Google reports 0.24 watt-hours (Wh), 0.03 grams of CO2-equivalent, and 0.26 milliliters of water per median prompt. The 0.03-gram figure uses market-based carbon accounting. The result is fleet-wide rather than model-specific, and does not necessarily apply to inference run on current-generation Google Gemini models.

In July 2025, Mistral published a detailed LCA for its Le Chat product, but these numbers are inapplicable to the Mistral-family models available on Bells Up because we run Mistral inference through Amazon Bedrock, not Le Chat. Mistral’s LCA reports 1.14 grams of CO2-equivalent and 45 milliliters of water for a 400-token Le Chat response using Mistral Large 2, and this number also includes upstream impacts such as server manufacturing. Mistral’s public report does not separate the portion attributable to serving electricity, so the data cannot support recalculating the total with an AWS grid factor in place of a French one.

For the remaining families, current model-level data is sparse or absent. Meta publishes Llama training information in its model cards, but does not disclose current production inference measurements for Llama 4. Anthropic does not publish per-query energy for the Claude generation Bells Up serves. OpenAI does not publish a methodologically complete figure for GPT-5.6. While Alibaba publishes Qwen architecture information, it does not provide any current per-query measurement. And while DeepSeek publishes unusually detailed training reports, it does not publish a Bedrock inference measurement. xAI does not publish current model-level energy for Grok.

Researchers have warned that declining provider transparency creates a second problem: demand for simple environmental figures can cause incomplete or decontextualized estimates to circulate as facts. Luccioni et al. describe this as “misinformation by omission.”

Outside researchers can measure open-weight models on known hardware. Closed models are harder. Jegham et al., How Hungry is AI?, provides modeled research estimates based on public API performance, inferred hardware, regional infrastructure data, and an assumed server-side batch size of eight. The authors’ sensitivity analysis shows that changing the batch-size assumption can change an inference-energy estimate by roughly 4 to 20 times. Its numbers are not readings from the providers’ electricity meters.

3. What the available evidence shows

The table below summarizes the strongest public information we found.

Family Best available evidence What the evidence supports What remains unknown for Bells Up
Google Gemini Google’s production study reports 0.24 Wh for a median Gemini Apps text prompt. Google’s median prompt used relatively little energy in May 2025. Current tier-level energy and the location serving a particular Bells Up request.
Meta Llama Model cards disclose architecture and some training information; inference figures for Llama 4 are modeled. Smaller active-parameter designs may reduce computation. Current per-query energy on Bedrock.
Alibaba Qwen Architecture and active-parameter information are public; inference figures are modeled. Small and low-active-parameter tiers are reasonable lower-compute candidates. Current per-query energy on Bedrock.
Anthropic Claude Jegham et al.’s modeled research estimate is about 2.8 Wh for Claude 3.7 Sonnet on its 1,000-input/1,000-output-token workload. The older Sonnet proxy falls below the study’s higher-energy estimates under that workload. Current Haiku, Sonnet 4.6, and Opus 4.6 energy.
Mistral Mistral’s LCA reports 1.14 gCO2e and 45 mL of water for a 400-token Le Chat response using Large 2. Mistral provides unusually broad life-cycle disclosure. The footprint of current Mistral models served through Bedrock.
OpenAI GPT Current GPT-5.6 tiers are unmeasured; available figures concern older models or undefined average prompts. Greater reasoning and longer outputs generally require more computation. Current Luna, Terra, and Sol energy and serving conditions.
DeepSeek Jegham et al.’s modeled research estimates are about 29 Wh for R1 and 9.1 Wh for V3 on its medium workload. R1 is a higher-energy estimate under the study’s workload and assumptions. Current R1 deployment on Bedrock and DeepSeek V3.2 energy.
xAI Grok No current provider measurement; outside figures are modeled. Current reasoning-model energy cannot be stated reliably. Per-query energy, water, and the location serving Bells Up’s API requests.

To show the scale without treating these estimates as measurements, 1,000 medium-workload queries at Jegham et al.’s modeled figures correspond arithmetically to approximately 2.8 kWh for the Claude 3.7 Sonnet proxy and 29 kWh for DeepSeek R1. As an order-of-magnitude reference, 29 kWh is roughly one day of electricity purchases for an average US household, using EIA’s 2022 national average. These are not measurements of current Bells Up models, and the study’s batch-size assumption alone can change an estimate by roughly 4 to 20 times.

4. Why Bells Up’s serving path matters

Because Bells Up uses Amazon Bedrock for five model families—Claude, Llama, Mistral, DeepSeek, and Qwen—we can identify AWS as the inference-service operator and identify the US regional context in which Bells Up submits those requests. Bells Up’s Bedrock endpoints are in Northern Virginia (us-east-1).

That endpoint does not establish the physical location that served a request or support assigning Virginia’s grid factor to it. Several Bells Up models may use AWS’s US cross-region inference profiles, and AWS may route those requests among supported US regions. AWS does not tell Bells Up which datacenter served an individual request, or disclose current per-model energy or hardware utilization. We can identify the serving path and possible regional context. We cannot state the electricity or emissions of a particular query.

Why the Colossus datacenters matter

This section addresses public concerns about Colossus as a provider- and facility-level issue. Except where the cited sources establish a specific connection, it does not identify the location or environmental footprint of a request sent through Bells Up.

For users concerned about environmental impact, SpaceX’s Colossus datacenters matter for a different reason. SpaceX’s June 3, 2026 SEC filing states that Colossus I and Colossus II were brought online almost entirely through on-site power generation and that Colossus II is expected to be powered primarily by a dedicated natural-gas plant. On-site gas generation may have climate implications from greenhouse-gas emissions and local environmental implications from other air pollutants. Those are related but distinct concerns. (SpaceX Form S-1/A, “Power Availability and Powered Shells.”)

Colossus can be described in unusual detail because it is addressed in public corporate filings and a detailed complaint. The absence of comparable detail elsewhere is not evidence of lower environmental impacts.

Colossus I and Claude

Anthropic announced on May 6, 2026, that it had agreed to use all compute capacity at SpaceX’s Colossus I facility: more than 300 megawatts and more than 220,000 NVIDIA GPUs. Anthropic said that the additional capacity would directly improve service for Claude Pro and Claude Max subscribers. Anthropic’s announcement does not identify which Claude requests run there or how the capacity is divided between inference, training, and other workloads.

A later SpaceX SEC filing described the agreement more broadly. It said Anthropic would have access to capacity across Colossus I and Colossus II, but does not identify which Claude workloads run at which facility. (SpaceX Form S-1/A, “Compute Services Agreements with Third Parties,” June 3, 2026.)

Bells Up’s Bedrock path is separate

Bells Up submits Claude requests through Amazon Bedrock. AWS describes Bedrock as AWS-operated and running on AWS-controlled infrastructure. Anthropic’s Colossus announcement does not state that Amazon Bedrock requests are served at Colossus. The public sources reviewed therefore do not establish that a Claude request sent through Bells Up is served there. For that reason, this paper does not attribute Colossus’s power source or facility-level impacts to that request.

Colossus II, gas turbines, and Grok

Colossus II raises a separate facility-level concern. In April 2026, the NAACP and Mississippi State Conference NAACP filed a complaint alleging that xAI violated the Clean Air Act in connection with a gas plant in Southaven, Mississippi. The complaint alleges that 27 gas turbines were installed and operated without the required air permit between August and December 2025; that the plant supplied power to the nearby Colossus II datacenter; and that the turbines had the potential to emit approximately 1,700 tons of nitrogen oxides per year, as well as formaldehyde and other pollutants. These allegations concern Colossus II, not Colossus I. (Complaint ¶¶ 4–8, 158, 205–215.)

SpaceX states that Colossus II is used to train Grok 5. The public record does not establish where xAI serves Bells Up’s Grok API requests. For that reason, this paper does not attribute Colossus II’s power source or facility-level impacts to Grok inference served through Bells Up. That uncertainty is not a reason to omit the facility-level issue. A buyer concerned about environmental impact may reasonably care about dedicated gas generation even when the available evidence cannot support a per-query emissions figure. (SpaceX Form S-1/A, “Compute Services Agreements with Third Parties,” June 3, 2026.)

5. Water

Google reports 0.26 milliliters of water for its median Gemini Apps prompt. Mistral reports 45 milliliters for a 400-token Le Chat response. The difference does not establish that one model uses 173 times more water than the other. The studies use different products, workloads, locations, and measurement boundaries.

The other providers do not publish current, model-specific water figures that can be applied to Bells Up queries. Facility-level water use may be important to the communities around a datacenter, but Bells Up cannot calculate a per-query number without operational data.

6. Conclusion

AI has a real environmental impact, and model choice affects the computation required for a task. The most practical step available to a Bells Up user is to match model capability and reasoning depth to the work, while treating published per-model figures as source-specific evidence rather than a precise cross-provider ranking. Bells Up can identify the serving path, explain what each source measured, and state plainly where the evidence ends.

7. Sources and further reading

  1. Google, “Measuring the Environmental Impact of Delivering AI at Google Scale” (August 2025): paper (opens in new tab).
  2. Mistral AI, “Our contribution to a global environmental standard for AI” (July 22, 2025): provider disclosure (opens in new tab).
  3. Jegham et al., “How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference” (arXiv:2505.09598, May 2025 and subsequent revisions): paper (opens in new tab).
  4. Amazon Web Services, Amazon Bedrock inference profiles: supported regions and models (opens in new tab).
  5. Amazon Web Services, Claude Platform on AWS and Amazon Bedrock: service comparison (opens in new tab).
  6. Anthropic, “Higher usage limits for Claude and a compute deal with SpaceX” (May 6, 2026): announcement (opens in new tab).
  7. Space Exploration Technologies Corp., Amendment No. 2 to Form S-1 (filed June 3, 2026): SEC filing (opens in new tab).
  8. NAACP and Mississippi State Conference NAACP v. X.AI Corp. et al., No. 3:26-cv-00074-MPM-JMV (N.D. Miss., filed April 14, 2026): complaint (opens in new tab).
  9. Luccioni, Gamazaychikov, Alves da Costa, and Strubell, “Misinformation by Omission: The Need for More Environmental Transparency in AI” (arXiv:2506.15572, June 2025): paper (opens in new tab).
  10. U.S. Energy Information Administration, “How much electricity does an American home use?” (2022 data): household-electricity data (opens in new tab).
  11. Hugging Face AI Energy Score and University of Michigan ML.ENERGY Leaderboard: AI Energy Score (opens in new tab) and ML.ENERGY (opens in new tab).