#12 of 30 · Open-weight lab
Meta (Llama / Meta AI)
Model cards disclose training emissions; Hyperion will be powered by new gas plants.
In plain words
How scores workMeta scores 56 out of 100 — middle of the pack. Some disclosure or clean-energy commitments, offset by rising emissions or blind spots. Confidence in this score is good confidence (B).
What helps
- Efficiency · Open weights (Llama 3.x, Llama 4) can be benchmarked directly; Llama 4 uses mixture-of-experts with 17B active parameters.
- Carbon · Matched 100% of electricity with clean and renewable energy every year since 2020.
- Water · Restored over 1.6 billion gallons in high- and medium-stress regions in 2024, toward restoring 200% of consumption in high-stress basins.
What hurts
- Efficiency · No per-prompt figure for Meta AI, which serves hundreds of millions of users.
- Carbon · Louisiana regulators approved three Entergy gas plants (≈2.2 GW) and a $550M transmission line to power Hyperion; Meta subsidises costs for 15 years, after which they fall to ratepayers. The campus has since been expanded to 5 GW.
- Transparency · Model cards report market-based training emissions as zero, resting on renewable-certificate accounting that the cards do not describe.
What nobody knows
- Per-prompt energy disclosed
- Per-prompt water disclosed
- Hardware reuse or e-waste reported
Meta is the only large lab that prints training emissions in its model cards (Llama 4 Scout 1,354 t, Maverick 645 t location-based) and its open weights let anyone measure inference. It used 18.4 TWh of electricity in 2024 and planned to end 2025 with over 1.3 million GPUs. Its footprint is 8.2 MtCO₂e market-based but 15.6 Mt location-based, and the Hyperion campus in Louisiana — now sized at 5 GW on 3,650 acres — will draw on three new Entergy gas plants approved in August 2025.
Key figures
Per-prompt and fleet metrics. Hover a title for what it means; open “How we got this number” for the working.
ecogpt.co estimate; no company figure.
How we got this number
Llama 4 activates 17B of its parameters per token (mixture of experts), roughly half the work of a dense frontier model that Epoch puts at ≈0.3 Wh. Meta AI serves hundreds of millions of users on this family. Uncertainty roughly ±3×.
On-site cooling only, so comparable with Google and OpenAI's figures.
How we got this number
0.15 Wh per prompt × 0.2 L/kWh (Meta's fleet WUE as reported by secondary sources). Counting the water evaporated to generate the electricity adds ≈0.27 mL.
Location-based: what the grid actually emitted. Renewable certificates would bring the market-based figure near zero.
How we got this number
0.15 Wh × 390 gCO₂/kWh (US grid, location-based).
15.6 Mt location-based.
How we got this number
Meta does not print a year-on-year change for emissions. Data-centre electricity rose from 14.98 TWh in 2023 to 18.06 TWh in 2024 (+20%), which is the best proxy for the location-based trend.
How we got this number
Fleet figure repeated by Koomey Analytics and others from Meta's reporting; not printed in the 2025 report itself.
Hardware & materials
The chips, power, copper and future e-waste behind the fleet — the part sustainability reports leave out
GPUs, TPUs or custom accelerators the company runs. Analyst figures are in H100-equivalents, company figures in physical chips.
How we got this number
Meta's own target was more than 1.3 million physical GPUs by end-2025, with ~1 GW of new capacity in the year.
Megawatts the AI fleet pulls. One megawatt runs roughly 750 US homes.
How we got this number
18.4 TWh a year ÷ 8,760 hours. Average draw implied by annual electricity; peak capacity is higher.
Terawatt-hours per year. One TWh is a year of electricity for about 95,000 US homes.
How we got this number
18.06 TWh of it in data centres; up 20% in a year.
CO₂e emitted making the chips and servers before they answer a single prompt.
How we got this number
2,250,000 accelerators × 514 kg CO₂e each (chip plus share of the host server). Emitted in manufacturing, before the first prompt is served.
Tonnes of server hardware that will be retired in three to five years.
How we got this number
2,250,000 accelerators × 17 kg of server hardware each. Tomorrow's e-waste unless reused; generative AI could produce 2.5 Mt of it a year by 2030.
Tonnes of copper in cabling, busbars and transformers.
How we got this number
2,100 MW × 27 t of copper per MW (cabling, busbars, transformers).
Diesel backup units or gas turbines at the data centres, permitted or observed.
How we got this number
2,100 MW ÷ 2.6 MW per generator, the average size permitted at US data centres. Permit counts are not published fleet-wide; this is the number a fleet this size normally carries.
Share of retired servers and parts given a second life.
Kilograms of antimony-equivalent, the life-cycle unit for scarce minerals consumed.
Score breakdown
Each pillar is 0–100; weights are fixed across the index
- Open weights (Llama 3.x, Llama 4) can be benchmarked directly; Llama 4 uses mixture-of-experts with 17B active parameters. Meta (model card)Hugging Face, Salesforce, Cohere, Carnegie Mellon
- No per-prompt figure for Meta AI, which serves hundreds of millions of users. Meta
- 2024 emissions: 8.2 MtCO₂e market-based versus 15.6 Mt location-based — the gap is covered by renewable certificates and PPAs. Electricity grew to 18.4 TWh, about 20% in a year. Meta
- Louisiana regulators approved three Entergy gas plants (≈2.2 GW) and a $550M transmission line to power Hyperion; Meta subsidises costs for 15 years, after which they fall to ratepayers. The campus has since been expanded to 5 GW. Louisiana IlluminatorStraight Arrow NewsData Center Dynamics
- Matched 100% of electricity with clean and renewable energy every year since 2020. Meta
- Training GPU-hours and location-based emissions published per model; annual report plus an environmental data index with location-based Scope 2 and electricity by year. Meta (model card)Meta
- Model cards report market-based training emissions as zero, resting on renewable-certificate accounting that the cards do not describe. Meta (model card)
- Net-zero value chain by 2030; Scope 1+2 down 42% by 2031 from 2021; water positive by 2030. Meta
- Pushed the Louisiana Public Service Commission to vote on the Hyperion gas plants before the administrative judge's recommendation; approved 4–1 over ratepayer objections. Louisiana Illuminator
- Bought a further 1,400 acres at Hyperion in late 2025 and raised the campus target from 2 GW to 5 GW. Data Center Dynamics
Models
What this company ships, and what each prompt costs
| Model | Weights | Energy / prompt | Water | Training CO₂e |
|---|---|---|---|---|
| Llama 4 Scout 109B total / 17B active (MoE) | Open | 0.10 Wh Estimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. | 0.020 mL Estimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. | 1,354 t DisclosedCompany figure. Published by the company itself. We take it as given and record the scope next to it. |
| Llama 4 Maverick 400B total / 17B active (MoE) | Open | 0.15 Wh Estimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. | 0.030 mL Estimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. | 645 t DisclosedCompany figure. Published by the company itself. We take it as given and record the scope next to it. |
| Llama 3.1 family (8B / 70B / 405B) 8B–405B dense | Open | 0.25 Wh Estimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. | 0.050 mL Estimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. | 11,390 t DisclosedCompany figure. Published by the company itself. We take it as given and record the scope next to it. |
Where the compute lives
Named sites from company disclosures and press reporting
- PrinevilleOregon, USA
- New Albany (Prometheus)Ohio, USA · First gigawatt-scale AI cluster.
- Richland Parish (Hyperion)Louisiana, USA · 5 GW on ≈3,650 acres; three new gas plants approved.
- LuleåSweden
- CloneeIreland
Timeline
Incidents and developments involving this company
Louisiana approves three new gas plants to power Meta's Hyperion
The Louisiana Public Service Commission voted 4–1 to let Entergy build three gas plants (≈2.2 GW) and a $550M transmission line for Meta's Richland Parish AI campus. Meta covers costs for 15 years; afterwards they fall to ratepayers.
MetaAI Energy Score launches star ratings for 166 models across 10 tasks
Hugging Face, Salesforce, Cohere and Carnegie Mellon publish an Energy Star-style leaderboard: GPU watt-hours per task, with five relative star bands. Open models are tested automatically; closed models through a sealed sandbox.
Salesforce, Hugging Face, Cohere, Meta, Mistral, Alibaba, DeepSeek, IBM