Reference · Sustainability

The carbon twin of $/M-token

Every token has an energy shadow. This is a modeled estimate of the electricity (kWh) and emissions (gCO₂) per million output tokens for AI models with disclosed active-parameter counts, normalized on a reference accelerator. Like our sizing and teardowns: modeled, assumptions published — best for relative comparison, not a billed figure.

Grid:Global avg · 400 gCO₂/kWh United States · 370 gCO₂/kWh EU · 250 gCO₂/kWh Low-carbon grid · 40 gCO₂/kWh

Assumptions: reference accelerator NVIDIA Rubin Ultra VR300 (32,000 GB/s HBM, 3600 W) · memory-bound decode at 50% MBU · batch 32 · 2-byte weights (FP16) · PUE 1.2 · grid 400 gCO₂/kWh (Global avg).

Model	Vendor	Active params	kWh / Mtok	gCO₂ / Mtok	$ / Mtok
Gemma 3 1B	Google	1B	0.004	1.9	—
Llama 3.2 1B	Meta	1B	0.004	1.9	—
Gemma 2 2B	Google	2B	0.008	3.8	—
Qwen 3 30B (MoE)	Alibaba	3B	0.012	5.6	$0.50
Qwen2.5 3B	Alibaba	3B	0.012	5.6	—
Llama 3.2 3B	Meta	3B	0.012	5.6	—
Gemma 3 4B	Google	4B	0.016	7.5	—
Mistral 7B	Mistral	7B	0.029	13.7	—
LLaMA 3 8B	Meta	8B	0.031	15.0	—
Llama 3.1 8B	Meta	8B	0.031	15.0	—
Gemma 3 12B	Google	12B	0.047	22.5	—
LLaMA 4 Scout	Meta	17B	0.066	31.9	$0.30
LLaMA 4 Maverick	Meta	17B	0.066	31.9	$0.80
o4-mini	OpenAI	20B	0.078	37.5	$4.40
Qwen 3 235B (MoE)	Alibaba	22B	0.086	41.3	$1.82
Gemma 3 27B	Google	27B	0.105	50.6	$0.33
DeepSeek R1	DeepSeek	37B	0.145	69.4	$2.50
DeepSeek V3	DeepSeek	37B	0.145	69.4	$0.80
DeepSeek-V3 (MoE)	DeepSeek	37B	0.145	69.4	—
Mixtral 8x22B (MoE)	Mistral	39B	0.152	73.1	—
Mixtral 8x22B	Mistral	39B	0.152	73.1	$6
GPT-4o	OpenAI	50B	0.195	93.8	$10
o3	OpenAI	50B	0.195	93.8	$8
Gemini 2.5 Flash	Google	60B	0.234	112	$2.50
LLaMA 3 70B	Meta	70B	0.273	131	—
Claude Haiku 3.5	Anthropic	70B	0.273	131	$4
Llama 3.1 70B	Meta	70B	0.273	131	—
Qwen 2.5 72B	Alibaba	72B	0.281	135	$0.40
Grok-2	xAI	80B	0.313	150	$10
Command R+	Cohere	104B	0.406	195	$10
Claude Sonnet 4	Anthropic	120B	0.469	225	$15
Mistral Large 2	Mistral	123B	0.480	231	$1.50
Gemini 2.5 Pro	Google	300B	1.17	563	$10
GPT-4.5	OpenAI	300B	1.17	563	$15
Grok-3	xAI	300B	1.17	563	$15
Claude Opus 4	Anthropic	350B	1.37	656	$75
LLaMA 3.1 405B	Meta	405B	1.58	759	—
Llama 3.1 405B	Meta	405B	1.58	759	—
Kimi K3	Moonshot	2800B	10.94	5,250	$15

Sorted most-efficient first. Only models with a disclosed active-parameter count are shown (39 of 46); most closed models don't publish parameters.

How it's modeled

Decode is memory-bandwidth-bound: each output token reads the model's active parameters from HBM once, amortized across the serving batch. So energy/token ≈ chip power ÷ (per-stream throughput × batch), where per-stream throughput = (HBM bandwidth × MBU) ÷ (active params × bytes/param). Emissions = energy × PUE × grid carbon intensity. Real serving efficiency varies widely with batch size, precision, hardware and grid — treat these as directional. Terms: MBU, PUE, grid carbon intensity, gCO₂ per token.

Model your workload's footprint precisely

For your actual model, chip, batch and region, the sizing report computes energy, cost and $/M-token together.

Sizing report →Model catalog →

Modeled estimate; methodology above. Not investment or sustainability-reporting advice. Methodology.