Bigger models, dirtier planet

Neural scaling laws describe how model performance improves predictably as parameters, data and compute increase, following a power-law relationship first formalised by Kaplan et al. at OpenAI in 2020^[1]Kaplan et al., OpenAIView source → and refined by Hoffmann et al. in the Chinchilla paper^[2]Hoffmann et al., DeepMindView source →. Chinchilla’s key finding, that compute-optimal training requires scaling parameters and training data together rather than just adding more parameters, corrected the prevailing over-parameterised approach and drove a wave of leaner, more efficient models.

What neither paper describes is the corresponding carbon scaling law. The relationship between model size and energy is not linear, and it is not even strictly proportional to compute. Moving from 7 billion to 70 billion parameters increases per-token inference energy by roughly 100 times^[3]Husom et al. / Caravaca et al.View source →. At the upper end of the open-source spectrum, Llama-3.1-405B requires about 1.861 watt-hours per response, against 0.032 watt-hours for an 8-billion-parameter model^[4]Wikipedia / University studyView source → producing the same number of output tokens, a 58-fold difference.

This super-linear scaling is structural. Larger models need more GPU memory, more inter-chip communication and more tensor parallelism, each added device bringing communication overhead. And generation is not free: the energy cost of autoregressive decoding, where every output token requires a full forward pass through the entire model, is weighted roughly four to five times higher than processing input tokens^[5]Seedling / Epoch AIView source →.

What mixture-of-experts does, and does not, fix

Mixture-of-experts (MoE) architectures activate only a subset of a model’s parameters for each token, routed by a gating network. The clearest public example is Mixtral 8×7B: 46.7 billion total parameters, with 12.9 billion active per token^[7]Mistral AI technical paperView source →. That gives it inference economics closer to a 13B dense model while keeping performance closer to the larger scale. Google has confirmed Gemini 1.5 Pro is a sparse MoE architecture^[8]Google / Gemini 1.5 Pro reportView source →that is cheaper to serve than Gemini 1.0 Ultra, though it has not disclosed the parameter counts. MoE is a core part of Google’s 33-fold per-prompt efficiency improvement between 2024 and 2025^[9]Google Cloud BlogView source →.

The gain is real, but it comes with nuances the broader narrative tends to omit. A 2025 study across Qwen3 variants found a 30B MoE model with 3B active parameters consumed energy 3.56 times lower per token than a comparable dense 32B model^[10]arXiv 2601.22076View source →, confirming active parameter count as the primary inference energy driver. Yet separate research on OLMoE found that MoE models can use up to 54 per cent more energy than a dense model with the same active parameters^[11]arXiv 2504.17674View source →, because the fused kernel operations in expert layers run substantially slower than equivalent dense matrix multiplications, 19 per cent slower at batch size 1 and 63 per cent slower at batch size 8.

MoE bends the curve. It does not break it.

The net position: a 235B MoE model with 22B active parameters still consumes substantially more energy than a dense 8B model, because routing overhead, memory bandwidth costs and communication patterns add their own load. MoE changes the energy-per-parameter relationship; it does not eliminate the scaling penalty.

Training costs: what has been measured

Training is a one-time event, but its carbon cost sets the floor and grows faster than linear with model size. The clearest data point remains Patterson et al. (2021): GPT-3 consumed roughly 1,287 MWh and emitted around 552 tonnes of CO₂ equivalent^[12]Patterson et al. (2021)View source → in Scope 1 and 2 operational emissions, equivalent to driving about 110 petrol cars for a year, and that figure excludes hardware manufacturing entirely.

BLOOM, at 176 billion parameters, produced just 24.7 tonnes CO₂ equivalent operationally^[13]Luccioni et al. (2023)View source →. The 20-fold difference from GPT-3 comes almost entirely from grid carbon intensity: BLOOM was trained on a French nuclear-powered facility at about 57 g CO₂e per kWh, versus an estimated 429 g for GPT-3’s US location. With embodied hardware included, BLOOM’s total rose to 50.5 tonnes, still ten times lower than GPT-3’s operational figure alone. For GPT-4, a lifecycle analysis from leaked configuration data, roughly 25,000 A100 GPUs over 90 to 100 days and over 50,000 MWh^[14]LudvigsenView source →, yields about 10,200 tonnes operationally and approximately 15,000 tonnes including embodied hardware^[15]Blubrom et al. (2025)View source →, the annual emissions of roughly 938 average Americans. In July 2025, Mistral published the first peer-reviewed full lifecycle assessment of a production model, placing Mistral Large 2 at 20,400 tonnes CO₂ equivalent^[16]Mistral AI / Carbone 4 / ADEMEView source → with independent ADEME audit.

The model size decision is an environmental decision

Mistral’s LCA was explicit about what the measurements imply for model selection: a model ten times larger generates impacts one order of magnitude larger for the same number of generated tokens^[16]Mistral AI / Carbone 4 / ADEMEView source →. At the individual query level, the difference between an 8B and a 70B model is about 1.67 watt-hours, roughly equivalent to leaving an LED bulb on for four minutes. At 2.5 billion queries per day^[6]AI Energy Score BenchmarkView source →, that difference compounds to hundreds of thousands of tonnes of CO₂ equivalent annually, more than the yearly emissions of a mid-sized city.

Efficiency is not the same as sufficiency

The efficiency counter-argument is real. Chinchilla showed prior frontier models were over-parameterised, and MoE has since partly decoupled parameter count from active compute. But the gains have not outpaced ambition. Epoch AI estimated compute in the largest training runs was doubling roughly every six months through 2025^[17]Epoch AIView source →, far faster than hardware efficiency. The buildout locks that in: $500 billion committed to US data centres through Stargate, $75 billion from Google in 2025 alone^[18]CNaught / Google 2025 ReportView source →, decade-long capital commitments. And Google’s own total emissions rose 51 per cent between 2019 and 2025 even as per-prompt energy fell, the efficiency gains consumed by growth in volume and scale.

The capability argument for frontier scale is weakening too. Llama-3.3-70B outperforms GPT-4-Turbo on standard benchmarks^[19]MarmelabView source → despite a roughly 25-fold reduction in parameter count, evidence that sufficiency, not scale, is the right frame for many production use cases. Researchers have proposed treating carbon as a first-class optimisation metric: the CarbonScaling framework calculates accuracy improvement per unit of additional carbon^[20]CarbonScaling (arXiv, 2025)View source →, finding the point where further scaling produces diminishing returns in emissions efficiency as well as capability.

The energy cost of a query is not just a technical parameter, it is a procurement decision. When an organisation runs frontier-scale inference for a task a 7B model handles adequately, it chooses to emit roughly 50 times more carbon per response. Sufficiency, deploying the smallest model adequate for the task, is the lever scaling efficiency alone cannot pull. It requires no new hardware and no architectural innovation, only that the question does this task require a frontier model? be asked before the API call is made.