#18 of 30 · Open-weight lab
DeepSeek
Radically cheap training, zero environmental reporting.
In plain words
How scores workDeepSeek scores 43 out of 100 — laggard. Little or no disclosure, fossil-backed capacity, or active disputes with regulators and communities. Confidence in this score is mostly estimates (C).
What helps
- Efficiency · Full training run of 2.788M H800 GPU-hours disclosed in the technical report; MoE activates 37B parameters per token.
What hurts
- Carbon · Training and first-party serving run on China's grid, with no renewable procurement disclosed.
- Water · No water data of any kind.
- Commitments · No climate or water commitments published.
What nobody knows
- Per-prompt energy disclosed
- Per-prompt water disclosed
- Annual environmental report
- Location-based Scope 2 reported
- Data-center locations public
- +3 more
DeepSeek-V3 was trained in 2.788 million H800 GPU-hours — about $5.6M at rental rates — and its mixture-of-experts design activates 37B of 671B parameters per token. SemiAnalysis puts the company's actual fleet at around 50,000 Hopper GPUs. The weights are open, so inference can be measured independently, but the company publishes no energy, water or emissions data and serves from the Chinese grid.
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
DeepSeek-V3 is a 671B mixture-of-experts model that activates 37B parameters per token — about two thirds of a dense frontier model's work at Epoch's 0.3 Wh. Uncertainty roughly ±3×.
On-site cooling only, so comparable with Google and OpenAI's figures.
How we got this number
0.2 Wh per prompt × 1 L/kWh (estimated WUE for Chinese colocation sites). Counting the water evaporated to generate the electricity adds ≈0.36 mL.
Location-based: what the grid actually emitted. Renewable certificates would bring the market-based figure near zero.
How we got this number
0.2 Wh × 580 gCO₂/kWh (Chinese grid, location-based).
Location-based estimate for the AI fleet only; not comparable with a full corporate inventory.
How we got this number
65 MW × 8,760 h × 80% utilisation × 580 gCO₂/kWh (Chinese grid). Compute only — nothing else the company does is counted.
How we got this number
SemiAnalysis reports orders for 'many more' China-specific H20s on top of the ≈50,000 Hopper fleet; a 50% year-on-year increase in compute is the conservative reading.
How we got this number
No disclosure. Chinese colocation sites typically use evaporative cooling; 1.0 L/kWh is a mid-range assumption.
How we got this number
China's grid generated about 35% of its electricity from non-fossil sources; DeepSeek discloses no renewable purchases.
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
≈10,000 H800, 10,000 H100 and H20 variants; ≈$1.6bn of server capex.
Megawatts the AI fleet pulls. One megawatt runs roughly 750 US homes.
How we got this number
50,000 accelerators × 1.3 kW each, all-in. Fleet power implied by the chip count.
Terawatt-hours per year. One TWh is a year of electricity for about 95,000 US homes.
How we got this number
65 MW × 8,760 h × 80% utilisation. Compute fleet only.
CO₂e emitted making the chips and servers before they answer a single prompt.
How we got this number
50,000 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
50,000 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
65 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
65 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
- Full training run of 2.788M H800 GPU-hours disclosed in the technical report; MoE activates 37B parameters per token. DeepSeek (arXiv 2412.19437)
- Training and first-party serving run on China's grid, with no renewable procurement disclosed. DeepSeek (arXiv 2412.19437)
- No water data of any kind. DeepSeek (arXiv 2412.19437)
- Training compute disclosed and weights open; no emissions, energy or water reporting; fleet size known only from analysts. DeepSeek (arXiv 2412.19437)Tom's Hardware
- No climate or water commitments published. DeepSeek (arXiv 2412.19437)
- No environmental enforcement actions on record. DeepSeek (arXiv 2412.19437)
Models
What this company ships, and what each prompt costs
| Model | Weights | Energy / prompt | Water | Training CO₂e |
|---|---|---|---|---|
| DeepSeek-V3 671B total / 37B active (MoE) | Open | 0.20 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. | 0.200 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. | 1,455 t DerivedDerived. Arithmetic on a disclosed or independent input using a published conversion factor. |
Where the compute lives
Named sites from company disclosures and press reporting
- Hangzhou clusterZhejiang, China · ≈50,000 Hopper GPUs (SemiAnalysis).
Timeline
Incidents and developments involving this company
