Hangzhou DeepSeek Artificial Intelligence Co., Ltd.
AI LabLegal name: Hangzhou DeepSeek Artificial Intelligence Co., Ltd. · HQ: CN · Private · Founded 2023
Chinese open-weight frontier lab funded by hedge fund High-Flyer. Its sparse mixture-of-experts models use far less compute per token than US frontier labs, but the company publishes no environmental data of any kind and runs on a coal-heavy grid. All figures are modeled.
Disclosure Quality
Score
3/100
Transparency Grade
T-5
near-total opacity
Grades run T-1 (most transparent) to T-5 (least transparent).
Red flags identified (4)
- RF-01 · Total energy consumption not disclosed
- RF-04 · AI-specific energy use not disclosed
- RF-02 · A greenhouse-gas emissions scope not disclosed
- RF-03 · Water withdrawal not disclosed
See the disclosure audit below for details.
Fiscal year 2026. Score computed from OpenButterfly T1–T6 disclosure framework audit. Confidence tier: C (Estimated, Documented Method).
Environmental Metrics
Fiscal year 2026. Each metric is shown at its most recent available year; rows from a different year are marked.
Confidence tiers: A measured and verified · B company-reported · C documented estimate · D uncertain estimate · E modeled · ND not disclosed.
Total Energy Consumption
438 GWh
≈ the annual electricity use of 42,000 US homes
method: known_capacity_model
Source: OpenButterfly pipeline
Scope 1 Emissions
5,081 tCO₂e
≈ the annual emissions of 1,105 passenger cars
method: scope1_ratio_model
Source: OpenButterfly pipeline
Scope 2 Emissions (market-based)
203 kt CO₂e
≈ the annual emissions of 44,000 passenger cars
method: grid_carbon_model
Source: OpenButterfly pipeline
Scope 3 Emissions
389 kt CO₂e
≈ the annual emissions of 84,000 passenger cars
method: scope3_ratio_model
Source: OpenButterfly pipeline
Water Withdrawal
526 ML
≈ 210 Olympic swimming pools of water
method: wue_model
Source: OpenButterfly pipeline
Property-Line Noise (estimated)
~57.2 dBA at property line
method: cooling_load_noise_model
Source: OpenButterfly pipeline
Disclosure Audit
Show 5 disclosure checks (T1–T6 framework)
Total electricity consumption disclosed
No environmental disclosure of any kind.
RF-01 (Total energy consumption not disclosed)
AI-specific energy consumption disclosed
Training compute (2.788M H800 GPU-hours for V3) known only from research papers, not environmental reporting.
RF-04 (AI-specific energy use not disclosed)
Scope 1 GHG emissions disclosed
No environmental disclosure of any kind.
RF-02 (A greenhouse-gas emissions scope not disclosed)
Total water withdrawal disclosed
No environmental disclosure of any kind.
RF-03 (Water withdrawal not disclosed)
GHG accounting methodology documented
No GHG methodology published.
Sources
- Insights into DeepSeek-V3: Scaling Challenges and Hardware Reflections· academic · trust: high · accessed 2026-07-18
V3 trained on 2.788M H800 GPU-hours; MoE 671B/37B active.
- DeepSeek — company profile· journalism · trust: medium · accessed 2026-07-18
High-Flyer ownership, Firefly cluster fleet; no environmental reporting.