Skip to content

#28 of 30 · Open-weight lab

Zhipu AI (Z.ai / GLM)

First 'AI tiger' to list; a reported 1 GW all-domestic-chip campus; nothing on emissions.

Beijing, China·z.ai ·Verified Oct 17, 2026
CMostly estimates
Index score
Mostly estimates

In plain words

How scores work

Zhipu scores 34 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 · GLM-4.5/4.6 are mixture-of-experts models with 32B active parameters; weights are open.
  • Transparency · Open weights and a listing prospectus that discloses compute spending by year.

What hurts

  • Efficiency · Ascend-based training and serving uses two to four times the compute of comparable NVIDIA hardware; nothing measured.
  • Carbon · Chinese grid, a 1 GW campus and no renewable procurement disclosed.
  • Water · No water data.

What nobody knows

  • Per-prompt energy disclosed
  • Per-prompt water disclosed
  • Training emissions in model cards
  • Annual environmental report
  • Location-based Scope 2 reported
  • +4 more

Zhipu (Z.ai) ships the open-weight GLM-4.5 and 4.6 family, listed in Hong Kong in January 2026 (HK$4.35 billion raised, 70% earmarked for model training) and reports compute-service costs that rose from RMB 15 million in 2022 to RMB 1.55 billion in 2024. It says GLM-Image was trained entirely on Huawei Ascend chips, and Bloomberg reported a 1 GW data centre running only domestic chips. Domestic chips need two to four times the power of NVIDIA's for the same result, and Zhipu discloses nothing about energy, water or emissions.

Products
GLM-4.6 / GLM-4.5 · GLM-Image · Z.ai chat · Zhipu coding agent
Hosting
Own and leased capacity in China on Huawei Ascend and NVIDIA chips; a reported 1 GW domestic-chip campus.

Key figures

Per-prompt and fleet metrics. Hover a title for what it means; open “How we got this number” for the working.

How to read the labels:Company figureCompany figure. Published by the company itself. We take it as given and record the scope next to it.Measured independentlyMeasured independently. Measured or estimated by researchers or analysts with no stake in the result.ecogpt.co 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.DerivedDerived. Arithmetic on a disclosed or independent input using a published conversion factor.Not disclosedNot disclosed. The company publishes nothing on this and no defensible estimate exists.— hover any label.
Energy per promptElectricity used to answer one typical request. Under 0.3 Wh is good. A 10 W LED bulb uses 0.3 Wh in about two minutes.
0.25
Wh / typical prompt
≈ 20.8 smartphone charges per 1,000 prompts
ecogpt.co 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.eWeekEpoch AI

ecogpt.co estimate; no company figure.

How we got this number

GLM-4.5 and 4.6 are 355B mixture-of-experts models with 32B active parameters, served on Huawei Ascend chips that need two to four times the compute of NVIDIA's for the same work. Uncertainty roughly ±3×.

Water per promptWater evaporated cooling the servers for one request. Scope matters: on-site cooling only, or also the water used to generate the electricity (which adds ≈1.8 mL per Wh).
0.25
mL / typical prompt
≈ 0.50 500 mL bottles per 1,000 prompts
ecogpt.co 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.eWeekEpoch AIKoomey Analytics

On-site cooling only, so comparable with Google and OpenAI's figures.

How we got this number

0.25 Wh per prompt × 1 L/kWh (estimated WUE for Chinese colocation sites). Counting the water evaporated to generate the electricity adds ≈0.45 mL.

Carbon per promptGreenhouse gas from one request on the grid where it actually runs. Below 0.1 g is clean. A kilometre in a petrol car is about 170 g.
0.145
gCO₂e / typical prompt
≈ 0.85 km driven (petrol car) per 1,000 prompts
ecogpt.co 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.eWeekEpoch AI

Location-based: what the grid actually emitted. Renewable certificates would bring the market-based figure near zero.

How we got this number

0.25 Wh × 580 gCO₂/kWh (Chinese grid, location-based).

Annual emissionsThe company's whole yearly footprint, in million tonnes of CO₂e. Clouds and device makers dwarf pure labs because they also count shops, logistics or products. Compare the trend, not the size.
1.22
MtCO₂e / year (AI compute only)
ecogpt.co 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.KuCoin News

Location-based estimate for the AI fleet only; not comparable with a full corporate inventory.

How we got this number

300 MW × 8,760 h × 80% utilisation × 580 gCO₂/kWh (Chinese grid). Compute only — nothing else the company does is counted.

Emissions trendDirection of travel against last year or the company's own baseline. Negative is good. Positive means the AI build-out is outrunning clean energy.
+200
% (1 GW campus ramping; IPO proceeds to training)
ecogpt.co 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.Dolphin Research (Longbridge)eWeek
How we got this number

Compute fees grew tenfold between 2022 and 2024, a 1 GW campus is partly operating, and 70% of IPO proceeds go to training.

Water per kWh (WUE)Litres of water consumed per kilowatt-hour of server electricity, across the fleet. 0.15 is excellent, 0.3 is good, 1.8 is the US data-centre average.
1.00
L/kWh
ecogpt.co 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.Koomey Analytics
How we got this number

No disclosure; mainland Chinese sites mostly use evaporative cooling. 1.0 L/kWh is a mid-range assumption.

Clean electricityShare of electricity matched with renewable or carbon-free sources. Annual matching counts certificates bought anywhere; hourly matching on the same grid is far stricter.
35.0
% non-fossil electricity (grid share, no procurement)
ecogpt.co 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.KuCoin News
How we got this number

China's grid is about 35% non-fossil; Zhipu discloses no renewable purchases.

Hardware & materials

The chips, power, copper and future e-waste behind the fleet — the part sustainability reports leave out

Factors we use
AI chips in service
200,000
Ascend accelerators (order of magnitude)
≈ 2,778 racks of 72 GPUs
ecogpt.co 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.Dolphin Research (Longbridge)KuCoin News

GPUs, TPUs or custom accelerators the company runs. Analyst figures are in H100-equivalents, company figures in physical chips.

How we got this number

Zhipu reports compute-service fees of RMB 1.55bn in 2024 and Bloomberg reported a 1 GW campus running only domestic chips, partly operating; a full 1 GW would hold about 700,000 accelerators. ±2×.

Power draw
300
MW (2026)
≈ 225,000 US homes' worth of power
ecogpt.co 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.KuCoin News

Megawatts the AI fleet pulls. One megawatt runs roughly 750 US homes.

How we got this number

Partial operation of a reported 1 GW domestic-chip campus plus leased cloud capacity.

Electricity a year
2.10
TWh / year
≈ 200,000 US homes for a year
DerivedDerived. Arithmetic on a disclosed or independent input using a published conversion factor.KuCoin NewsSemiAnalysis

Terawatt-hours per year. One TWh is a year of electricity for about 95,000 US homes.

How we got this number

300 MW × 8,760 h × 80% utilisation. Compute fleet only.

Carbon baked into the hardware
103
ktCO₂e embodied in chips and servers
≈ 22,348 petrol cars driven for a year
DerivedDerived. Arithmetic on a disclosed or independent input using a published conversion factor.Dolphin Research (Longbridge)KuCoin NewsarXiv 2607.05408NVIDIA developer blog

CO₂e emitted making the chips and servers before they answer a single prompt.

How we got this number

200,000 accelerators × 514 kg CO₂e each (chip plus share of the host server). Emitted in manufacturing, before the first prompt is served.

Future e-waste
3,400
tonnes of servers to retire within ~5 years
≈ 19 empty Boeing 747s by weight
DerivedDerived. Arithmetic on a disclosed or independent input using a published conversion factor.Dolphin Research (Longbridge)KuCoin NewsNVIDIAThe Register

Tonnes of server hardware that will be retired in three to five years.

How we got this number

200,000 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.

Copper in the build
8,100
tonnes of copper in the build
≈ copper for 97,590 electric cars
DerivedDerived. Arithmetic on a disclosed or independent input using a published conversion factor.KuCoin NewsBHPS&P Global Market Intelligence

Tonnes of copper in cabling, busbars and transformers.

How we got this number

300 MW × 27 t of copper per MW (cabling, busbars, transformers).

Fossil generators on site
115
backup diesel generators (implied)
≈ 299 MW of backup capacity
DerivedDerived. Arithmetic on a disclosed or independent input using a published conversion factor.KuCoin NewsLatitude MediaLatitude Media / Floodlight

Diesel backup units or gas turbines at the data centres, permitted or observed.

How we got this number

300 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.

Hardware reused or recycled
—
No hardware reuse or e-waste figure published.
Not disclosedNot disclosed. The company publishes nothing on this and no defensible estimate exists.

Share of retired servers and parts given a second life.

Mineral depletion
—
Only Mistral has published a mineral-depletion figure.
Not disclosedNot disclosed. The company publishes nothing on this and no defensible estimate exists.

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

Energy efficiency — How much electricity one answer costs.weight 20% · confidence C
55
  • GLM-4.5/4.6 are mixture-of-experts models with 32B active parameters; weights are open. eWeek
  • Ascend-based training and serving uses two to four times the compute of comparable NVIDIA hardware; nothing measured. KuCoin News
Carbon intensity — How clean that electricity is, and whether total emissions are rising.weight 25% · confidence C
25
  • Chinese grid, a 1 GW campus and no renewable procurement disclosed. KuCoin News
Water — How much water the data centres consume, and whether it is replaced.weight 15% · confidence D
30
Transparency — How much the company actually publishes.weight 20% · confidence C
30
Commitments — Whether it has dated, measurable climate and water targets.weight 10% · confidence C
10
Conduct — How it behaves: permits, lawsuits, gas turbines, community water.weight 10% · confidence C
55
  • No enforcement actions on record. eWeek

Models

What this company ships, and what each prompt costs

ModelWeightsEnergy / promptWaterTraining CO₂e
GLM-4.6
355B total / 32B active (MoE)
Open0.25 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.0.250 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.5,000 t 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.

Where the compute lives

Named sites from company disclosures and press reporting

OperationalConstructionPlanned
  • Beijing HQBeijing, China
  • 1 GW domestic-chip campus (reported)China (location undisclosed) · Bloomberg report; partly operating.