A single H100 GPU weighs about 3 kilograms. To make it, TSMC and its suppliers will move roughly 30 kilograms of raw material, consume an estimated 1,200 to 1,500 kWh of electricity in fabrication, and generate somewhere between 150 and 250 kilograms of CO2e before the chip ever runs its first inference. That’s one chip. Meta alone ordered 600,000 of them in 2024.
The conversation about AI’s environmental impact keeps stopping at the data center door. It counts the watts that go in and the heat that comes out, then declares the problem understood. The problem is not understood. It’s being hidden in the supply chain, in the mining concessions, in the fab water, and in the landfills of Ghana and Pakistan where the hardware goes to die.
This piece is a full lifecycle teardown of what AI actually costs the planet, from the rare earth mines to the e-waste dumps. And it starts with one sacred cow that the entire industry has been grazing on for a decade: the belief that clean cloud equals clean AI.
The Full Lifecycle, in Five Stages
Most AI carbon accounting stops at operational emissions, which are the electricity consumed by GPUs at inference and training time. That’s one of five stages, and it’s often the smallest one over the lifetime of the hardware.
Extraction. Rare earth elements, cobalt, tantalum, tungsten, gold, and high-purity quartz. Mining for these is carbon-intensive, water-intensive, and geographically concentrated. The Democratic Republic of Congo supplies roughly 70% of the world’s cobalt, much of it from artisanal mines with documented labor abuses. Rare earth processing is dominated by China, which has historically tolerated environmental damage that would be illegal in the EU.
Fabrication. Semiconductor fabs are among the most resource-intensive industrial facilities ever built. A single leading-edge fab consumes 50 to 100 MWh per hour of operation and uses ultrapure water at rates comparable to a mid-sized city. TSMC’s annual electricity consumption is roughly 22 TWh, more than the entire country of Iceland. The embodied carbon of a leading-edge chip is dominated by fab electricity and process gases, including fluorinated compounds with global warming potentials in the thousands.
Packaging and assembly. Advanced packaging (CoWoS, HBM stacking) is itself energy-intensive and uses scarce materials. HBM memory stacks are a significant and growing fraction of the embodied carbon of an AI accelerator.
Operation. This is the part the industry talks about. Training a frontier model can consume tens of GWh. Inference at scale, summed across millions of daily queries, is now comparable to or larger than training. The IEA projects data center electricity could double by 2026, with AI workloads as the primary driver.
Disposal. GPUs and servers have useful lives of 3 to 5 years in hyperscale settings. After that, they enter a global e-waste stream. Roughly 62 million tonnes of e-waste were generated globally in 2022, with a formal recycling rate under 25%. The rest is landfilled, incinerated, or informally processed in countries with weak environmental enforcement.
The lifecycle is not a circle. It’s a line with an exit at the end, and the exit is where most of the accounting goes dark.
The Sacred Cow: “Clean Cloud Equals Clean AI”
Here’s the belief I want to dismantle, and it’s the load-bearing assumption of nearly every AI sustainability report published in the last three years.
The claim goes like this: our data centers run on renewable energy, therefore our AI is clean.
This is mostly wrong, for three reasons.
First, renewable energy procurement is not the same as renewable energy consumption. Most hyperscalers buy renewable energy credits (RECs) or sign power purchase agreements (PPAs) that match their annual electricity consumption on paper. But the grid doesn’t care about your contract. If you draw 100 MW from a grid that’s 40% coal at 3 AM, you’re drawing 40 MW of coal power, regardless of what you bought in a different state at a different time. Marginal emissions are what matter, and RECs don’t fix marginal emissions.
Second, operational emissions are only part of the story. The embodied carbon of the hardware, the fab electricity, the mining, and the eventual disposal are all real emissions that no PPA can offset. For a GPU-heavy workload with a 3-year lifetime, embodied carbon can be 30 to 50% of total lifecycle emissions. When you hear a company say “our AI is carbon neutral,” ask whether they’re counting embodied carbon. Usually they’re not.
Third, the disposal stage is nearly always excluded. E-waste is a global externality, and it lands mostly on countries that had nothing to do with the AI being trained. A GPU that ran in Iowa for three years is a toxic liability in Accra for the next thirty.
If you’re a builder or an operator, the practical implication is this: a clean cloud contract is not a clean AI footprint. You have to measure the whole lifecycle, or you don’t know what you’re shipping.
The Numbers Nobody Puts in the Deck
Concrete figures, because the discourse is vague.
TSMC’s 2023 electricity consumption: roughly 24.8 TWh, or about 6% of Taiwan’s total electricity. TSMC has committed to net-zero by 2050, but its absolute emissions are still rising.A 3nm wafer: roughly 1.5 to 2 tonnes CO2e per wafer at the fab, before packaging. A single wafer yields hundreds of chips, so per-chip embodied carbon varies, but the aggregate is enormous.Cobalt: DRC produces ~70% of global supply. A 2023 study found that industrial cobalt mining in DRC emits roughly 1.5 to 2.5 tonnes CO2e per tonne of cobalt, plus water contamination and tailings issues.Rare earth processing: China controls roughly 60% of mining and 85% of processing. The Bayan Obo mine alone has left a 5-square-mile tailings lake and elevated radioactivity in surrounding areas.Water: a leading-edge fab uses 2 to 4 million gallons of ultrapure water per day. TSMC’s total water consumption is comparable to a city of 1 to 2 million people.E-waste: global total hit 62 million tonnes in 2022, growing at roughly 2.6 million tonnes per year. Formal recycling captures under 25%. The rest is informal, often toxic, and mostly in the Global South.Embedded carbon per H100: estimates range from 150 to 250 kg CO2e, driven mostly by fab energy, HBM stacking, and substrate manufacturing.
Add the operational side: a single frontier training run can emit 500 to 1,500 tonnes CO2e depending on grid, model size, and duration. Multiply by the number of labs doing this. Multiply by inference at global scale. The numbers compound faster than any renewable procurement strategy can offset.
Why the Lifecycle Is Broken at Both Ends
The extraction end and the disposal end share a common feature: they happen far from where the value is captured, and far from where the accountability lives.
Extraction happens in DRC, Chile, Indonesia, and China, under regulatory regimes that have historically prioritized output over environmental protection.Disposal happens in Ghana, Pakistan, Nigeria, and India, where informal recyclers burn cables and acid-bath circuit boards to recover metals, exposing workers and communities to lead, mercury, and dioxins.
The middle of the chain, the fab and the data center, is where the money and the media attention live. That’s not a coincidence. It’s a structural feature of how global supply chains distribute harm.
If you’re building AI products, the implication is uncomfortable but clear: you are downstream of a supply chain you don’t control and can’t easily audit. “Carbon neutral” claims almost never cover the ends of that chain. Neither do most ESG reports.
What Actually Moves the Needle
Let’s talk about real interventions, not slogans.
Extend hardware life. The single highest-leverage move for embodied carbon is to use GPUs longer. A GPU that runs for 5 years instead of 3 amortizes its embodied carbon over 67% more compute. This is not glamorous, and it conflicts with the upgrade cycle the industry is addicted to. Do it anyway.
Match workloads to marginal grid carbon. Run training when and where the grid is cleanest. This is schedulable in a way that inference often isn’t. A 30% reduction in marginal emissions per training run is achievable with existing tooling.
Track tokens-per-joule as a first-class metric. Quantization, distillation, speculative decoding, and MoE routing can cut inference energy by 3 to 8x with minor quality loss. If your dashboard shows latency and cost but not energy per token, you’re flying blind on the metric that will constrain you next.
Demand supply chain transparency from vendors. Ask your hardware provider for embodied carbon per unit, fab location, energy mix, and end-of-life program. Most will not have answers. The ones who do are worth paying a premium for.
Plan for disposal before you buy. If you’re deploying at scale, build a hardware retirement plan. Refurbish, resell, or certify destruction through a responsible recycler. The default path is a container ship to Accra.
If you want a code-level starting point for lifecycle-aware instrumentation:
Rough lifecycle-aware carbon accounting for an AI workload
def lifecycle_co2(
gpu_hours, # total GPU-hours for the workload
gpu_power_w, # average wall power per GPU
pue, # facility power usage effectiveness
grid_g_per_kwh, # marginal grid carbon intensity
embodied_kg_per_gpu, # embodied carbon per GPU
gpu_lifetime_hours, # expected service life
gpus_used, # number of GPUs provisioned
):
operational_kg = (
gpu_hours * (gpu_power_w / 1000) * pue * grid_g_per_kwh / 1000
)
embodied_kg = (
gpus_used * embodied_kg_per_gpu * (gpu_hours / gpu_lifetime_hours)
)
return {
"operational_kg_co2e": operational_kg,
"embodied_kg_co2e": embodied_kg,
"total_kg_co2e": operational_kg + embodied_kg,
}
If your team is only tracking the operational_kg line, you're missing somewhere between 30 and 50% of the footprint depending on workload shape and hardware generation.
The Mental Model
Here’s the lens I want you to keep.
AI’s carbon footprint is not a data center problem. It’s a global supply chain problem with a data center in the middle. The industry has spent a decade optimizing the middle and ignoring the ends. That worked when the ends were small relative to operation. It doesn’t work now, because embodied carbon is rising as chips get more complex, and disposal externalities are compounding as deployment scales.
The correct unit of analysis is not “kWh per inference.” It’s total lifecycle carbon per unit of useful work, measured end to end, including the mine and the landfill. Very few teams measure this. The ones that do will be structurally advantaged when regulation, procurement standards, and customer expectations catch up. The ones that don’t will be exposed.
The interesting question is not whether AI can be carbon neutral. It’s whether the industry will treat the full lifecycle as its responsibility before it’s forced to. History suggests it won’t. That’s the gap worth closing yourself, in your own stack, starting this week.
What’s your read? If you’ve tried to measure embodied carbon in your own infrastructure, where did you get stuck? And if you haven’t, what would it take for you to start? Drop it in the comments
Top comments (0)