Carbon in the Age of AI Chips: What the Semiconductor Industry Needs to Know This Earth Day

Stephen Russell: Senior Technical Fellow, TechInsights

Every April, Earth Day prompts a flurry of corporate sustainability pledges and green-tinted press releases. But for the semiconductor industry in 2026, the conversation has moved well past pledges. Carbon accountability is now a procurement requirement, a regulatory expectation, and increasingly a design constraint. This Earth Day, TechInsights is releasing a new sustainability report, Carbon in the Age of AI Chips. Authored by TechInsights Senior Technical Fellow Stephen Russell and Senior Sustainability Analyst Lara Chamness, the report examines where semiconductor emissions are actually coming from, why AI is accelerating the problem faster than most reporting methods can track, and what engineering, procurement, and sustainability teams can do about it right now.

Here’s a preview of what’s inside:

The Scale of the Problem Is Getting Harder to Ignore

Start with the headline numbers. Fabrication emissions are projected to reach 186 million metric tons of CO₂e in 2026, a record high, rising to approximately 247 million metric tons by 2030. Leading-edge technologies below 4nm will account for 26% of total emissions this year, climbing to 42% by 2030. Those are not abstract figures. They represent real consequences of real decisions: which fab to use, which memory configuration to specify, which supplier to source from.

What makes 2026 feel genuinely different from prior years is the convergence of three forces pushing carbon upstream into product decisions. Advanced manufacturing keeps getting more energy- and resource-intensive, especially at leading-edge logic and high-layer 3D NAND. AI demand is driving unprecedented silicon and memory intensity per system, not just more units but fundamentally heavier systems. And procurement teams are being asked, with increasing urgency, to defend supplier choices with traceable carbon logic rather than slide-deck narratives.

Manufacturing Carbon Is a Strategic Variable, Not a Fixed Cost

One of the report’s central arguments is that manufacturing carbon should not be treated as a black box or a rounding error. It is a strategic variable, and it responds to specific decisions.

The report’s Sustainability Matrix maps carbon hotspots across device types and toolsets. For advanced logic nodes, Scope 2 emissions driven by electricity are concentrated in lithography. For 3D NAND, dry etch can account for nearly half of total manufacturing emissions, driven by high-power plasma processes and high global warming potential gases. Some of those gases carry a 100-year GWP of around 25,000 times that of CO₂.

Perhaps the most striking case study involves backside power delivery (BSPD), a major scaling innovation that many assume carries a straightforward carbon penalty due to added process complexity. The reality is more nuanced. In an illustrative comparison of Intel 18A manufactured in the United States versus TSMC N2 manufactured in Taiwan, the Intel process results in lower manufacturing CO₂e per die. Not because it is simpler, but because the U.S. grid is cleaner. The electricity mix where a chip is fabricated can outweigh the complexity of the manufacturing process itself. That is a finding with immediate implications for anyone making sourcing or fab-selection decisions.

AI Hardware Is Scaling Emissions Faster Than Shipments

The report’s treatment of AI accelerators is where the numbers become genuinely striking. TechInsights’ Global AI GPU Manufacturing Carbon Emissions Forecast shows that by 2030, manufacturing emissions from AI GPU production are projected to rise more than twelvefold, from approximately 1.8 million metric tons CO₂e in 2024 to 21.6 million metric tons CO₂e. AI GPU manufacturing is expected to account for roughly 8.7% of total semiconductor die fabrication emissions by 2030. The average accelerator is expected to exceed one metric ton of CO₂e per unit by 2029.

The driver is not primarily bigger logic dies. It is memory, specifically high-bandwidth memory (HBM). The average AI accelerator is expected to integrate roughly 250 HBM dies by 2030. NVIDIA’s Rubin Ultra-class designs are projected to approach approximately 1 TB of HBM through higher stack counts and heights. As Stephen Russell notes, the AI-driven surge in HBM and advanced memory is likely to raise semiconductor manufacturing emissions in absolute terms, increasing memory wafer starts and adding process complexity even as leading manufacturers improve efficiency per transistor.

There is a subtler dimension the report explores carefully: yield. Stacking dies compounds yield loss, and in tall-stack HBM scenarios, stacking yields above 93% are necessary to prevent emissions per usable stack from rising sharply. That makes yield learning and process control first-order sustainability levers, not just cost levers.

The Client Device Story: Carbon Paid Upfront

The report’s third major focus is consumer and enterprise devices, where on-device AI is often framed as an operational efficiency win. Fewer cloud calls, lower network load, specialized local hardware: all genuine benefits. But the manufacturing emissions for those devices are paid upfront, and they are concentrated in places that might surprise you.

Using teardown-based analysis of AI PCs including recent Microsoft Surface and ASUS Zenbook models, the report finds a consistent pattern: memory and storage, not the headline processor or NPU, account for the majority of embodied carbon in AI PC platforms. Across the examples evaluated, memory accounts for roughly 43% to 57% of packaged-IC CO₂e, while the applications processor accounts for only about 14% to 21%.

The supplier concentration finding is particularly actionable. In one Surface Laptop 7 configuration, three suppliers account for approximately 73% of packaged-IC carbon. In a Zenbook S14 model, three suppliers account for roughly 84%. A small number of parts and vendors determine most of the footprint, which means platform configuration and supplier selection are among the highest-leverage carbon choices a product team can make.

What Can Actually Be Done

The report identifies a clear set of high-leverage actions: pursuing cleaner electricity and power purchase agreements, reducing yield loss and rework especially late in the flow and in stacked memory, substituting low-GWP gases with higher abatement efficiency, improving tool energy and utilization, and optimizing bit density and platform configuration.

Semiconductor sustainability has become a decision problem. The highest-impact choices around fab location, memory configuration, supplier mix, and platform architecture are made before a product ships, carrying carbon consequences that most legacy reporting methods cannot capture. The full report, Carbon in the Age of AI Chips, is available now.

LINK: Carbon in the Age of AI Chips | Earth Day eBook | TechInsights

Stephen Russell: Senior Technical Fellow

As Senior Technical Fellow for Sustainability at TechInsights, Stephen provides expert insight into carbon footprint across the entire technology life cycle, from raw materials through product manufacturing, use and end of life. Stephen also works on unique initiatives to characterize Scope 3 emissions in the use phase of consumer electronics products, with further reaching implications for data center and automotive applications.

Stephen is internationally recognized for technical research contributions and collaborations. These include being awarded best paper 2018 for the IEEE Journal Transactions on Power Electronics paper “High Temperature Electrical and Thermal Aging Performance and Application Considerations for SiC power DMOSFETs”. He led an exploratory research project in gallium oxide for power devices, presenting findings to the Royal Institution, London. While working in industry, he led the development of a new silicon IGBT product line and instigated a research and development project to use silicon carbide JFETs in circuit protection applications.

Also Read:

Kirin 9030 Hints at SMIC’s Possible Paths Toward >300 MTr/mm2 Without EUV

Cost, Cycle Time, and Carbon aware TCAD Development of new Technologies

5 Expectations for the Memory Markets in 2025