The Hidden Environmental Cost of Advanced AI Chips
7 Min Read August 5, 2025
As artificial intelligence accelerators grow more powerful to meet the demands of modern workloads, one consequence is generally accepted...
As artificial intelligence accelerators grow more powerful to meet the demands of modern workloads, one consequence is generally accepted: These chips consume enormous amounts of energy during use. But what's often overlooked is the substantial environmental footprint created long before the chip is ever powered on. What about the emissions from manufacturing these chips?
If we explore the semiconductor emissions and water usage associated with the production of three advanced AI chips: AMD's MI300X, NVIDIA's Blackwell, and Intel's Gaudi, it's easy to see how architecture, die count, process node, and high-bandwidth memory (HBM) integration dramatically influence their manufacturing impact.
Chip Complexity Drives Emissions and Water Use
The MI300X leads the pack in complexity, with an astounding 129 individual dies. Unsurprisingly, it also tops the list in semiconductor manufacturing carbon emissions. Blackwell, with 90 dies, comes in second, followed by Gaudi with 81 dies. Note that a standard package yield of 98% was assumed for all three products.
But semiconductor emissions are only part of the story. Water consumption, a critical and often scarce resource, follows a similar pattern. Producing just one MI300X requires more than 40 gallons of water, compared to just over 30 gallons for a Blackwell and under 20 gallons for a Gaudi. To put this into perspective, 40 gallons is enough to supply a person with drinking water for nearly three weeks.
Figure 1: Semiconductor cradle-to-gate emissions. Emissions include: die manufacturing, assembly & test, and transportation. Does not account for HBM stacking yield loss.
Architecture Matters
Why such differences? Much of it comes down to design. The MI300X uses a chiplet-based architecture with eight smaller GPU dies built on a 5 nm process, supported by a complex HBM stack and additional auxiliary dies. While this modular approach has performance advantages, it also drives up assembly, testing, and interconnect complexity, adding to the environmental burden. In fact, MI300X shows the highest share of assembly and test emissions, making up 8% of its total.
In contrast, both Blackwell and Gaudi use two larger GPU dies. Blackwell is built on a 4 nm process, while Gaudi uses 5 nm. The HBM configurations in both are broadly similar. GPU dies account for more than 50% of Blackwell's total semiconductor carbon footprint. In contrast, the majority of Gaudí's manufacturing emissions, nearly 60%, come from its HBM dies. This is due to modeled lower yields for its HBM die compared to Blackwell, highlighting how even subtle differences in manufacturing efficiency can lead to big environmental differences.
Why This Matters
These differences underscore how architectural decisions, such as die count, size, process node, and packaging, can significantly impact carbon and water footprints. Improving yield across all die types is one of the most powerful ways to reduce emissions for these chips.
Designing Smarter, Not Just Faster
As AI and high-performance computing continue to evolve, performance gains will always be a priority. But it's time to weigh sustainability just as heavily. The takeaway is simple: the more complex the chip, the larger its environmental and semiconductor carbon footprints. By understanding and addressing the manufacturing impacts early in the design phase, companies can build next-generation chips that are both high-performing and lower-impact.
