The Carbon Surge Behind AI's Expansion

The Numbers Tell the Story

Google's total carbon emissions climbed 25% year-over-year, while Amazon's increased 16%, according to sustainability reports both companies released this week. The scale of those jumps is striking for two firms that have spent years touting renewable energy investments and net-zero targets. Neither report explicitly names artificial intelligence as the culprit, yet the infrastructure required to train and run large language models sits at the heart of the problem.

At DailyTechWire, we've tracked the energy appetite of AI systems since the first wave of transformer models hit production. What these reports confirm is that the theoretical concern has become a measurable reality. The buildout is no longer incremental; it is reshaping the carbon profiles of the world's largest technology companies faster than their decarbonization strategies can adapt.

Where the Emissions Come From

The bulk of the increase does not stem from electricity consumed in offices or even from the direct energy use of data centers, though that remains significant. Instead, the surge originates in Scope 3 emissions, the category covering goods and services a company purchases and the lifecycle impact of products it sells. For Google and Amazon, that means data center construction materials, semiconductor manufacturing, and the embedded carbon in GPUs shipped from Asian fabs.

Google's Scope 3 emissions rose by 2.1 million metric tons in the past year, doubling the level recorded in 2019, the baseline year the company uses for climate accounting. Amazon added more than 1.2 gigawatts of data center capacity in the fourth quarter alone, according to the company, more than any competitor in a single quarter. That kind of expansion requires steel, concrete, cooling systems, and racks of high-performance chips, all of which carry carbon baggage.

Capital goods represent the largest contributor within Scope 3 for both firms. Data centers under construction involve cement and steel, two of the heaviest-emitting industries globally. While startups in both sectors are developing low-carbon alternatives, those technologies have not yet reached the volume or cost point needed to serve hyperscale infrastructure projects. The result is that every new facility locks in emissions before a single watt of electricity flows through it.

The Semiconductor Bottleneck

Graphics processing units and high-bandwidth memory chips power the training runs and inference workloads that define modern AI. Manufacturing those components is energy-intensive, and the majority of leading-edge production happens in Taiwan, South Korea, and China, where grids remain heavily reliant on coal and natural gas. The chemicals used in semiconductor fabrication, including certain fluorinated gases, are also potent greenhouse agents, with global warming potential thousands of times higher than carbon dioxide on a per-molecule basis.

As demand for AI accelerators has surged, so has the embedded carbon in the hardware supply chain. Amazon and Google are both large buyers of Nvidia's H100 and H200 chips, as well as custom silicon from their own design teams. Each purchase adds to the Scope 3 tally, and neither company has direct control over the energy mix or process chemistry at the foundries that produce those chips.

This dynamic creates a structural challenge. Even if Google and Amazon achieve 100% renewable energy coverage for their own operations, the carbon footprint of their supply chains will continue to grow unless chipmakers and materials producers decarbonize at a similar pace. That coordination is difficult when the supply base spans multiple continents and regulatory regimes.

Renewable Energy Reaches Its Limit

For years, purchasing renewable power through power purchase agreements allowed tech companies to offset the emissions from their electricity use. Google and Amazon have both signed deals totaling gigawatts of wind and solar capacity, and those agreements have kept direct energy-related emissions relatively stable. But AI's power demands are outpacing the availability of renewable supply in many regions, particularly where data centers are concentrated.

In response, some hyperscalers have begun investing in natural gas plants to ensure reliable baseload power. Google has explored partnerships with gas providers, and other firms have floated the idea of co-locating nuclear reactors with data centers. While nuclear power offers low-carbon generation, it comes with long lead times and regulatory complexity. Natural gas, by contrast, is fast to deploy but adds to the carbon ledger.

The shift toward fossil fuel backup signals that the renewable-only strategy has hit a ceiling. Battery storage can smooth out intermittency for some workloads, but training large models requires sustained, high-density power that current battery economics struggle to support. The result is a growing reliance on dispatchable generation, much of which remains carbon-intensive.

The Carbon Intensity Metric

Both companies devote space in their reports to discussing carbon intensity, a measure of emissions per unit of revenue. The metric has appeal because it allows a company to grow its business while still showing improvement, as long as revenue outpaces emissions. China has used a similar approach in international climate negotiations, arguing that emissions per unit of GDP matter more than absolute totals.

For Google and Amazon, carbon intensity provides a more favorable narrative than absolute emissions. Both firms have increased revenue substantially, so even with emissions rising, intensity may have improved or remained flat. But from an atmospheric perspective, absolute emissions are what count. A ton of CO₂ contributes the same warming effect whether it was generated efficiently or wastefully.

The focus on intensity also raises questions about how companies prioritize growth versus environmental goals. If AI-driven revenue expands faster than emissions, the intensity metric improves, even if the planet sees more carbon. That misalignment is not unique to tech, but it becomes more visible when companies tout climate leadership while expanding carbon-heavy infrastructure.

What It Will Take to Close the Gap

Reaching net-zero under these conditions will require more than incremental adjustments. Both Amazon and Google will need to accelerate renewable energy procurement, fund the commercialization of low-carbon steel and cement, and purchase large volumes of carbon removal credits to offset what cannot be eliminated. Each of those steps carries cost and complexity.

Carbon removal, in particular, remains an immature market. Direct air capture and enhanced mineralization technologies are still measured in thousands of tons per year, while the offsets needed by hyperscalers will eventually reach millions. Scaling those solutions will require sustained capital and patience, neither of which is guaranteed in a competitive environment where AI performance and speed to market drive strategic decisions.

Supply chain decarbonization is another lever, but it depends on coordination with partners who may not share the same timelines or incentives. Chipmakers are under pressure from multiple customers and regulators to reduce emissions, but the capital cycles in semiconductor manufacturing are long, and retrofitting fabs is expensive. Progress will happen, but not at the pace the current AI buildout demands.

The Tension Between Ambition and Reality

The sustainability reports from Google and Amazon reflect a broader tension in the technology industry: the desire to lead on climate action while simultaneously pursuing the most energy-hungry product category in a generation. Both companies have made genuine investments in renewables and efficiency, and their climate teams are working within real constraints. But the scale and speed of AI deployment have outrun those efforts.

What remains unclear is whether the companies will adjust their AI strategies to fit within their climate commitments, or whether the commitments will be quietly relaxed to accommodate AI's trajectory. The reports do not answer that question, but the direction of the emissions curves suggests the latter is more likely unless something changes soon.

For now, the data centers keep going up, the chips keep shipping, and the carbon footprint keeps expanding. The question is no longer whether AI has an environmental cost. It is whether the companies profiting from it are willing to pay what it takes to bring that cost down.