Microsoft’s latest datacenter experiment features an unconventional twist: timber. In Virginia, the company is building two facilities using cross-laminated timber (CLT) to reduce the environmental impact of construction. On paper (or screen), it sounds impressive—emissions are expected to be 35% lower compared to steel-only designs and up to 65% less than concrete-heavy ones. However, the timing raises questions about whether this eco-friendly move can genuinely make a dent, given the enormous energy requirements that Microsoft’s AI operations impose.
The tech giant claims these CLT panels, composed of layers of timber glued and pressed for maximum strength, will replace traditional materials in large parts of the datacenters. Yet, the concrete foundation remains, and steel beams will still provide structural support. While the carbon savings from reduced use of heavy materials are laudable, the real issue is whether such a change can balance the escalating power consumption driven by AI models like ChatGPT.
The AI Energy Dilemma
The push to use more sustainable building materials comes at a time when Microsoft’s data infrastructure is ballooning. Earlier this year, the company revealed it had surpassed 5 gigawatts of installed datacenter capacity, with more than 500 megawatts added in just nine months. These numbers reflect a massive investment in AI, which requires extraordinary amounts of electricity to train and operate.
Generative AI models, reliant on thousands of GPUs running in parallel, are power-hungry beasts. Critics argue that swapping concrete for timber, while better than nothing, might be more symbolic when placed against the backdrop of AI’s insatiable appetite for energy.
April brought an announcement of even more aggressive expansion plans. Microsoft wants to double its datacenter growth by mid-2024 and triple it in the first half of its fiscal year 2025, rolling out over 200 megawatts each month.
GPU Expansion and Growing Carbon Concerns
Over the past year, Microsoft has ramped up its GPU capacity, more than doubling it to power complex AI models and extending its AI clusters to 98 global locations. The infrastructure investments are part of a strategy to maintain a competitive edge in AI, which has become a multi-billion-dollar endeavor. However, these datacenters guzzle enormous amounts of electricity, with carbon emissions tied not only to the construction but also to the continuous operation of servers.
This raises a crucial point: Are the sustainability efforts enough? The use of timber, while a step forward, pales compared to the environmental impact of powering AI infrastructure. Microsoft’s multibillion-dollar investments in GPU resources underscore this paradox, suggesting that the green construction efforts may not sufficiently offset the environmental damage.
Project Natick: Underwater Datacenters and Mixed Results
The company isn’t new to exploring alternative datacenter models. In June 2024, Microsoft concluded Project Natick, a decade-long experiment that involved submerging datacenters off the Scottish coast. The results showed that underwater servers had one-eighth the failure rate of land-based systems and could be deployed in 90 days—a fraction of the time traditional builds require. Yet, despite these promising findings, Microsoft chose not to continue with the subsea approach.
Noelle Walsh, head of Microsoft’s Cloud Operations + Innovation division, explained that while the project won’t move forward, insights from underwater environments are informing other areas, such as liquid immersion cooling. This method involves submerging servers in non-conductive liquid, making cooling systems more efficient. Still, critics note that these innovations, like the use of timber, address only the margins of a much larger issue: AI’s massive and growing energy footprint.
Symbolic or Substantial? Microsoft’s Environmental Investments
Microsoft often highlights its $1 billion Climate Innovation Fund, launched in 2020 to accelerate the development of low-carbon technologies. By mid-2024, the fund had invested $761 million in initiatives like hydrogen-powered steel and carbon-sequestering concrete. But how impactful are these investments when placed alongside the vast emissions generated by AI infrastructure? Brandon Middaugh, who manages the fund says that they are not only funding these projects but also arecommitting to use these materials themselves. Yet, this commitment often feels dwarfed by the scale of the AI operations.
One of the fund’s biggest bets is on Stegra, a Swedish firm producing green steel with renewable hydrogen. This method could cut emissions by 95%, a significant improvement over coal-based steel production. But steel and cement still account for 15% of global emissions. Technologies like CarbonCure, which traps CO₂ in concrete, and Prometheus Materials’ algae-based cement are promising but still experimental. Small pilot projects in Microsoft’s datacenters are testing these materials, yet they barely offset the environmental impact of a rapidly expanding AI ecosystem.
The Timber Trend: A Meaningful Shift or Greenwashing?
Cross-laminated timber, while a step in the right direction, might not be the game-changer it appears to be. CLT has been used in Europe for years, and Microsoft’s use of it at scale began with its Silicon Valley campus in 2021. Prefabricated CLT panels do cut down on construction time and labor costs, but they remain pricier than conventional materials. Structural engineer David Swanson stressed that the Virginia datacenters are undergoing rigorous safety testing. But even with advanced fire resistance and concrete reinforcements, the carbon savings seem minor compared to the ongoing energy use of AI.
Microsoft’s scale does give it the leverage to push sustainable technologies, argues Thomas Hooker of Thornton Tomasetti, a structural firm involved in the project. But whether this influence is wielded meaningfully or mainly for optics remains debatable. The data-heavy future Microsoft is building comes at an environmental cost that no amount of sustainable construction can easily offset.
Last Updated on November 7, 2024 2:13 pm CET
