Beyond Data Center Decarbonisation Hype: Sustainable Solutions in Clean Firm Power and Grid-Enhancing Technologies

Given how AI is transforming industries and our daily lives, we can be sure that there’s significant work to be done to enable decarbonisation in our increasingly energy-hungry world. Although data center emissions projections vary greatly among sources, McKinsey predicts that data center emissions could more than triple by 2030, with AI, of course, being the key driver behind growth in data center capacity.

With this, AI is also creating an opportunity for data center operators to think creatively about shaping computational loads. Diverging from traditional cloud data centers that host applications or user data and thus require full backup and redundancy, AI training-focused data centers do not have such requirements. Because model training can be paused and resumed, the infrastructure doesn’t require 100% uptime and has potential to serve as a flexible grid asset during periods of peak demand or grid stress. Although it’s not realistic for GPUs/TPUs to fully ramp up and down since they account for on average 40% of CapEx in AI training data centers, they can shift for a couple of hours a day to avoid system peak. In fact, offering curtailable load for the data center itself and/or its associated backup generation can help data center operators secure a grid connection faster, which is increasingly critical as the grid becomes more constrained. More on how we’re taking this into account in terms of investment targets later.

So with AI largely responsible for increasing data center usage, while also ushering in a new era of flexible data centers, where are data center operators on energy efficiency? It turns out, hyperscalers have done most of the hard work on optimising energy efficiency within the data center. This has led to a global power usage effectiveness (PUE) of ~1.5, with most hyperscalers reporting PUEs of ~1.1. PUE measures how much energy the entire data center facility uses compared to powering the IT equipment itself, so a PUE of 1.0 would represent a perfectly efficient data center.

‍

Since most of the low-hanging energy efficiency fruit has already been plucked, where does this leave us on how to decarbonise data centers? Our take: the most interesting opportunities are no longer on optimising what’s inside the data center, but rather on harnessing decarbonisation opportunities outside of it (i.e., clean firm power and grid-enhancing technologies).

This is also largely supported by hyperscalers. During Latitude Media’s Transition AI conference last fall, the former Director of Energy at a hyperscaler underscored the need for clean firm power. He emphasised that big tech intends to hit their sustainability goals and will pursue all clean energy solutions, focusing on net new gas generation with carbon capture and storage (CCS), nuclear, and geothermal. Addressing the importance of updating our grid, the Head of Data Center Energy at another hyperscaler emphasised that one of the biggest bottlenecks in enabling AI is making the grid visible to appropriately assess risk. She noted that widespread adoption of technologies like dynamic line ratings (DLR) will help with this, but utilities need to be able to capture as many data points as possible to enable grid orchestration in the near-term.

Here’s how we’re thinking about the levers of data center decarbonisation we’re largely bullish on:

‍

Solutions Overview: Clean Firm Power

Bridge/Backup Power: We’re interested in technologies like fuel cells and linear generators that can support data centers pre-grid connection and then transition to backup power post-connection. We also see potential for one of our existing portfolio companies, Reverion, to use their plug and play 100kW units that electrochemically convert biogas or hydrogen into electricity to serve as affordable bridge power/backup generation. And although we don’t see this as an early-stage investment target for us, we like a new co-location strategy, coined “power couples” by RMI, to pair solar, wind, and batteries with existing natural gas peaker plant infrastructure. The aim is to repurpose an existing gas connection, which is inherently underutilised when the grid is not experiencing a peak, and integrate it with the renewables, while also creating backup power for the data center through the gas plant. Overall, we see this as a faster, cheaper, and greener way to meet data center demand as geothermal and nuclear scale up, in comparison to net new gas generation with CCS. Beyond this, it’s also a more logical solution given gas turbine backlogs projected through 2030.

Geothermal: We’re always assessing developer solutions across enhanced geothermal systems (EGS) and advanced geothermal systems (AGS) that can enable power production anywhere, as well as technology provider plays that enable a piece of the geothermal tech stack (e.g., high temperature sensors and electronics, plasma drill bits). For developer plays, we need to see a team with superstar fundraising abilities as they will likely need to raise hundreds of millions to scale up (see Fervo which just raised another $206M), coupled with commercial milestones that can be reasonably achieved to unlock each subsequent fundraising round.

Nuclear Fission: We reassessed gen IV reactors / SMRs for power production and ultimately decided against investing at this time. We see promising reactor technologies in development, our favourite being molten salt reactors that offer unique refuelling and safety benefits (check out Stellaria and Thorizon); however, we ultimately couldn’t get behind the unclear regulatory requirements and the reality that none of these next-generation reactors have been built outside of China and Russia to date.

Nuclear Fusion: We still don’t see fusion aligning with a time-to-impact that matches our fund’s exit timelines. See our full thinking here.

‍

Solutions Overview: Grid-Enhancing Technologies

Advanced Conductors: We’re interested in superconductor cables — room-temperature solutions being the holy grail, and also looking at high-temperature solutions with realistically scalable liquid nitrogen cooling strategies. We’ve also looked into photonic coatings that can be applied to existing or new lines to significantly increase line capacity. Such solutions are particularly exciting because they offer the lowest cost percentage capacity improvement per km and avoid the CapEx of full reconductoring. Stay tuned for an investment announcement here soon!

Advanced Transformers: We’ve assessed next-gen transformer solutions that reduce energy losses and offer condition monitoring through embedded sensors and IoT connectivity. However, given the global transformer shortage, with lead times up to three years, we’re most interested in monitoring solutions that can extend the life of and squeeze more out of transformers already on the grid.

Grid Orchestration: Today’s grid is currently overbuilt and underutilised. Grids are over-dimensioned for worst-case peak load scenarios (e.g., the coldest winter nights and hottest summer days), which results in low average utilisation often around 20–50% in Europe and North America as these peaks occur just a few hours per year. To reduce the excessive CapEx spend that results from such overbuilding, we like solutions that can make the grid more visible to assess bottlenecks and enable flexibility. In general, we’re most excited about software-only orchestration solutions that will enable faster utility adoption and can be recovered as OpEx spend.

This is particularly interesting given that utilities in much of Europe and the US can now recover such investments in their regulated return rate, which is shifting from a CapEx only model to a total expenditure (TotEx) model that combines CapEx and OpEx in a single cost pool. Within grid orchestration, we’ve also assessed dynamic line rating (DLR) solutions as this space is relatively advanced. There are several mature players today so when considering an early-stage player we need to see exceptional traction with a DLR solution that is 5–10x better than incumbents, coupled with additional monitoring solutions (i.e., fire and ice detection).

Grid Expansion: It’s no secret that one of the biggest bottlenecks in clean energy deployment today is the interconnection queue, with it taking on average 5+ years for a project to connect to the grid in Europe and the US. Although regulation certainly plays a big role in slowing down the process, we’re interested in solutions that speed up time to connection by helping developers understand where the grid has available capacity and/or helping utilities process interconnection requests faster, often through interconnection automation and power flow simulation.

‍

We’re more passively considering the following levers and are more bearish on these investment opportunities, though we’re happy to be convinced otherwise:

Redesigned Chips: Although there are new CPUs and GPUs that promise significant energy savings (20–30%+), we find the climate angle to be blurred here. Since AI is limited not by chips, but by energy, and the AI game is a race to build the most powerful data centers, hyperscalers will likely plug in more chips while consuming the same amount of energy, rather than dialling down consumption. This brings us back to being more bullish on investing in solutions that ensure clean firm energy generation instead. We’re also not convinced that many startups can outcompete incumbents, though we do see players like NVIDIA becoming an acquirer in some cases.

System Optimisation: With the global PUE for data centers at 1.5 as of last year, we don’t see room for early-stage players to offer significantly better solutions to increase energy efficiency at the system level and thus offer something differentiated enough to secure large contracts with hyperscalers.

Enhanced Cooling: From what we’ve learned, it’s difficult for new entrants to get into conversations with data center developers at this point in a way that’s compelling, given the tried and true solutions already on the market today. It’s clear the market is moving to liquid cooling solutions, but within this space, NVIDIA is a gatekeeper that tells already conservative developers what type of cooling technologies are best for their chips, which makes operators unlikely to try alternative solutions given chips’ high price tags. Ultimately, new entrants today need to compete not only with legacy liquid cooling players like Schneider Electric and ASUS, but also need to have a significant edge against startups that were founded in the last 5–10 years.

Redesigned/Modular Data Center Developers: There are a range of options to rethink sustainable data centers from the ground up or retrofit existing data centers. The range includes infrastructure developers focused on equipping sites with clean energy, to prefabricated, modular data centers, all the way to moonshots like underwater data centers. Within this category, there are also companies focusing on waste heat utilisation, which is still largely unsolved in most data centers today. Over time, we’ve become more bearish on this space as we’ve yet to be fully convinced that companies operating within it are a true VC play. We see many exciting players, including Cloverleaf Infrastructure, who are making strides for decarbonisation, but have gone right to growth and are clearly an infrastructure case. So far, we’ve been most interested in players that are designing their solutions to be flexible grid assets, as this will likely enable faster interconnection.

—

Although there’s a lot of hype to sift through on this topic, we’re focused on doubling down on technologies that we’ve already known are critical for climate impact and can now enjoy the powerful tailwinds of data center decarbonisation. As mentioned, stay tuned for our first investment in grid-enhancing tech and please reach out if you’re working on a solution that we’ve missed — we’d love to chat!

‍