In Northern Virginia’s data center alley, the world’s densest cloud hub, it’s becoming clear that there simply isn’t enough power to go around. The U.S. Department of Energy estimates that by 2030, the area could face more than 430 hours of power outages as the power grid struggles to keep up with the growth of AI and data centers. The current rate is 2.4 hours of power outages per year.

This isn’t an isolated problem. Around the world, appetite for AI and cloud services is ballooning quicker than existing energy infrastructure can keep up. While the power squeeze is making headlines, most of the attention is paid to cooling innovations and grid upgrades – solutions which address the symptoms, not the cause.

The real culprit at the heart of this crisis is the inefficiency of the hardware itself.

The scale of the energy crisis

Data centers face a double challenge. First, there’s not enough clean energy being generated fast enough to meet demand. Second, even when renewable power is available, it often can’t be delivered to where it’s needed. For example, wind energy generated in Scotland can’t easily reach London’s data hubs due to limited transmission capacity.

This has resulted in an energy catch-22: there is a desperate need for more energy, but the energy we are generating can't be used effectively.

According to Bloom Energy research, there’s now a 12- to 24-month gap between when data centers need power and when the grid can actually deliver it. This is obviously not good enough. In response, countries such as Ireland, where data centers consume more than 20% of total electricity, are considering allowing data centers to build their own private power lines, while the U.K. is relaxing planning rules for new pylons. Around the world, there are clear signs that transmission, not just generation, is now a major bottleneck.

Managing the symptoms, not the cause

To cope, operators are turning to ever more creative solutions: immersion cooling, district heating systems, and even underwater data centers. Some are building on-site power generation, but this approach is hampered by wildly fluctuating demand – the power usage of AI training with GPUs can vary by 50% within just a few minutes. These are clever fixes, but they only treat the symptoms, not the root cause.

The real issue is that the hardware powering today’s AI and cloud workloads is astonishingly power-hungry. Power density is soaring: Nvidia now projects racks consuming up to 250 kW each, with this set to rise to as much as 900kW each, but most existing data centers weren’t designed for this. As a result, many facilities are underutilized or simply can’t expand, no matter how much demand there is.

A different approach: Shrink the problem

Instead of asking how we feed the beast, we should be questioning how we can shrink it. What we need is a new generation of energy-efficient computing and networking infrastructure. One promising solution is all-optical networking, which slashes energy use by tackling the fundamental inefficiency in data transmission.

Right now, many networks rely on electro-optical switches. These work by converting light to electricity, conducting their switching within the electronic domain then converting the signal back to light for onward transmission. While this approach has worked well so far, at scale it demands huge amounts of energy and is subject to high latency penalties.

But now, there’s a solution. Fully optical packet switching routes data entirely in the optical domain, eliminating electronic conversions. By using nonlinear photonic crystals and encoding routing information onto a control light beam, destination addresses are imprinted directly onto data streams. This enables passive, energy-efficient packet switches, which operate at transceiver speeds, removing key network bottlenecks.

This innovation is a rethinking of how data moves inside data centers, enabling lower latency, higher throughput, massive bandwidth, and significantly reduced power consumption.

Economic and competitive stakes

Today, access to power is the new land grab. There are cases of old cement works being converted to data centers simply because they have a 50 MW grid connection. For operators, cutting power needs by a factor of 10 could unlock growth opportunities that are currently blocked by grid constraints. In this new reality, energy efficiency goes beyond the domain of sustainability to become a core competitive advantage.

However, this desire for ever-expanding grid capacity must be integrated with making data center technology fundamentally more efficient. New technologies have now emerged that slash energy consumption while increasing performance, and this should be promoted and prioritized by industry and government to ensure we don’t reach the physical and economic limits of legacy digital infrastructure.

If we want AI and cloud computing to keep growing, we must stop accepting inefficient hardware as the status quo. Solving energy inefficiency at the source, rather than endlessly treating its symptoms, is the only way forward for radically efficient computing. It’s time for industry and policy leaders to prioritize smarter, leaner technology that can power the future without overwhelming the grid.