The surge in AI workloads has driven massive demand for compute infrastructure – from GPUs to new cybersecurity standards. But no part of the stack has been hit more urgently than data centers. Hyperscalers are continually constrained by access to power (and the ability to transmit it) as they scale to meet demand, provoking a flood of startups and new solutions vying to solve one of the most pressing challenges of the AI age.

To dig into these constraints and the innovations reshaping them, we sat down with Phill Lawson-Shanks, Chief Innovation Officer at Aligned Data Centers and 53 Stations Wayfinder. With decades of experience building hyperscale environments, Phill has helped power the internet as we know it – and is now helping re-architect the systems that will power what’s next. 

Read on for five key takeaways from the conversation.

1. Transmission is the Biggest Bottleneck in AI Infrastructure

“The problem isn’t generation, but distribution,” Phill told us. “It’s about getting power from where it is to where it’s needed.”

Take Ashburn, Virginia, home to the largest concentration of data centers on the planet. The grid there, like many others, can’t keep up with a spike in demand that started during the dawn of COVID lockdowns. Phill described how hyperscalers like Microsoft and AWS accelerated the crisis by suddenly activating massive power commitments. 

One example: in 2020, Microsoft Teams saw a 775% increase in cloud services in just one week and reached 75 million daily users shortly after. Companies exercised dormant rights to additional power, overwhelming utilities that had thin-provisioned for slower ramp-ups. The spike exposed just how outdated our transmission systems are.

Phill’s not alone in sounding the alarm. The U.S. added only 386 miles of new high-voltage transmission lines in 2023, down from over 1,700 miles/year a decade ago. Projects can take 10-15 years to complete, provoking many companies to enter the field timing to use existing rights-of-way and replacing older cables with higher-transmission lines.

2. Nuclear Innovation is Outpacing the Systems that Govern it

Phill sees promise in new small modular reactors (SMRs) and sodium-cooled designs that are walk-away safe and can even reuse spent fuel. “They’re ready. They’re safer, smaller, and smarter – but they’re stuck in a slow, paper-based certification process,” he said.

The certification process remains slow and manual, but change is underway. Microsoft is partnering with the Nuclear Regulatory Commission (NRC) to deploy AI tools that can help automate the agency’s document-heavy licensing workflow. 

Meanwhile, momentum is building: more than 25 small modular reactor (SMR) designs are now in the pipeline. In January, NuScale became the first SMR in over a decade to receive NRC design certification, and commercial deployment is expected by 2029. To accelerate that timeline, the U.S Department of Energy has pre-certified 16 nuclear sites across the U.S. and hyperscalers are already moving to secure land near these future sources of generation.

3. Data Centers are Reaching Megawatt Density

To handle massive AI workloads, today’s racks are hitting power levels that would’ve once served entire facilities. While traditional cloud racks handle 10–50 kW, new AI-dedicated racks are hitting 250-600 kW and are projected to top 1 MW soon. “We’re already designing for 1.2 megawatt racks,” Phill told us. “They’re more like machines than buildings.”

Power distribution is shifting from 220V to 415V, and eventually to high-voltage direct current. Liquid cooling is becoming table stakes, not a nice-to-have. And in many cases, operators are moving away from massive centralized builds toward smaller, modular deployments that can deliver dense compute closer to metro areas – within 10 kilometers of a cloud on-ramp, as Phill noted. It’s not just more power, but an entirely new architecture.

4. AI is Rewiring the Data Center Build Process

For all the precision in data center operations, the construction process often runs on guesswork. Visibility into schedules, supply chains, and labor capacity is limited – and surprises get expensive fast. Phill explained how better tooling is helping teams move from reactive to proactive. (This is a shift 53 Stations is betting on, too – our portfolio company Kaya AI is bringing the same kind of data discipline to procurement.)

These intelligence platforms, originally developed for megaprojects in Europe, are now reshaping how data centers get built – flagging material shortages, schedule conflicts, and labor mismatches before they cost millions.This kind of operational intelligence aligns with broader trends we’re tracking around how infrastructure is planned and delivered.

5. Sustainability will be an Infrastructure Advantage

Energy demand is rising, but sustainability pressures haven’t gone away. Phill pointed to sodium-ion batteries as one answer – safer, cheaper, and made from materials 500x more abundant than lithium.

He’s already watching companies like Natron, which recently announced a $1.4B U.S. gigafactory in North Carolina. The facility is expected to produce 600MW of batteries annually, with no thermal runaway risk and a cycle life twice that of traditional lithium-ion. Phill noted that technologies like this are gaining traction as data centers look to decarbonize backup power, which has long relied on diesel gensets.

“Our clients haven’t stopped tracking every gram of carbon,” Phill said. “They’re just not talking about it as much.”

The Bottom Line

Infrastructure is the hidden layer shaping the future of AI, cloud, and compute. Founders operating in this space – whether they build software, workflows, or materials – need to know where the pressure points are and what’s coming next. At 53 Stations, we’re actively investing in this future. If you’re working on challenges at the intersection of AI, power, and industrial systems, we’d love to hear from you.