The Power Stack Delivering the Next Trillion in Compute

 Photo courtesy of Celesta Capital 

AI infrastructure spending is on track to reach $600B this year. Much of the conversation is already about chips, photonics, and networking. Far less attention is being paid to the infrastructure and deep tech needed to power it all.

I spent a recent Tuesday afternoon at Deep Tech Week in San Francisco, at the Celesta Capital session: Battle for the Data Center: The Hardware Powering the Next Trillion in Compute, listening to panels alongside leaders from Intel, SambaNova, Coherent, Lumentum, Eliyan, Upscale AI, and Velaura. Intel CEO Lip-Bu Tan also joined Michael Marks for a fireside chat. 

The investors on stage — from Cota Capital, Maverick Silicon, and Intel Capital — were unusually candid about where they see the next decade of value capture in compute. Talk of GPUs and photonics was everywhere, but one hardware/deep tech issue that was perhaps under the radar, but well discussed throughout sidebar conversations in the room: power. 

There is More than One Power Problem Gating Data Centers

When the industry talks about the power problem in AI data centers, it usually means two things: there isn't enough of it, and it takes too long to get connected to it. In several major data center markets, grid connection queues now exceed four years. That alone is reshaping where capacity gets built and how quickly operators can bring it online.

Buried inside those two issues is a third one that often gets considerably less airtime: how power actually gets delivered to the chip, in real time, at the speeds AI workloads now demand. This happens through multiple layers of a Power Stack across many different hardware systems. Without an effective Power Stack, everything else becomes “stranded investments” that quickly depreciate.

Training and inference don't draw power smoothly. They ramp racks of GPUs from 30% of rated capacity to 100% within seconds, sometimes within milliseconds or less. The grid wasn't engineered for that. Traditional backup systems weren't either. Substations get stressed, equipment gets damaged, and operators end up over-provisioning generation just to ride through the swings. 

That's where energy storage shifts from backup infrastructure to something the Power Stack architecture is actually designed around. Morgan Stanley recently called battery storage "a system-level necessity", and the Power Stack is where that volatility gets absorbed. 

Conventional BESS isn't the Answer

By 2030, data centers are projected to make up 83% of behind-the-meter commercial and industrial BESS deployments. However, that deployment will not be done with conventional or legacy BESS platforms. Most battery energy storage systems on the market today were architected for a different problem: utility-scale arbitrage, solar firming, behind-the-meter peak shaving. Their established design choice is to push every joule that comes out of the batteries through a power converter. 

That works, but each converter is expensive, and each conversion step has an efficiency cost. Stack enough of them in series and the inefficiencies compound across the system. There's also an architectural rigidity. 

Conventional BESS designs assume a specific battery chemistry, a specific pack format, and a relatively narrow set of use cases. For a data center operator who needs both peak-shaving and millisecond-scale buffering, and who would like to deploy new lithium cells today while keeping the option to integrate second-life packs as supply scales, that rigidity is a meaningful constraint.

Traditional Power Stacks have multiple power components that transform voltage, direct power flow, clean up the electricity, protect equipment, and provide energy storage.  These distinct systems from different manufacturers have to be orchestrated and be compatible.

This is Why we Built Volt Harbor

Most BESS platforms convert every watt that leaves the battery. We built MAC-BESS™ around a different premise. There is a fundamental tradeoff between energy storage and the size of power electronics. You don't need to convert every watt that leaves the battery. You need the right converters in the right places, coordinated by software that understands the Power Stack functionality and the heterogeneity of the underlying batteries.  

In today’s Power Stacks, power is routed by mechanical switchgear and high voltage is transformed by steel transformers. Fast UPS backups remain separate from large-scale battery storage. Multiple distinct power electronics convert DC to AC and vice-versa, condition “dirty” electricity, and transform the voltage some more. This delays installation and commissioning. This makes the operation and maintenance more expensive.

Our Medium Access Control architecture — patented and modeled on computer networking — uses identical low-power converters and embedded computing together with potentially different batteries as building blocks to integrate much of the Power Stack functionality in a single turnkey unit. The energy storage isn’t the system, it’s the enabler.

In peer-reviewed testing, that approach can deliver as much as 94% energy utilization, compared with about 78% for conventional partial-power processing and 23% for full-power processing. We also generate AC directly from the battery side, eliminating a discrete DC-AC inverter stage and bringing overall power-conversion cost to roughly 10% and 20% of a conventional system.

Where the Trillion Gets Captured

The Celesta panels kept circling back to a single question: where in the stack does the next trillion dollars of compute value get captured? A lot of the smart money on stage was being placed on silicon, photonics, networking, and advanced packaging. Every one of those bets deserves the attention it's getting.

The Power Stack is still up for grabs. And the operators I talked to know that adding capacitors or generators is just buying time. What actually solves it is treating storage and power electronics as a single system, architected around how these workloads actually behave. 

That is what we are building at Volt Harbor. And walking out of Deep Tech Week, I was more convinced than ever that this layer is going to be one of the decisive pieces of the trillion-dollar AI infrastructure buildout.