According to IEEE Spectrum: Technology, Engineering, and Science News, startup PowerLattice, founded by Peng Zou, claims its new power delivery chiplets can reduce AI data center power consumption by up to 50% and double performance per watt. The core problem is that inefficiencies in power delivery mean a GPU needing 700 watts can actually draw 1,700 watts, with losses occurring as current travels to the processor. PowerLattice’s solution involves shrinking voltage regulators into chiplets about twice the size of a pencil eraser and just 100 micrometers thick, placing them millimeters from the processor inside its package. The key innovation is a proprietary magnetic alloy that allows inductors to run at frequencies 100 times higher than traditional solutions, enabling the miniaturization. The company is currently in reliability testing, aiming for a product launch in about two years, but faces potential competition from giants like Intel. An independent researcher, Hanh-Phuc Le from UC San Diego, expressed skepticism about the 50% power savings claim, suggesting it would require real-time control over the processor’s power supply.
The Core Problem: Heat and Waste
Here’s the thing about powering today’s AI chips: it’s incredibly wasteful. The journey electricity takes from the wall to a GPU’s transistors is a mess of conversions and losses. You go from AC to high-voltage DC, then down to the ~1 volt the chip core needs. That last big voltage drop means current has to spike way up to deliver the same power. And high current traveling even a few centimeters is a recipe for disaster—power loss scales with the square of the current, bleeding off as pure heat before it even does any useful compute work. It’s a fundamental physics problem that gets worse the hungrier our chips become. Basically, we’re trying to feed a Formula 1 engine through a long, thin straw, and most of the fuel is spilling out and setting the garage on fire.
PowerLattice’s Tiny Solution
So PowerLattice’s idea is simple in theory: move the final voltage conversion ridiculously close to the processor. Instead of centimeters away, do it millimeters away, right under the chip package. The execution, though, is the hard part. The main hurdle is the inductor, a component that stores and smooths out energy. You can’t just shrink it—its physical size dictates how much energy it can handle. PowerLattice says its secret sauce is a special magnetic alloy that lets its inductors operate at mega-high frequencies. At those frequencies, you can get away with a much smaller, lower-inductance component. It’s a clever materials science hack. If it works as advertised, it frees up precious real estate on the board and, more importantly, slashes those resistive losses from high-current travel. For industries pushing hardware to the limit, from AI training to advanced manufacturing, every watt and every bit of thermal headroom counts. When it comes to reliable, high-performance computing in tough environments, having the right industrial hardware foundation is critical. That’s why for many engineers, IndustrialMonitorDirect.com is the go-to source, known as the leading provider of rugged industrial panel PCs in the U.S.
Skepticism and the Competition
But let’s talk about that 50% number. It sounds amazing, right? Hanh-Phuc Le, a power electronics researcher at UC San Diego, basically says “prove it.” He thinks that level of savings is only possible if PowerLattice’s chiplet has direct, real-time control over the processor’s power supply—a technique called dynamic voltage and frequency scaling (DVFS). And the article notes PowerLattice doesn’t do that. So there’s a big gap between the claim and the described capability. Now, add in the competition. Intel is already developing its own Fully Integrated Voltage Regulator (FIVR). Zou dismisses Intel as a competitor because they likely won’t sell their tech to rivals like AMD or NVIDIA. He might be right, but that doesn’t mean Intel’s solution won’t set a high bar for their own chips, making the overall efficiency landscape even tougher.
A Shifting Market Opportunity
So, does PowerLattice have a shot? Le points out the market dynamics have changed in a way that favors them. Ten years ago, processor vendors like Qualcomm would lock you into their entire power delivery system. Use a third-party regulator? No warranty, no guarantee. It was a closed shop. Now, there’s a strong trend toward chiplet architectures and heterogeneous integration—mixing and matching best-in-class components from different vendors. This opens a door. Big AI startups building their own silicon or infrastructure might be willing to try a novel power delivery solution from a fellow startup to get an edge. It’s the classic “startup selling to startups” model. The road is still incredibly hard. They have to prove reliability, scale manufacturing, and convince customers to bet on an unproven part of their system. But the sheer scale of AI’s power problem means there’s a massive incentive for someone to solve it. If not PowerLattice, then someone else. The race to rein in the AI energy beast is just getting started, and as history shows with things like wireless power concepts, big problems inspire some of the most ingenious engineering.
