For years, the AI arms race was all about who could cram the most GPUs into a rack. Now the real bottleneck isn’t compute — it’s electricity. And in this new phase of the AI boom, the quiet winners aren’t the cloud giants stacking servers, but the companies selling the power chips that keep those racks from melting down.
This is the heart of AI’s power race: as data centers push past old limits, the focus is shifting from bigger models to better power delivery. And that’s where next‑generation power semiconductors are starting to matter more than the hardware they live inside.
The AI power race moves from GPUs to volts
The numbers behind this shift are brutal. Power consumption per rack for Nvidia’s Hopper generation sat around 40 kilowatts. With the newer Blackwell platform, that figure has already jumped to roughly 120 kilowatts per rack.
And that’s just the beginning. Next‑generation AI systems are expected to demand between 600 kilowatts and as much as 1 megawatt per rack — comparable to the electricity used simultaneously by about 1,000 households. At that scale, power isn’t a line item on a data center budget; it becomes the defining constraint.
The traditional 48‑volt architecture that has underpinned server power delivery is straining under those demands. As you push more current through it, transmission losses spike, heat builds, and efficiency tanks. That’s why an 800‑volt high‑voltage direct current (HVDC) architecture is increasingly seen as not just attractive, but practically unavoidable for future AI data centers.
Why power semiconductors are the new AI gold rush
On paper, you might expect this to spark an explosion in spending on transformers, switchgear and other chunky power equipment. But that’s not what’s happening. Large cloud providers are tightening capital expenditure, even as they scramble to scale AI infrastructure. They’re hunting for efficiency, not just more steel and copper.
That’s where power semiconductors step into the spotlight. Inside each power supply unit (PSU), dozens or hundreds of chips quietly handle voltage conversion, switching and control. As architectures move to 800 volts and thermal and efficiency margins shrink, these parts get far more complex — and far more valuable.
Wide‑bandgap semiconductors like silicon carbide (SiC) and gallium nitride (GaN) are built for exactly this environment. They can withstand higher voltages and temperatures than conventional silicon, switching faster while wasting less energy as heat. That makes them ideal for AI racks that will soon be sipping power at 800 volts instead of 48.
Crucially, these advanced power semiconductors are more expensive than the legacy silicon components they replace. That means that even if overall spending on heavy power equipment grows modestly, the value captured by chipmakers embedded inside that hardware rises much faster.
Follow the bill of materials, not the megawatts
A concrete example makes this clearer. One major AI power supply leader, Delta Electronics, has already seen what next‑gen AI demand does to costs. For its upcoming 12‑kilowatt PSU, the component cost per watt is up by around 43 percent compared with its current 5.5‑kilowatt model.
That spike isn’t coming from metal boxes or fans. It’s largely driven by a sharp increase in the amount — and sophistication — of power semiconductors baked into the unit. More power per rack at higher voltages forces PSU makers to stuff in more SiC and GaN devices, plus smarter control chips and higher‑end substrates to carry the load safely.
If you zoom out, the pattern is simple: every new wave of AI servers demands more and better power supply units, and every new PSU design packs in more high‑value silicon. Demand for the boxes themselves grows, but demand for the advanced chips inside them grows even faster.
For anyone trying to understand where the real profit pools sit in AI infrastructure, the lesson is obvious: follow the bill of materials, not just the megawatts.
Near‑term vs. long‑term winners in AI power chips
This doesn’t mean every part of the power semiconductor stack wins at the same time. The roadmap for AI platforms suggests a staggered payoff.
Nvidia’s next‑generation AI platform, due in 2026, will still lean on the familiar 48‑volt architecture. That keeps most of today’s board‑level power management designs in play a little longer. Companies providing control ICs, regulators and other chips that orchestrate power on the GPU board itself are likely to see their earnings lift first as those systems ramp.
The more radical shift — a full move to 800‑volt HVDC — is expected to kick in with platforms launching after 2027. The real earnings impact from that architectural change should start to show up around 2028, as deployments scale and high‑voltage designs become the norm rather than the outliers.
That staggered timeline creates two distinct opportunity windows:
- Short term (through 2026): Strength for suppliers of board‑level power management chips tuned to 48‑volt systems.
- Longer term (2027 and beyond): A bigger payoff for makers of wide‑bandgap power devices, high‑voltage control chips and premium substrates designed for 800‑volt data center architectures.
Across both periods, the direction of travel is the same: more value migrating into silicon.
Why power equipment makers may lag behind
So where does this leave traditional power equipment suppliers — the companies building the racks of gear that connect data centers to the grid?
They will still grow with overall AI demand, but the outlook is far less explosive. Cloud providers are intent on keeping the cost of power infrastructure per megawatt under control. That means sweating efficiency gains out of each generation of equipment instead of simply paying more for every new watt delivered.
In practice, that caps how much prices for big iron — transformers, distribution units, backup systems — can rise, even as AI workloads balloon. Makers of those systems will be pushed to deliver more capacity at similar or only slightly higher cost per megawatt.
Inside those same boxes, though, advanced chips are quietly taking a larger slice of the cost structure. Every time a design team shifts from silicon to SiC or GaN, or adds more sophisticated control silicon to hit efficiency targets, the semiconductor content per watt climbs.
That’s why analysts increasingly argue that investors should focus more on power semiconductor manufacturers than on the branded power equipment vendors. The brains and brawn hiding on the circuit boards are where more of the incremental dollars are going.
The invisible backbone of AI data centers
It’s easy to get distracted by GPU benchmarks and parameter counts, but as AI racks race toward the megawatt era, power semiconductors are becoming the invisible backbone that makes it all viable.
Manufacturers in three areas look especially well positioned:
- Board control chips: The ICs that manage voltage regulation, switching and safety on AI accelerator boards, particularly in 48‑volt systems shipping over the next few years.
- Power semiconductor devices: SiC and GaN transistors and diodes built to handle 800‑volt operation efficiently in high‑density data centers.
- High‑performance substrates: Advanced materials that can reliably carry high voltages and currents while managing heat in ever‑tighter spaces.
These aren’t the glamorous parts of the AI stack, but they’re becoming essential. Without them, the industry can’t safely or efficiently move from 40‑kilowatt racks to 1‑megawatt monsters.
What this means for the next phase of AI
The AI boom is entering a harder, more physical stage. It’s no longer just about abstract models in the cloud; it’s about voltage, heat and the cost of every extra watt. As racks leap from tens of kilowatts to the edge of a megawatt, the power architecture under the floor and on the boards becomes a first‑order concern.
In that world, power semiconductors are set to punch far above their historical weight. They won’t be the headline names on AI press releases, but they will quietly decide which data centers can scale, which cloud providers keep energy costs in check, and which investors picked the right side of the power race.
If phase one of the AI era belonged to GPU makers, phase two increasingly belongs to the companies mastering power delivery at 48 volts today and 800 volts tomorrow.
What This Means
The real winners in AI’s power race are shifting from the builders of visible infrastructure to the makers of the chips that make that infrastructure possible. As data centers chase 800‑volt efficiency and megawatt‑class racks, the value of every watt is going up — and so is the value of the silicon that controls it.
Photo: National Museum of American History / Public domain via Wikimedia Commons | Photo: Gage Skidmore from Peoria, AZ, United States of America / CC BY-SA 2.0 via Wikimedia Commons
