Edition #3. Every Tuesday, I write about the AI supply chain constraints you need to know about, and which companies are benefiting.

For the last year, the AI power debate has mostly lived outside the data center.

PPAs. Substations. Transformers. Gas turbines. Nuclear restarts. Interconnection queues.

That layer still matters. If you cannot energize the site, nothing else matters. But once you look inside the data center and at a rack level, there is another constraint forming.

Once electricity reaches the data center, it still has to be converted, switched, protected, cooled, and delivered to the GPU. That job belongs to power semiconductors.

And as AI racks move from tens of kilowatts toward hundreds of kilowatts, and eventually megawatt-class designs, the power tree is becoming a semiconductor problem.

We covered the grid-side power constraint in Edition #1 and the photonics chokepoint in Edition #2. The semiconductor problem inside the rack is the next layer down.

Power semiconductors are basically ultra-fast electrical switches.

They turn electricity on and off thousands or millions of times per second to convert it.

From AC to DC
From high voltage to lower voltage
From unstable power to stable power
From building-level power to chip-level power

An AI data center has many conversion steps, and needs a chain of them between the grid transformer outside the building and the GPU inside a rack.

Medium-voltage AC comes in from the substation, gets rectified to DC, stepped down through the data hall, dropped again at the rack, and finally regulated to the ~0.7V the GPU needs.

Each stage uses a different type of power semi, with high-voltage modules (1.2 to 10kV+) at the front, mid-voltage (650V to 1.2kV) in the rack, low-voltage devices at the load.

Now the rack itself is rebuilding. Today's AI rack runs ~120 kW at 48V DC. The next generation, which is already in NVIDIA's reference design and being deployed at every major hyperscaler, runs at 800V DC distribution.

Higher voltage means lower current for the same power, which means less copper, less heat, dramatically smaller conversion losses. That matters when a single rack pulls 600+ kW and a campus burns 500 MW.

The catch is that at 800V, the silicon parts that ran the 48V rack don't work well enough. Switching losses get too high and thermals become unmanageable.

That's where these two chemical compounds come in:

These are materials that switch faster, run hotter, and lose less energy in conversion. SiC owns the higher voltages (650V+). GaN owns the lower-voltage, faster-switching sockets inside the rack.

Net effect: the architecture transition rebuilds the dollar value of every layer of the power tree, and most of the new dollars flow to SiC and GaN.

ON Semi put a number on it this quarter: "At the 120-kilowatt rack today, roughly $9,500 of content. At the 800-volt or high-voltage rack, roughly $115,000 of content." It can sell ~12x worth of components on the same compute.

And the demand for these power semis has already started to show up. Three power semi companies that had their earnings in the last couple of weeks show this:

So, yes, the demand is starting to grow but it is not all smooth sailing. A few supply constraints exists and it bites depending on which voltage tier you're looking at. A few include:

None of these bottlenecks are binding hard yet. But each one tightens with every quarter of 30%+ AI growth, and a few of them are gated by clocks that can't be sped up once the demand catches them.

Watch qualifications more than order books. In power semis, the real signal is when designed-in content turns into shipped volume, because the highest-content sockets often stay constrained by long qualification cycles before they can ramp.

Here’s how the constraints moved in the other layers:

DRAM memory, particularly DDR5 and LPDDR5, tightened further this week as agentic AI continues to push the demand for server CPUs.

The CPU-to-GPU ratio captures the shift. Training racks used to run 1 CPU for every 8 GPUs. Inference dropped that to 1:4, and agentic workloads are pushing it toward parity and beyond. Intel CEO Lip-Bu Tan, on Cramer this week: "It used to be the CPU to GPU ratio for training was 1-to-8. And now, because of inference and agentic AI… that becomes 1-to-4 and 1-to-1, and some people even talk about 4-to-1." DDR5 DRAM is the main memory type used by both AMD and Intel CPU servers.

NVIDIA's Vera CPU rack runs on LPDDR5, which is the same DRAM that goes into smartphones and given the increase in demand, NVIDIA is set to overtake the largest smartphone OEMs and become the world's single biggest LPDDR buyer. Citrini Research projects 2027 AI CPU LPDDR demand at ~6,000 million GB, more than Apple (2,966M) and Samsung (2,724M) combined.

A few important earnings this week. Read our pre-call briefs here and stay prepared.

Want the live data behind The Chokepoint? Tessara is the AI infrastructure operating system for investors. We track what's binding in the supply chain and what it means for what you own. 270+ companies across compute, memory, foundry, networking, and power. Apply for early access →

See you next week,

Teng & Arvind

This article is for informational and research purposes only. It is not financial advice, investment advice, or a recommendation to buy or sell any security. Tessara Research does not publish price targets. The views expressed here reflect our analysis at the time of publication and may change as new evidence arrives. Readers should do their own research and consult a qualified financial adviser before making investment decisions.