Welcome back! Today is one of those newsletters in which I’ve learned just as much as what you’re going to learn yourself. And it also has a huge upside because it’s an article that will set a precedent for you about data centers for the foreseeable future.

Because to allow progress to continue, data centers will undergo a profound repurposing, especially regarding power equipment, the great forgotten bottleneck.

The deadline? As soon as next year.

This implies a huge amount of player and opportunity reshuffling, which means the lamest and most predictable part of the entire AI stack has just become exciting.

I’m going to tackle everything to describe where data centers are headed, the challenges that lie ahead, and the financial opportunities that emerge in the process, including the hottest new market emerging from AI and the company that, as of today, represents my next investment; a company that is equally risky as it has huge upside potential, probably the most out of all the ones I’ve talked about.

Let’s dive in.

As you surely know, we use accelerators (e.g., GPUs) to train and serve AI models. But to get those GPUs to do their job, we need a lot more stuff around them: memory, storage, CPUs, interconnect gear, and, of course, all the power-related equipment: VRMs, PSUs, busbars, sidecars, breakers, switchgear, converters, rectifiers, transformers, and many more.

Building a data center is not like building a warehouse; everything is connected. But until now, we were in easy mode, as AI has followed the traditional approach to building data centers.

But for the beasts that are arriving as soon as next year, current data centers are not enough, not even close; we need something else.

But why? Well, absurd amounts of power in very small places.

Just five years ago, the state of the art in accelerators was the NVIDIA A100 (the chip that, by the way, gave us the first ChatGPT).

These chips were located in servers in groups of 8, and each server required 6,500 watts at max power, less than what my burger joint’s fryer needs to operate.

This meant that you could house these chips in traditional data centers, which could allow between 5-10kW per rack (a rack is the chassis where the AI server, or other non-AI servers, are housed.

A data center offering 15-20 kW per rack was considered pretty dense at the time, before the AI revolution. However, things have changed.

Today, an NVIDIA GB300 NVL72, the most powerful AI server the world has to offer, can “offer” peaks of 150 kW, more than a tenth of a megawatt.

This has already implied having to build data centers from scratch just to be able to house servers of this size, and led to truly enormous data centers that can draw hundreds of megawatts and, in some cases, gigawatts, in order to house just a few thousand of these servers.

But wait, it gets worse.

With the 2027 server lineup, AI servers like NVIDIA’s Rubin Ultra or AMD’s MI550 series, we cross 500 kW per rack, reaching 600 kW per single AI server, with a clear line of sight to 1 MW servers before the end of the decade (Interestingly, China already crossed that line with the Huawei CloudMatrix, a 500kW beast. However, that is a corridor-scale server, not a single rack).

But before I explain to you why this completely changes the picture of how you approach building a data center, we must first answer: why, in the name of Jesus, do we need that much power?

In AI hardware land, you have two ways of “progressing”:

Not trying to linger too much on this topic, you have to think about accelerators as two workers in a factory. The memory worker feeds work to the compute worker, which performs the job and returns the result. It’s a back-and-forth between both.

Usually, you optimize one or the other, but in AI, you need both because, no matter how fancy Anthropic or OpenAI sound, we only really know how to progress by packing more compute and data. That’s the only play in the book.

This means that each accelerator has grown significantly in size along both axes, compute and memory. For instance, if we compare the current NVIDIA chip in production, the B200, to the 2027-slated Rubin Ultra chip, it’s literally half the size (approximately 2.5-ish more compute and 4-ish the memory required).

But what’s more striking is that the power required has more than doubled between these two chips (orange and dark blue below):

2,500W doesn’t sound like a lot considering your hairdryer requires that or more. The problem is that size is still literally the surface area of a little more than a credit card, generating incredible amounts of power… and heat, in a very concentrated place.

And when you factor in that transistors operate at less than 1 volt (~0.7 V), using Watt's law, we have 3,571 Amps of current flowing through this credit card-sized chip.

Not remotely comparable, but for reference, you only need 0.1 A, or 35,000 times less, and a stroke of bad luck, to kill a human, and your (~2,300 W) hairdryer requires around 10 Amperes to work in Europe (~20 in the US), only 3,500 times less than the numbers we’re throwing around here.

And once we start talking about logic-to-logic hybrid bonding, as Huawei intends to do, we can make double use of our GPUs as ovens, too. But I digress.

But having servers that require such large amounts of power and, indirectly, current means one thing: losses.

The power a data center needs has to come from a source. That source can be on-site, as most Hyperscalers would love to do, in order to avoid the grid altogether and avoid dealing with angry neighbors with higher electricity bills.

In reality, however, most of the data centers remain connected to the grid, and many new projects will still require grid connection:

This means data centers receive from an external source a huge amount of power at very high voltage, using alternating current, or AC.

But why AC? What is AC?

Most power plants are about heating something that vaporizes water. This water gets funneled through a turbine, which starts to move. At the end of the turbine, we have a rotor with magnets that, by default, create a magnetic field. Around those magnets, we have static wire coils.

As the turbine rotates, so do the magnets, creating a changing magnetic field around the coils. Using Faraday’s Law, the changing magnetic field induces a voltage in the wires, creating current, which is then sent to the world.

Interestingly, because each coil sometimes sees the magnet's north pole and sometimes its south, the current “changes direction” continuously, creating the alternating flow, giving us ‘alternating current’.

As mentioned, to minimize transmission losses, electricity is transmitted at very high voltages, so that the current is as small as possible. We’re talking about tens of thousands of volts, thousands of times more than what GPUs need.

Thus, we need to step down those voltages. For that, we use a transformer, a piece of equipment that hasn’t changed much in more than 100 years (until now, as we’ll see later).

This machine puts two coiled wires side by side, the second shorter than the first, and we apply the same intuition as in the generator; as the current changes direction, the change in current direction changes the magnetic field, which induces a current in the other wire (the ‘iron core’ helps keep the magnetic field constrained rather than spreading out).

In a step-down transformer, the second coil is shorter, so the induced voltage is smaller.

With this transformer (many of them in a sequence, actually), we gradually step down the voltage to around 54 V (you usually have other transformers in between), which is the legacy voltage we work with in normal data centers. Overall, the process looks a lot like this:

We don’t have to go deep with the intermediate steps like UPS/Battery and switchgear (the former is to ensure chips keep receiving electrons if the grid fails, the latter is about protection and isolation).

The important part here is steps 5 and 6.

Currently, that power equipment is inside the rack NVIDIA ships to customers, meaning the rack receives AC at a certain voltage, which is rectified (converted to DC) and stepped down once or even twice.

This power equipment that does this requires space in the rack, which means less space for compute.

For example, the Blackwell GB300 NVL72, NVIDIA’s most powerful server, and the one being deployed as we speak by your favorite data center constructor, requires 8 power shelves to handle the steps outlined above.

The issue? We’re basically reaching the limit. If we wanted to do the same with the 2027 Kyber Rack (the Rubin VR200 NVL576, also known as ‘Rubin Ultra’), you would need the entire server for power, literally, leaving absolutely zero space for compute.

We need an alternative. And we need it fast.