Below Compute Is Electricity

Founders often ask us: in this AI wave, are the companies doing power conversion at the table?

The view up front. They are, and their seat keeps moving up. Why? Because the bottleneck in AI is shifting from compute to power. We are fairly sure of this one.

The method for judging what will be scarce next in a cycle is crude: find the slowest ruler. Chips turn over by the quarter, data centres are built by the year, and one grid expansion often takes several years more. The slowest of the three curves is the ceiling. Equipment can be bought by paying more, but grid time cannot be bought. As things stand, the slowest is electricity.

Rack power in data centres is moving from a hundred kilowatts towards several hundred, and on to the megawatt class (figures illustrative). A megawatt-class rack draws roughly what several hundred households draw at once, all squeezed into one cabinet. The traditional power chain was designed for an age of steady loads: low voltage, many stages, losses at every stage. Compute iterates by the month; the grid expands by the year. So when we look at a new data centre, we are used to asking about the denominator first: how much power can this site get in total? The gap is not in technology. It is in the calendar.

The Ceiling Is Not the Chip, It Is the Ability to Move Heat Away

So how much further can power density be pushed? Look at cooling first. What sets the ceiling on a single rack is in the end neither the chip nor the power supply. It is whether the heat can be moved away. Every notch that thermal management moves forward, density immediately cashes in a new step. Power supply can be brute-forced with money, but heat is a physics problem. So for the next step in compute infrastructure, we are used to looking at the cooling route before the chip model. Reverse the order and the arithmetic comes out wrong.

The Middle Layer Is Being Squeezed from Both Ends

One thing is worth even more attention. Players on the energy side are working inwards, players on the rack side are working outwards, and both have their eyes on the power conversion layer in the middle. That is no coincidence. First, the middle layer's gross margin is on the thick side for the whole chain. Second, its technical barrier is being rebuilt by new devices. Third, it is the last piece each side needs to close its own loop. So will the company standing in the middle be eaten? It will. Unless it becomes the integrator first. The middle is not a safe zone.

In other words, the entry point of AI infrastructure is moving down: from the chip down to power electronics, and down again to energy. A new generation of power equipment is no longer a better transformer but a system that folds transformation, rectification, control, protection and software into one cabinet, which the industry calls a solid-state transformer, or SST, where solid-state means power semiconductors replace the iron core and the copper windings. A megawatt-class unit may hold more than two thousand power chips, with electric, thermal and magnetic fields and control all coupled inside one shell. You cannot make that by assembling parts.

So what does it take? Engineering capability tempered by mass production. Mass-production DNA means somebody has already paid the tuition on a million units somewhere else. When we look at a hardware team we like to ask a dumb question: where have you seen real mass production? The answer is usually in the CVs, not in the prototype. This skill can almost never be acquired quickly inside the industry itself. It has to be brought in by people. It is usually trained first in the automotive industry and then carried over to data centres. Automotive-grade mass production is the proving ground for this layer.

There is a quieter variable as well: incumbents are not necessarily willing to revolutionise themselves. Capacity, supply chains and customer relationships are all tied to the old architecture, and turning means writing off the profit already in hand. Challengers win by having no baggage, not by being cleverer. But this window will not stay open forever.

AI is the fuse; the upgrade of the grid and the energy system is the larger market underneath. The two sentences should be heard separately: the fuse decides the upside, the market underneath decides the floor. Say only the first and these assets get treated as cyclical equipment. Say only the second and the urgency of the moment is lost. Say both and the account is complete.

That is also why we put "compute, power and energy, optical interconnect and edge computing infrastructure" into a single direction — four things on one chain, not four separate targets. When you look at AI Infra, do not fix only on compute and optics. Look one layer further down. The business one layer down is quiet, but hard.