The Datacenter Is The Accelerator

There is a fractal nature in modern computing that is only becoming more and more apparent as workloads have long since outstripped the capacity of a single server. Just looking at the renderings of the massive datacenter that social networker Meta Platforms is building in Louisiana only makes this more obvious, and in a visually and architecturally satisfying way.

You can see it in the feature image above and in the top view of the expanse of the datacenter, which is being built to support AI workloads and which presumably will be stuffed to the rafters with GPUs and homegrown AI accelerators, shown below.

Doesn’t that datacenter look just like a modern CPU? To be specific, like a second generation AMD Epyc processor composed of a central I/O and memory controller tile with four banks of compute tiles that are coupled to it? In this case, there is no external HBM memory analog, which might be stored in skyscrapers on the periphery of the compute halls of this datacenter, but there is something on the upper right that could turn out to be storage of some kind.

The resemblance is not just uncanny, but absolutely canny. Centralized networking to provide coherence across the four compute halls, and wrapped around the outside are what appear to be transformers and power generators that are required to keep the megawatts flowing as power draw peaks as massive AI training jobs are fired up. This looks just like the capacitors on a modern GPU or other kind of AI accelerator, and perform an analogous function.

The feed and speeds and slots and watts of the datacenter that Meta Platforms is building in Richland Parish in northeast Louisiana were not divulged. What we do know is that the Richland Parish facility will be built between now and 2030, and will weigh in at 4 million square feet of capacity in its several buildings. Meta Platforms has 27 datacenters worldwide, with 22 of them located in the United States (not including this one).

In its 2024 sustainability report, which you can see here, Meta Platforms said it had 11,700 megawatts of contracted renewable energy across its datacenter portfolio in 2023, and has a total of 6,700 megawatts in the US. Just two days, when Meta Platforms was talking about starting to use nuclear power, it said the power consumption worldwide was now more than 12,000 megawatts, and said that its goal was to add 1,000 megawatts to 4,000 megawatts of new nuclear generation capacity to its US datacenters, starting in the early 2030s. Meta Platforms has clean and renewable offsets for all of the power used in its datacenters worldwide, but don’t be mistaken in believing that the datacenters are actually powered by clean and renewable power directly in all cases. That is not practical in all of the places that Meta Platforms places datacenters.

Exactly how much compute and storage will be housed in the Richland Parish facility remains to be seen, but we can get a sense of it in two ways. First, the amount of power being allocated, and second, the money being spent for the datacenter.

The northeast Louisiana facility, which is known as Franklin Farm and which has a total of 2,250 acres for future expansion, will be the largest one that Meta has ever built. Having all of those AI accelerators in one place greatly reduces the latency of the calculations that span those GPUs. The facility will apparently have three natural gas plants with a total capacity of 2,200 megawatts according to some rumors, but a statement from Entergy, which has been commissioned as the power supplier for the facility, says it is bringing in 1,500 megawatts of power and that Meta Platforms will have a match of 1,500 megawatts of clean and renewable energy as an offset.

Meta Platforms said the Franklin Farm bit barn will cost $10 billion, but most of that cost is going to be for accelerators and the hosts, storage, and networks that support them. If you assume that 90 percent of the cost of an AI facility goes into the IT gear, then it is $1 billion to build out the facility and $9 billion is allocated for gear. Say that a little more than half of the cost of the IT gear is for accelerators, now you are at $5 billion. At an average price of $25,000, that works out to 200,000 AI accelerators. If you assume this will be crammed with future homegrown MTIA accelerators that cost half as much, then you are talking $400,000 accelerators.

We suspect it will be the latter, not the former. Sorry Nvidia and AMD. Good news for Marvell or Broadcom, though.

Anyway, back to the fractal nature of this. The accelerator – whatever it is – is like a core, and a server is like a chiplet, and a rack is like a socket and a row is like a chassis to one way of looking at it. But really, maybe a datacenter hall is like a chiplet, and a row is like a CCD, and an accelerator is like an arithmetic unit, and a streaming processor is like a transistor. . . .

At the very least, a CPU socket today looks like a NUMA server from two and a half decades ago, writ very small. And a NUMA server looks like a federated NUMA supercomputer from the same time. It doesn’t just look like that — it is that.

Another observation. There is always much talk about how these datacenters stimulate the local economy, and with more than 5,000 construction workers, that is certainly true until the facility is built out and then opened in phases between a few years from now and 2030. Meta Platforms says the Franklin Farm facility will “support over 500 operational jobs,” but that does not mean they will all be on site. The parking lots, in the top center of the big building complex and then offset above that, look like they hold maybe 300 cars by our count. There are visitors and contractors that come in and do HVAC and electrical work eating up some of those spaces every day, we presume.

We presume those green areas surrounded by trees are ponds used for holding the intake and output water to cool the datacenter. Meta Platforms says it has a goal of being water positive by 2030, which means restoring more water than it consumes, and says that the facility will use less water than the site does today – presumably as a farm.

We would love to see the data on this bit of magic. Meta Platforms also says that it will invest $200 million in “local infrastructure improvements,” which Entergy says covers road and water systems, but the altruism of Meta Platforms here is obviously self serving to a certain extent. You need good roads to haul in zillions of truckloads of stuff as the datacenter halls fill up, and you need a good water supply to cool all this gear. We want to know where the sludge goes as the water is purified before it is used to cool the datacenter. At millions of gallons per day, which we figure this datacenter will need to cool, that precipitate from the input water will need to go somewhere.