The construction of data centers are a contentious issue because operating them consumes an inordinate amount of water and energy. A working data center may place a strain on the local power supply, raising the cost of electricity for everyone. They can strain the local water resources as well leading to problems for residents. A facility in Georgia allegedly took 29 million gallons of water without authorization over a period of 15 months and it wasn’t detected until nearby residents complained of a lack of water pressure. There are 50,000 residents in Tahoe who will lose power in May 2027 because their providers have decided to serve data centers for Google, Apple, and Microsoft.
These are just the short term impact of data centers, how bad could it get in the long term? Will anyone do anything about it?
Both of those things are really regulatory failures more than any sort of inherent evil with data centers. Right now they use inordinate amounts of water and electricity, because those are the cheapest ways to do it… because there are very few rules or regulations preventing or limiting it.
I suspect if states and localities started passing laws that said that only purified wastewater could be used, or drew thresholds and conditions on electrical use, they’d raise the efficiency tout suite on both things. Or go places where neither is in short supply.
The article doesn’t say it wasn’t authorized, it says the county lost track of the connection because it fucked up an IT issue. That doesn’t really have anything to do with data centers. 29 million gallons over 15 months really isn’t that much water. It’s about 230 households’ worth, in a county with over 100,000 people.
The water usage of data centers always seems massively overblown and driven by people thinking all numbers with more than 6 digits are equally huge (and you’re guaranteed to get huge numbers when you measure water in tiny human-sized units).
A data center is going in about 10 miles south of me, so I’ll let you know how it is once it’s up and running.
At present, the mere construction of it is blocking and cracking roads with continuous transit of heavily laden dump trucks and equipment hauling devices, and the noise of 24 hour construction (including floodlights all over the 170 acre site) has been driving the locals crazy. Finally the company agreed to only work 12 hours a day instead of 24, and only 8-10 hours on weekend days.
I don’t know how accurate it is, but I’ve read that once they finish building them and provisioning them, about the only jobs for the locals are for security guards and some maintenance. I assume they would at least have a manager for the data center, but it seems likely those managers aren’t going to be from the local community.
I worked down the hall from a compute ranch, which is a relatively small data center (thousands of CPUs.) Big enough that we needed our own power line. Relatively few people worked there, and we weren’t in competition with McDonalds to get them. So the local government is going to be sorely disappointed if they expect a long term hiring surge.
I drive by there every week or two, going back and forth from my house to my parents’ house. It’s adjacent to the town where my mother grew up.
It is an enormous project: it’s right alongside the interstate, being built on a stretch of what was previously farmland. It’s easily over two miles from north to south, and the number of construction machines is staggering.
South Memphis is home to Elon Musk’s XAI Colossus data center. There’s the first problem. The city’s leaders trusted Elon Musk. The most immediate problem is dealing with pollution in the area. Just across the state line in Mississippi, Musk has installed at least 46 methane gas turbines to run the center. These were illegally installed without federal air permits that would require effective pollution controls. Of course, Musk said this was temporary - two years ago. The unpermitted turbines spew toxic nitrogen oxides, linked to asthma and other breathing illnesses, and formaldehyde, linked to cancer. Lawsuits have recently been filed to try to shut them down.
Last summer he said he was going to buy a power plant overseas, transport it to Memphis and rebuild it to provide additional energy to run Colossus. No word on how that plan is coming along. He has also promised to build a greywater recycling facility. That plan has been halted.
Only surprising to the city’s leaders, Musk hasn’t lived up to any of his promises. Our electrical grid and water system are at serious risk. In the current environment, I don’t expect him to be held accountable for anything and the city is going to suffer immensely.
On that Georgia article, it’s helpful to note that the water usage was from the construction, not the actual datacenter running:
QTS told Politico the 29 million gallons were consumed during temporary construction activities, including concrete work, dust control, and site preparation. The company markets a “closed-loop” cooling system for its data centers, which recirculates the same water rather than drawing from the municipal supply. Once operational, QTS said its facilities would only require water for domestic needs like bathrooms and kitchens.
If Moore’s law holds up, the enormous power and cooling requirements should be temporary. The first step of AI is to build a computer with number crunching capability equal to the human brain. We seem to be just barely there. Historically, that capacity has been doubling every few years since the 1960’s. Back in my days there, the UC Berkely campus had a computer shared by all of the undergraduates. It ran at somewhere around 1 MHz and could address 1.75 megabytes of core memory. The computer in our telephones today beats that by five or six orders of magnitude.
Moore’s law put a number on that exponential growth in capacity. It will have to reach a limit someday but, it does not seem to be just yet.
Here in Missouri, there’s one issue that seems to unite everyone, even the ones who aren’t knee-jerk opponents:
Who actually will own and operate these things?
Representatives come into an area, buy options on land (sometimes they skip that step and just walk into a meeting with county authorities), sign the landowners to NDA’s, then walk in to county authorities with a bunch of paperwork listing shell companies as owners. Who’s behind them? Microsoft, Apple, Amazon, China?
Moore’s Law is not a true law but an observation. And it does not apply to only silicon. Silicon has slowed down because of economic issues, but there are now chiplets, smaller chips connected more efficiently than the usual way on boards, which let compute capacity still double.
Adding processors to naturally parallelizable tasks such as AI is not going to stop any time soon.
I’ve seen Moore’s Law traced backwards in time before the IC, before even the computer.
Plus, Moore’s law at one point gave guidance to people planning new technology. I was in a meeting one where this explicitly happened, and the Roadmap for Semiconductors certainly used it as a starting point.
It is an extrapolation of history into the future. As an exponential growth, it will have to end at some point. It has lasted longer than people expected. The improvement in silicon and possibly other semiconductor IC’s is due to continuing reductions in line width. Smaller transistors take less power and you can fit more of them on a chip.
Packaging improvements like multiple chip packages and chiplets to let us fit more transistors into a given space but don’t help all that much on power consumption.
The progression is not necessarily linear. The technology industry goes in cycles. It has its ups and downs. Each boom seems to create a new batch of companies. Some of them last more than one cycle and grow big. Even the big ones can run out of steam though. Intel has been kluging the 8080 CPU for about 50 years now but those optimized for graphics like Nvidia’s seem to be a better match for the requirements of AI. Microsoft seems to be past their peak.
The final end to Moore’s law may be economic. We don’t seem to be there yet though. There may be a point where everyone has all of the computing power they want. AI is currently driving demand. If there is not something before then, it will be when linewidths get too close to atomic dimensions. Submicron was not the end but, sub-nanometer could be.
I’m not an expert in AI but I know human nature pretty well. There will never be a day when everyone in the world stops acting like Veruca Salt and wanting something more, right away.
When you get to small enough device sizes, a major power factor is leakage which occurs even when the transistor is not switching. Leakage grows exponentially as gate size decreases, and leaky gates are a gigantic factor. Especially since hot (leaky) gates are faster.
You can control this to some extent by the voltage which you use, and part of the test for the processors I worked on was running them at decreasing voltages until they fail. The lowest voltage for which they work is burned into memory elements and used to direct the smart power supplies on the board, so the processors on a processor board might all run at different voltages.
Since chiplets can give you the same number of gates with less aggressive technology they save power. They also improve yields, since smaller chips yield better.
A lot of the advances in transistors today is reducing area by going 3D. Same goes for chiplets. If you can stack a memory chip on a cpu chip (or gpu chip) you can go faster by reducing internconnect lengths with a smaller footprint. Of course you have to worry about heat.
I doubt Moore’s Law is ever going to end, but it has slowed strictly due to economics already. Fab costs grow more than linearly with technology, and if you roll out new processes too quickly you can never recover the costs of the previous one. Like I said, I’ve been at the meetings.
There are all kinds of additional problems at the nanometer level. Any decently sized on-chip memory these days has error detection and correction circuitry, and is also built to be redundant. During test failing cells are detected and extra rows or columns are swapped in to replace them. Yields would be awful otherwise - I had the numbers on my old job. Now people are worried about flops suffering from the same problem. I haven’t seen anything demonstrating this has happened yet, but I’m not privy to that kind of info anymore, but I know of a lot of research about it.
Again, building 2.5 or 3D systems with fast interconnect is likely a lot more cost effective, and allows Moore’s Law to keep going without improving silicon.