There are folks who are euphoric about AI and there are skeptics and there are folks who don’t know where they fall, but fall somewhere in between.
Irrespective of where you fall, all of us have to be aware of some facts. There are lots of AI data centers are being build. One GW datacenter takes about $50 B to build, and consumes as much electricity as a city of 1 million people. It is not just the electricity consumption but water consumption and other environmental issues too. Let us set aside all of that.
If you look at the structure of these deals, the AI companies (the likes of OpenAI, etc) are protected, i.e., if for some reason AI sputters, these companies are not the ones holding the bag, but others, from real estate companies to utilities to local cities and thus people who were never part of any of these deal, that is you.
While some cities are fighting back for their local resources like water, electricity , many are yet to wake up to the financial ruin that will come their way. Just one example, the cost for the grid buildout, which are passed to the consumers via base rate alone will bankrupt most cities that have build these AI data centers, even if those DC’s operate at 50% capacity.
At this point AI is too big to fail. If it fails, it is not the AI companies that will suffer the biggest financial ruin, the entire economy will bear the cost much more.
It is scary how little these financial risks are discussed in MSM, or by congress or even city halls, who are going to be left with a massive cost.
Here is another risk, this is based on Hyperion data center, Louisiana. $META is developing 2 GW, funded by
$27.3 B of bonds, 74%
Out of $9.5 B of equity, Blue Own $7.5 B, 19%
$META $2.5 equity is in-kind contribution of land and development, 7%
The campus is owned by SPV Beignet Investor LLC
$META Lease ends in 4 years
So, $META offered 16 years of Residual value guarantee, i.e., if the building is not worth to pay bondholders $27.3, then $META will pay the difference, and/ thus the bonds received $META’s AAA bond rating.
Now, to bring power to the data center, the local utility, entergy has to spend billions in building the new grid, this will increase the cost for all the neaby consumers.
Virginia courts have already ruled that the consumers has to pay for the base rate increase (remember, Ashburn, VA is a big cloud data center region).
For reference, typical enterprise data center consumed 30 MW, and AI data center consumes 2000 MW, or 65+ data center power.
This is not a one-off deal structure, this is the template that is followed in most, if not all DC buildouts, and some even more onerous.
Republicans are cheering these DC buildouts, and democrats don’t know how to fight, or too busy because President keeps them busy by distracting them on frivilous issues. In the end, We the people, are bag holders.
Separately, this is the reason I think AI companies have little to lose, but overall economy, city finances, every household will bear the cost, whether AI wins or loses.
This post is because, the cost-benefit analysis of building these DC’s from hyperscaler point of view, an attempt to understand how much $META is at risk. The unfortunate realization is the cost to the public. There are no benefits to the local communities. There are no real, significant employment generated by these DC’s. They are sitting in your neighbourhood, sucking power, water, causing environmental issues and increasing your electricity cost and water cost.
If there is any benefit, you can articulate that, don’t be in a rush, take your time to find one. I could not find any.
To choose a local example, IIRC, Leap has found AI (which is dependent upon large data centers) useful in developing his game. On a larger scale, I have read that the militaries of both China and the US have found it quite useful.
The question is not whether AI has any benefits, but whether hosting a AI data center has any benefits to the local community. Clearly, No.
One can argue nuclear weapons has military benefits, a deterrent against unnecessary, costly war, etc. I am sure you don’t want nuclear testing on your backyard. Don’t paint it as extreme, sometimes only extreme cases help bring clarity to an absurd argument.
AI is dependent upon large data centers. The data centers have to be located somewhere. Unless one goes all NIMBY, it is an interesting problem to solve.
It’s like nuclear plants: put them where nobody lives and nobody cares. I suggest North Dakota.
In fact, you can do cooling without a lot of water (think: air conditioners) if you have enough electricity, which means you should put them next to nuclear plants. All you need to do is string the respective wires to each: electric for the power plant (assuming there’s power left over) and data lines. And you could use the same rights-of-way for both, just separated by a relatively few feet and bingo, nuclear powered AI, and then when there’s an accident they both go up in smoke and a hailstorm of radioactive debris, but luckily nobody cares because it’s in North Dakota.
Easier. Not cheap. Radiative cooling requires a lot of mass, and mass in space is expensive. There’s no plausible economic justification for space datacenters. The modest economic advantages of having 24-hour solar are completely outweighed by all the economic disadvantages.
No, you can’t. You can only radiate heat in the vacuum of space. The least efficient way to possibly do it. Stupendously less efficient than conduction or convection. So inefficient in fact that it is a stupid idea to put data centers in space. No, you CAN’T cool something easily in “cold dark space”. Because there is nothing of any mass for that heat to migrate to. Like air, or water.
So many reasons why space data centers are a bad idea, with cooling being one of the top reasons.
