Rendered at 19:36:20 GMT+0000 (Coordinated Universal Time) with Cloudflare Workers.
andrewla 22 minutes ago [-]
I can't read the article itself (or maybe listen to it? Hard to see what this page actually is). I read Harvard's summary [1] of the article. The preprint is also available [2]
One thing to note is that the framing "Just 400 US Data Centers" makes it seem like 400 out of the gazillion data centers use this much electricity, but the original research seems to be that they found exactly 403 "Hyperscale Data Centers" and evaluated them.
The framing is technically correct (since there exist non-hyperscale data centers) but is clearly intended to be deceptive.
Also note like just the top 10 steel plants equal michigan as well...
undersuit 3 minutes ago [-]
>"Just 400 US Data Centers" makes it seem like 400 out of the gazillion data centers use this much electricity, but the original research seems to be that they found exactly 403 "Hyperscale Data Centers" and evaluated them.
I would assume you wouldn't select the 400 smallest data centers and 400 random data centers would be weird to write an article about, so I defaulted to thinking the 400 largest(with available data). Was this assumption wrong? Where is the deception? Is it Michigan?
unrented7977 10 minutes ago [-]
If the framing were "top 400 datacenters use more electricity than Michigan" would you still call it deceptive?
The point still stands that 400 of the gazillion datacenters in the US use more energy than the entire population and industry of an entire state. That's a ridiculous amount of energy to pour into any number of datacenters.
renehsz 34 minutes ago [-]
Datacenter power consumption wouldn't be an issue if that power came from local renewable sources.
Unfortunately, the gas-fired turbines that are being built now feel like quite a regression. Or the additional load on the public grid, which in some places barely has any spare capacity left.
It's not an unsolvable problem though. A sustainable hyperscale datacenter can be built. There just isn't enough of an incentive as long as buying gas and polluting the air is cheaper (because the cost of both is externalized on society)
WarmWash 38 seconds ago [-]
>Datacenter power consumption wouldn't be an issue if...
...they didn't use them for AI.
The panic over DC's has nothing to do with DC's. It's entirely about fear around AI and knee-capping AI progress. It's like when conservatives suddenly care about native grasshoppers when they find out a solar farm is gonna be put on the land.
bradleybuda 1 hours ago [-]
The GDP of Michigan is $734B. The combined cloud computing revenues of Google, AWS, and Microsoft (which make up a large fraction of this datacenter work, though not all of it) are $290B. So yes, we should expect the power usage of these things to be in the same order of magnitude.
missedthecue 50 minutes ago [-]
Why calculate the cloud computing revenues and not the economic activity facilitated by the existence of cloud computing? You're measuring multiple orders of economic activity by using Michigan GDP but only one order using google, aws, and msft cloud revs.
coryrc 18 minutes ago [-]
We're also not counting the petroleum powering the cars everyone is driving and the natural gas used as the primary source of building heat.
Electricity is useful for so many things, but it's not needed in huge amounts in an ordinary Michigander's life today. I wish they'd switch to EVs and heat pumps though, but electric rates make the latter cost more (EVs are cheaper).
bluGill 11 minutes ago [-]
Michigan should install heat pumps. However it gets cold enough that a heat pump will not work for several days every winter (and even more nights), so while heat pumps will help they still need something and natural gas ends up our best bet...
shimman 59 minutes ago [-]
No, I wouldn't expect that at all. Google, AWS, and MSFT do not care about helping society. They do not care about making sure children get a good public education. They do not care that workers are treated with dignity. These three can't even bother to hire Americans for jobs in America. They just care about making money.
The state of Michigan provides more value for humanity than those three companies easily.
In fact we should explicitly tax those companies to provide free public school lunches, universal childcare, and universal college/secondary/vocational training programs.
For fun, look at what these three companies are explicitly paying for their data center build out and look at the cost of what it would cost to provide every American school child free lunch.
These companies are sick and should be justly destroyed.
bluedino 51 minutes ago [-]
> These three can't even bother to hire Americans for jobs in America. They just care about making money.
Are you talking about the auto companies?
jayGlow 13 minutes ago [-]
we already tax these companies besides that the US spends on average around $17,500 annually per student among the highest in the world. How much more would we have to spend to provide free lunches for everyone? is there any guarantees that if we increased taxes we would actually solve these issues?
53 minutes ago [-]
richwater 56 minutes ago [-]
> should be justly destroyed
Ah, the truth emerges. The smokescreen about taxes, appeal to emotion, etc are all just rooted in that you want to seize and destroy companies you personally disagree of.
Computer0 30 minutes ago [-]
Who’s side are you on? Seizing and destroying is the peaceful option that will be presented.
djeastm 1 hours ago [-]
Somewhat related: Data centers used 22 percent of Ireland's metered electricity in 2024, up from 5 percent in 2015.
there’s a good reason for this and it relates to ireland and US corporate tax laws. Basically the rules changed for corps to move hqs there , and instead they needed to prove “economic substance” in ireland to establish EU hqs, which for tech companies meant data center builds.
Goodliff 48 minutes ago [-]
I work in AI, but this is obviously not sustainable unless renewable energy is used.
But according to AI Multiple:
Globally, only about 22% to 27% of the actual electricity consumed by AI and data center facilities comes directly from renewable energy sources like solar, wind, and hydropower.
They recommend the obvious, "Prioritize inference efficiency over training efficiency" since most power is used for inference now.
They reference research that shows that over 80% of AI compute is now used for inference.
Your message makes a lot of unsubstantiated assumptions.
j45 46 minutes ago [-]
There is one more thing to consider - have our phones and software not become more efficient with computational power and battery life?
I don't see that referenced anywhere often enough in the AI datacenter swirling.
It's like saying mobile phones from the 1990's would never become more efficient, nor software, nor chips.
scottLobster 37 minutes ago [-]
Most of those efficiency gains came from shrinking transistors, and we've already hit the physical limits of Moore's Law with atomic-scale transistors.
That's not to say there aren't more efficiency gains to be found, but the low-hanging fruit has been thoroughly picked.
mw888 38 minutes ago [-]
"Work in AI" is one of those vague claims that could mean you work in a field relevant to the topic at hand, infrastructure, or that you simply use AI a lot at work. If your role was the more relevant one, one'd expect you to claim it specifically.
Your point is so basic it requires no expertise to reason about, however, anyways.
maerF0x0 15 minutes ago [-]
The demand creates incentive for people to add capacity to the grid. IMO what's really needed is a combination of capital to invest into the grid plus incentives for the public to use less. (Eg: allow DCs to pay marginal bids / $ to those who voluntarily reduce their usage. Allow a DC to pay me, say, $10 to free up $10 worth of marginal electricity usage, they effectively pay double the market rate and I have a benefit from DCs)
ck2 3 minutes ago [-]
I do not understand how "AI" can be profitable, it doesn't make sense
even the expensive hardware has to be replaced every few years
smallerize 1 hours ago [-]
What I want to know is - what are they all doing all the time?
phoronixrly 14 minutes ago [-]
Busy with inference for all the LLMs shoved into everything against users' wishes?
jakwnd 40 minutes ago [-]
Hold on I'll ask chatgpt.
Every google now comes with an AI answer too.
AuthAuth 36 minutes ago [-]
put that in a number i can understand like chinese steel plants or bitcoin mining
j45 47 minutes ago [-]
I wish we could see datacenters as a percentage of other industries and how much electricity they use.
Silamoth 58 minutes ago [-]
Always good to see more hard data on this topic! I’d be curious to play with their dataset. I toyed around with doing some data center research a couple months ago. My vision was to compare data center counts and population densities near data centers built before and after the dawn of LLMs. And do some other stuff. But I had trouble finding reliable data on where data centers are located (some lists exist, but they’re not perfect, so I was doing manual validation) and when they were built (permitting and whatnot provided decent bounds). In the end, there was too much manual work. I didn’t feel like doing that all weekend for several weekends in a row, and I have a day job to worry about during the week. Though this might be motivation to look back at this and do some automation…
Regarding the actual result, I’m appalled but not surprised (though obviously it’s good to have concrete data). We environmentalists having been raising this concern for at least a year. More people are catching on - this seems to have helped the Blue Wave this election season. But a lot of tech bros still don’t get it, particularly here on HN. Maybe this will help change some minds.
spicyusername 31 minutes ago [-]
And how much would we expect that number to be? I suspect nearly everyone is sufficiently ignorant about both datacenters and Michigan that they have no clue.
Useless rage baiting.
luckydata 1 hours ago [-]
"just"?
honeycrispy 35 minutes ago [-]
400 data centers are using more power than what 10 million people use to power their lives.
What do you mean "just"?
qwe----3 14 minutes ago [-]
Just 10 steel plants = michigan, how does that make you feel inside?
rwz 29 minutes ago [-]
400 data centers probably serve a lot more than 10 million people though.
One thing to note is that the framing "Just 400 US Data Centers" makes it seem like 400 out of the gazillion data centers use this much electricity, but the original research seems to be that they found exactly 403 "Hyperscale Data Centers" and evaluated them.
The framing is technically correct (since there exist non-hyperscale data centers) but is clearly intended to be deceptive.
[1] https://hsph.harvard.edu/news/hyperscale-data-centers-energy...
[2] https://arxiv.org/pdf/2606.05420
I would assume you wouldn't select the 400 smallest data centers and 400 random data centers would be weird to write an article about, so I defaulted to thinking the 400 largest(with available data). Was this assumption wrong? Where is the deception? Is it Michigan?
The point still stands that 400 of the gazillion datacenters in the US use more energy than the entire population and industry of an entire state. That's a ridiculous amount of energy to pour into any number of datacenters.
Unfortunately, the gas-fired turbines that are being built now feel like quite a regression. Or the additional load on the public grid, which in some places barely has any spare capacity left.
It's not an unsolvable problem though. A sustainable hyperscale datacenter can be built. There just isn't enough of an incentive as long as buying gas and polluting the air is cheaper (because the cost of both is externalized on society)
...they didn't use them for AI.
The panic over DC's has nothing to do with DC's. It's entirely about fear around AI and knee-capping AI progress. It's like when conservatives suddenly care about native grasshoppers when they find out a solar farm is gonna be put on the land.
Electricity is useful for so many things, but it's not needed in huge amounts in an ordinary Michigander's life today. I wish they'd switch to EVs and heat pumps though, but electric rates make the latter cost more (EVs are cheaper).
The state of Michigan provides more value for humanity than those three companies easily.
In fact we should explicitly tax those companies to provide free public school lunches, universal childcare, and universal college/secondary/vocational training programs.
For fun, look at what these three companies are explicitly paying for their data center build out and look at the cost of what it would cost to provide every American school child free lunch.
These companies are sick and should be justly destroyed.
Are you talking about the auto companies?
Ah, the truth emerges. The smokescreen about taxes, appeal to emotion, etc are all just rooted in that you want to seize and destroy companies you personally disagree of.
https://www.cso.ie/en/releasesandpublications/ep/p-dcmec/dat...
But according to AI Multiple:
Globally, only about 22% to 27% of the actual electricity consumed by AI and data center facilities comes directly from renewable energy sources like solar, wind, and hydropower.
They recommend the obvious, "Prioritize inference efficiency over training efficiency" since most power is used for inference now.
They reference research that shows that over 80% of AI compute is now used for inference.
https://aimultiple.com/ai-energy-consumption
I don't see that referenced anywhere often enough in the AI datacenter swirling.
It's like saying mobile phones from the 1990's would never become more efficient, nor software, nor chips.
That's not to say there aren't more efficiency gains to be found, but the low-hanging fruit has been thoroughly picked.
Your point is so basic it requires no expertise to reason about, however, anyways.
even the expensive hardware has to be replaced every few years
Every google now comes with an AI answer too.
Regarding the actual result, I’m appalled but not surprised (though obviously it’s good to have concrete data). We environmentalists having been raising this concern for at least a year. More people are catching on - this seems to have helped the Blue Wave this election season. But a lot of tech bros still don’t get it, particularly here on HN. Maybe this will help change some minds.
Useless rage baiting.
What do you mean "just"?