[HN Gopher] At Amazon's biggest data center, everything is super...
___________________________________________________________________
At Amazon's biggest data center, everything is supersized for AI
Author : pseudolus
Score : 83 points
Date : 2025-07-11 10:03 UTC (12 hours ago)
(HTM) web link (www.nytimes.com)
(TXT) w3m dump (www.nytimes.com)
| pseudolus wrote:
| https://archive.ph/9ytmy
| 2OEH8eoCRo0 wrote:
| > After building seven data centers in Indiana, Amazon plans to
| build 23 more.
|
| Jesus
| v5v3 wrote:
| Large numbers of the 8 billion people in the world are still
| not digitally literate.
|
| The demand from all the poorer countries with large populations
| is only going to go up as their kids grow up with tech.
| altcognito wrote:
| I'd imagine they would build those data centers closer to
| those populations.
| manquer wrote:
| It is not for inference so population or market does not
| matter, this is only for training and only for Anthropic,
| which is pretty unprecedented in the history
| computing/internet.
|
| > all the poorer countries with large populations
|
| It won't move the needle even if this infra is exclusively
| for inference.
|
| These projects are based on a economic thesis. Companies are
| investing $10s-100B / year expecting to make the money and
| more back.
|
| At this scale of money RoI will not come from parts of the
| world with per capita GDP in the low thousands or less even
| if there are billions of them and all of them are expected to
| use these new tools.
|
| This is same reason Apple is the richest company in the world
| despite selling a fraction of droid devices. Every mobile app
| developer knows all money is only in iOS. The top 10% people
| in the world as a market can make a company much more than
| the rest 90% combined.
| dude250711 wrote:
| Once the overall AI bubble bursts/deflates, would they be able to
| re-purpose it for other workloads e.g. cloud gaming?
|
| Though, maybe Midjourney or other niche AI companies could buy it
| for the right price I guess.
| progbits wrote:
| The server/ML GPUs are not great for gaming, they strip out all
| the specialized shader/pixel units to cram in more general
| compute cores.
|
| In theory you could write games with just those but aside from
| the amount of work needed I'm not sure if the performance would
| be good enough, the specialized texture samplers etc can be
| faster than general purpose compute shaders.
|
| Also, for cloud gaming you want very low latency, so few GPUs
| all over the world in local POPs, not a lot of GPUs in few
| large data centers.
| tyingq wrote:
| "Trainium accelerator" doesn't sound like it would be terribly
| useful outside it's current niche. It's not clear how much of
| the spend is that kind of thing versus general purpose compute
| and storage.
| coredog64 wrote:
| For the previous generation, AWS had a split where Trainium
| == model training and Inferentia == inference. With the
| current generation, they are also recommending smaller
| Trainium configurations for inference.
|
| Even if nobody trains another model, the Trainium chips will
| be useful for inference of existing models that have been
| recompiled for it.
| closewith wrote:
| This feels like saying when the dot-com bubble burst, could the
| servers be re-purposed as mainframes. Whatever about valuations
| and individual stocks, AI/LLM workloads are only going up.
| 0cf8612b2e1e wrote:
| I am more on the AI skeptic side (LLMs are not a path to
| general intelligence, most positions cannot be 100% replaced
| with any version of product I have seem to date), but I fully
| agree. _Some_ form of AI text generation is likely to be here
| forever. Maybe it gets vastly more efficient per clock cycle,
| but compared to N years ago, all of the tech vendors need
| more compute to offer this service.
| kjkjadksj wrote:
| It doesn't matter if its general intelligence. What matters
| is if they can get workers and consumers using it every
| day. Which increasingly seems to be the case.
| Nasrudith wrote:
| Putting aside possible specialization of hardware and
| differences for applications, cloud gaming was tried and
| largely flopped. Related for the power intense use-cases where
| it could be useful to outsource had too much latency involved.
| Remember Strada?
| DrBazza wrote:
| > is the first in a new generation of data centers being built by
| Amazon, and part of what the company calls Project Rainier, after
| the mountain that looms near its Seattle headquarters
|
| Near both the Casadia fault that's produced magnitude 9
| earthquakes _and_ a chain of active volcanoes. Both of which are
| statistically likely. I do wonder what contingency plans Amazon
| (and Microsoft) have in the event of a megathrust earthquake.
|
| https://en.wikipedia.org/wiki/Cascadia_subduction_zone#Forec...
|
| https://en.wikipedia.org/wiki/Mount_Rainier#Modern_activity_...
| readthenotes1 wrote:
| It is named project rainier, but the facilities are in Indiana.
|
| If you're speaking metaphorically, the mega thrust would be
| some cataclysmic world event like China invading Taiwan. In
| which case they won't need to earn natural gas to power or
| waste fresh water to cool...
| jltsiren wrote:
| Megathrust earthquakes are pretty common around the world.
| Some of them cause major damage, such as the ones in 2004 and
| 2011. But even then, the impact is mostly regional.
| PNewling wrote:
| > Both of which are statistically likely.
|
| It's seems a _little_ dishonest to not include that
| 'statistically likely' is on a geological time scale.
| progbits wrote:
| > The local utility will largely use natural gas to generate the
| additional electricity needed to power Amazon's data center
|
| Sad, but expected.
| sigmoid10 wrote:
| At this point we have to be glad it's not coal. After all we
| have a president that just passed an executive order to foster
| the "Beautiful Clean Coal Industry"
| blibble wrote:
| > At this point we have to be glad it's not coal.
|
| it will be if the coal price drops far enough
| gadders wrote:
| It's either that or nuclear.
| bflesch wrote:
| False equivalence bias
| gadders wrote:
| Dunning-Kruger effect
| progbits wrote:
| You say that as if nuclear was bad?
| gadders wrote:
| Not at all. I think it's probably the best option if you
| want guaranteed electricity.
| lordswork wrote:
| Far better than xAI's data center being powered by mobile
| Diesel generators
| perihelions wrote:
| xAI is natural gas, not diesel.
| polski-g wrote:
| Decades of lies about nuclear has its downsides.
| thegreatpeter wrote:
| For something to be considered "sad", there needs to be an
| available, practical alternative. What should they have chosen
| instead?
| joelthelion wrote:
| We could not build new data centers and focus on fixing the
| real issues instead.
|
| We could always revisit this decision once clean power is
| actually available.
| TOGoS wrote:
| Cover the roof in solar panels and use the power to do
| something useful instead of running bullshit generators.
|
| Or maybe a park with some oak trees and a frisbee golf course
| idk.
| yahoozoo wrote:
| > The facility will consume 2.2 gigawatts of electricity --
| enough to power a million homes. Each year, it will use millions
| of gallons of water to keep the chips from overheating. And it
| was built with a single customer in mind: the A.I. start-up
| Anthropic, which aims to create an A.I. system that matches the
| human brain.
|
| Sentences like this never ceases to amaze me. All of this juice
| to attempt to match what a single human brain can do with its
| relatively low resource requirements.
| mc32 wrote:
| I think you're conflating things. It's not a single human
| brain. It's processing power to provide the human trove of
| knowledge and reasoning at the beck and call of millions of
| people nearly simultaneously. No single person would be able to
| do that.
| threetonesun wrote:
| No, but we had done a pretty good job of storing and indexing
| it all on this thing called the Internet.
| matt123456789 wrote:
| And for the same purpose before that we had, or rather
| have, since it is still very much in use today, the Dewey
| Decimal System. Peering further onward still into the
| annals of library science, one can find even more creative
| and revealing methods of indexing human knowledge. In doing
| so, one might even be inclined to believe that the state of
| the science has come so far as to be considered solved.
| Alas.
| threetonesun wrote:
| I think if you went to the library and asked for
| information on "how to build a chicken coop" and the
| librarian took 60 books related to chickens and building
| and farming, cut up the words in them, then arranged them
| in a way they found satisfying, you might start going to
| a different library.
| blibble wrote:
| > It's processing power to provide produce near unlimited
| spam and millions of pictures of shrimp jesus
|
| ftfy
| tartoran wrote:
| It's a laudable goal to match human intelligence but we can't
| ignore the cost for too long. If humans can create the same
| reasoning for cheaper why would we chug along on this path?
| And this may be the case when the AI hype bubble pops and
| real economics take over.
| mc32 wrote:
| This is not the intelligence of one person, even the most
| intelligent person but rather the intelligence (wisdom) of
| the whole crowd -the whole world democratized (I.e.
| available to everyone for a nominal fee).
|
| That's an important difference.
| h1fra wrote:
| I thought datacenter water cooling was usually a closed loop? I
| keep reading conflicting info about this
| 2OEH8eoCRo0 wrote:
| Nope. Evaporative cooling.
| sleepydog wrote:
| For very hot data centers, evaporative cooling is still
| popular. This is from 2012 but I doubt much has changed.
|
| https://blog.google/outreach-
| initiatives/sustainability/gett...
| aitchnyu wrote:
| Are new water-guzzling DCs and nuclear plants built on
| water sources unlikely to be affected by climate change?
| placardloop wrote:
| 13 years is an incredibly long time for something as fast
| moving as data center development. I guarantee that a _lot_
| has changed. I know AWS in particular has gone through
| multiple entire revisions of their DC designs, and I recall
| a talk from some of their engineers saying how AWS actually
| found it more economical to use less cooling and let their
| DCs run hotter than they used to.
|
| Here's a recent article from AWS about using closed-loop
| systems for their AI data centers:
| https://www.aboutamazon.com/news/aws/aws-liquid-cooling-
| data...
| 2OEH8eoCRo0 wrote:
| Data centers may change but the physics of cooling
| doesn't.
|
| It's more economical to run chips hotter but at the end
| of the day you'll still have heat that needs dissipating
| and it's hard if not impossible to beat evaporative
| cooling in terms of cost.
| placardloop wrote:
| This is like someone in 1800 saying "at the end of the
| day you still have transportation needs and it's hard if
| not impossible to beat horses and carriages in terms of
| cost".
|
| Literally just do a google search. There are advancements
| every day that improve upon evaporative cooling to make
| it use less and less water and energy, and alternative
| methods other than evaporative cooling.
| butlike wrote:
| Bleeding-edge advancement and commercially-viable
| solutions are not apples to apples.
| 2OEH8eoCRo0 wrote:
| If there were then datacenters would use them ;) There
| must be a catch eh?
| jandrewrogers wrote:
| > Each year, it will use millions of gallons of water to keep
| the chips from overheating.
|
| Growing corn on that same 1200 acres would require on the order
| of a _billion_ gallons of water. People have no sense of just
| how much water agriculture consumes.
|
| In all probability, putting a data center on farm land is
| greatly reducing water consumption.
| raldi wrote:
| Doesn't a typical McDonald's require millions of gallons of
| water each week just for the beef alone, without even
| considering things like drinks and ice and boiling potatoes
| and washing floors?
| acquiesce wrote:
| By pointing out the obvious you miss out jumping on the
| bandwagon of everyone writhing in pain over how devastated
| they are for the environmental impact. Whenever AI is brought
| up you have to remind people how unsustainable it is and how
| it's all about to fall apart and how the value isn't there,
| among other topics. Off topic/on topic I'm reminded of death
| reports during COVID for particular states. When I went to go
| and look at the statistics of years prior to compare, they
| were at par if not less than the few previous years. Just an
| example of uncontrolled (mostly manipulative) outrage in
| action. I'd like to expect better of HN but I don't: I've
| lost all respect for the moderation (dang and now tom) here
| over the years.
| libraryofbabel wrote:
| Yeah. The power use is impressive, although still only 1/10th
| of the output of the largest power station in the world, the
| three gorges dam. The water use is not really a big deal at
| all. I don't know why journalists inevitably have to mention
| both. And even the power numbers require much more careful
| contextualization.
|
| I wish I had a good balanced article to link to when these
| discussions come up, one that really digs into the power use
| question and compares it to the power consumed by other human
| activities and capital projects. Does anyone know of one?
| gdbsjjdn wrote:
| Let me fix that for you: all that juice to exploit human's
| cognitive biases and trick them into paying engineers and
| founders ungodly sums of money.
|
| I don't think any rational person outside of the hype cycle
| thinks LLM s are emulating a human brain in a meaningful way.
| golly_ned wrote:
| Considering the human brain evolved over hundreds of millions
| of years of evolution (what a lot of data!) that almost seems
| cheap to do in the span of a number of months to train a model.
| v5v3 wrote:
| >A year ago, a 1,200-acre stretch of farmland outside New
| Carlisle, Ind., was an empty cornfield. Now, seven Amazon data
| centers rise up from the rich soil, each larger than a football
| stadium.
|
| Can't they find brownfield sites instead of fields
| micromacrofoot wrote:
| it's probably not actually rich soil, maybe a writer's flourish
| - post harvest fields are pretty much barren of nutrients and
| life
| maxerickson wrote:
| The county it is located in is about 290,000 acres.
|
| Indiana is about 23 million acres.
| lordswork wrote:
| This comment got me wondering whether loss of farmland in the
| US is a serious issue.
|
| It looks like there's about 800 million acres of farmland in
| the US and we're losing about 2 million acres per year to the
| land being repurposed. Despite that, crop production has more
| than tripled in the past 70 years due to technological
| advances.
|
| That said, economic effects, loss of farmland, and climate
| change have contributed to slower growth and higher variability
| of crop yields recently.
|
| In the past decade there's been a modest 0.8% annual increase
| in crop production despite losing about 2 million acres per
| year.
| TheNewsIsHere wrote:
| One of the major problems facing American agriculture is that
| there are fewer and fewer farmers/farming families.
|
| Farming is extremely money and labor intensive and there's a
| lot of upfront investment with a lot of long-tail return, and
| it's not "sexy" the way (for example) AI is, so there's not a
| bottomless pit of cash to shovel into the furnace for a quick
| buck turnaround.
|
| Independent farmers tend to seriously rely on good weather
| and a lot of advantageous tax treatment.
|
| Of course massive agri-business would very much love to
| continue to fill more and more of the void left by the
| shrinking independent farming population. That has its own
| problems.
| antisthenes wrote:
| It will swing back one day. Maybe in a few generations.
| kfajdsl wrote:
| Has there ever been a swing back to small, independent
| producers of anything after an industry has consolidated?
| antisthenes wrote:
| You're thinking too short-term if you're asking that
| question.
| clvx wrote:
| The problem is not the amount of land but if that land is
| economically viable. Farm already has low margins.so, if you
| grab a good location to build a data center and push the farm
| land even further away from population centers, then you are
| pretty much killing family farms.
| kasey_junk wrote:
| There are virtually no family farms left in the US.
| Especially central Indiana corn farmers. 1200 acres
| wouldn't be a financially viable corn farm if it were
| family farmed.
| nhecker wrote:
| I think this is really close. My hunch is that agricultural
| land is just simply cheaper to acquire and convert, as
| compared to industrial land which may or may not have all
| kinds of remediation or razing that needs to happen to it
| first.
| aaronbaugher wrote:
| Yeah, it's not a significant amount of land. But it does seem
| like these companies prefer to take quality farmland instead
| of unused land. There's something similar happening near my
| town, where a company wants to put a big solar array on some
| prime farmland, and the locals are asking why that spot and
| not getting an answer. It might save a small amount on
| development, as farmland is fairly flat and has no trees to
| remove, but that's miniscule in the overall budget for these
| things, and rough ground would be much cheaper to buy. But
| the corporation behind it and the government entities
| involved are digging in their heels and insisting on using
| farmland, without any explanation why.
|
| So it does seem like some of the people making these
| decisions just like the idea of taking farmland out of
| production for some reason. Maybe they just don't like
| farmers or modern farming methods. If that's their motive,
| they may not realize how tiny an effect they're having on the
| total, because most non-farmers don't really understand how
| much land is out there.
|
| It's like the people who say Bill Gates is trying to control
| the food supply because he owns something like 270,000 acres
| of farmland. Even that just isn't that much, not enough for
| him to control anything larger than the horseradish market.
| ssl-3 wrote:
| What unused land? Where?
|
| In my part of Ohio, everything big enough for a meaningful
| solar farm is "prime" farmland or coveted and necessary
| forest or wetlands -- unless it's already used for
| business, or housing, or infrastructure to support
| civilization.
|
| Perhaps my perspective is simply very limited, but: In my
| estimation, there is no unused land to use.
|
| Even the big pile of dirt I drive by twice a day: It does
| stuff. It gets bigger and smaller as some pay to get rid of
| extra dirt from their project, and others pay to buy some
| of that dirt for a different project. Someone somewhere
| manages that pile of dirt.
| anon7000 wrote:
| Also whether the farmland requires irrigation. In Ohio,
| there's not much irrigation. In Indiana, you see it a lot
| more. In Idaho, it's basically a hard requirement
| perihelions wrote:
| > _" Can't they find brownfield sites instead of fields"_
|
| You mean precisely like xAI did in Memphis? Build on the site
| of an abandoned factory--a historically disadvantaged, low-
| income industrial zone?
|
| I couldn't imagine anyone having anything negative to say about
| that.
| wpm wrote:
| As far as "industrial" neighbors go, a datacenter is a pretty
| good one to have vs an abandoned husk vs an active
| polluter/nuisance, at least if that datacenter isn't run by a
| selfish psychopath that tacked on a bunch of gas burning
| generators to power the thing.
| lesuorac wrote:
| This seems pretty bad faith.
|
| The problem with what xAI is doing in Memphis is they were
| originally running "temporary" generators for over a year
| (effectively permanently ...) without meeting emission
| requirements a permanent generator would need to.
|
| Sure, it's not xAI's fault the local power grid isn't
| suitable for their datacenter but the problem here is that
| xAI is effectively not following the law.
| j-bos wrote:
| I don't know about Indiana, but in Texas companies buy farmland
| because it's often available near population centers and
| infrastructure. And the land is available because the existing
| owners do just enough farming/ranching to qualify for the lower
| tax category.
| placardloop wrote:
| This particular site is being built basically in the parking
| lot/backyard of a steel mill, so it sort of is brownfield. I
| have a hunch that "empty cornfield" is just the author using
| some artistic freedom instead of saying "an empty field in the
| Midwest".
| allturtles wrote:
| It was definitely farm fields growing crops, you can look in
| Google Earth and see for yourself.
|
| Edit: Anywhere in the Midwest where there is a large expanse
| of flat ground will almost always be growing crops or have a
| town built on it. Undeveloped land is almost all hilly and
| forested or floodplain, nowhere where you could put a data
| center.
| xhkkffbf wrote:
| Why? Hold outs, zoning, greedy politicians and more. I drove
| through a blighted neighborhood. There were 60-70% unkempt
| houses and maybe 10-15% in great shape. Some won't want to move
| for their own reasons and they'll refuse to sell. Some will
| just be trying to get more money. Others will be sincere. It's
| very hard to unsubdivide land.
|
| And then there are the zoning and city regulations which
| naturally resist. You would think they would welcome a new
| taxpayer, but they're really in love with the regulations that
| someone wrote decades ago. The odds of getting them to bend is
| anyone's guess, but in most cases a bribe or ten is going to be
| required.
| tartoran wrote:
| Im a bit clueless on this subject but I've been wondering why
| isn't this heat used productively to, say, generate electricity.
| Why aren't AI datacenters for example use the heat to provide the
| municipality with hot water for example?
| jandrewrogers wrote:
| Electricity from heat is generated by temperature differences.
| Large differences are much more efficient than small
| differences and data centers are "small differences" for these
| purposes.
|
| Unless you can use the heat directly, it won't be worth it to
| convert it into electricity.
| foepys wrote:
| Using heatpumps you only need a little electricity to heat
| quite a few homes with district heating.
|
| But there probably is no nearby district to be heated.
| delfinom wrote:
| And you only need heating for at the very most, 6 months of
| the year in the US. Though that's shrunk depending on
| region in recent years.
| foepys wrote:
| True for heating but warm water is in demand all the
| time.
| coredog64 wrote:
| The limits are described by Carnot's theorem: https://en.wiki
| pedia.org/wiki/Carnot%27s_theorem_(thermodyna...
|
| Also important to remember when calculating this for yourself
| that the temperatures must be represented using an absolute
| scale (i.e. Kelvins or degrees Rankine).
| cyprien_g wrote:
| There are already some datacenters doing this:
| https://www.techtarget.com/searchdatacenter/tip/Data-center-...
| But for now, it's still the exception, probably because of the
| cost of such systems.
___________________________________________________________________
(page generated 2025-07-11 23:02 UTC)