- Joined
- May 18, 1997
- Messages
- 58,239
![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
-
Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.
How AI will fail
- Thread starter rinaldo00
- Start date
- Joined
- Aug 3, 2004
- Messages
- 24,217
This should be basics in terms of laws....especially when some power companies are cutting off customers purely to power DC's....
in Alberta here, the new $13Billion Meta DC - while they said it would run its own gas powered plants...now turns out those plants wont be done before the DC is ready,so they will leech off our grid now to power for potentially 2-4 years..
And you can bet, our rates will likely go up because of it...
mullet
2[H]4U
- Joined
- Aug 19, 2004
- Messages
- 3,363
I knew the whole global warming bs would end when they all wanted their new baby Ai.
New Amazon Data Center Stokes Worry It Would Be the Most Polluting Power Plant in the U.S
New Amazon Data Center Stokes Worry It Would Be the Most Polluting Power Plant in the U.S
Likely at lower pay than before, and also only temporary until they think the ai is ready.
All those coal miners can go back to mining coal once ai has taken their coding jobs. The server farms will need it.
The people who want their AI aren't the same people who were pushing wanting to clean up the environment. Your comment makes it seem like "they" are a group that changed their mind.
Prince Valiant
Gawd
- Joined
- May 3, 2023
- Messages
- 812
https://sustainabilityexchange.amazon.com/credits
Seems like his comment was on point. Am I missing something?
The idea of a 7.5 gw datacenter sound really "big" to me, enough to feel like this is a way off presentation of the situation by the NYTimes.
The permit allow for a lot, but that not at all the plan for the upcoming power plant and datacenter to be built there (phase one is using about 4-8% of the legal capacity). And there is a nice play here, buy off the grid that terrible, build your own that terrible.
It is part of the future plan, but that future is a lot closer than you think, because you think reasonably, and the scale and pace of AI stuff right now is not reasonable.
For example Texas has 470gW plus requests for connection to the grid for next year. Of that ~376 is for data centers. The current grid is about 93GW. Also look at the physical size. The reference NVIDIA H100 rack was 40kW. THe reference GB100 rack was about 100kW. The GB200/300 rack is about 130-150kW. The goal is a megawatt per rack. The reason why GB200/300 is not as big a jump is because they need to come up with new power standards to go bigger, and those are targeting the megawatt rack. The roadmap implies it will go from 500/600 on the initial production iteration and then scale up to WW.
PJM was looking at 253GW or increased data center demand. That's on about 183GW of existing capacity.
Neither of those are planning to turn on next year, but it's effectively booking the next available watt. The line has forming and is already deep to get that next watt. PJM at least provides some sense of scale and timeline. that 253GW of demand spans to 2046 and they do not have the capacity to deploy that much power in that time without a lot of up front cash (maybe not even then), so they have been talking about datacenters funding their own build out.
that because the projects and that numbers city come from projects that are a lot like the one being talked about, with phase one being about 5-10% of that figure and some projects will have 4-5 demands for different site and will be approved and use a maximum of one (the one with the fastest timeline or best price) actual demand is way lower here.
- Joined
- Sep 7, 2011
- Messages
- 10,179
These corrupt fucks are trying to undermine the free market by using socialist politicians to steal via legislation, giving power and water to greedy humans when Data Centers are desperate. Does nobody even care about shareholder value? Profit incentives? What has this country come to.
Also: Poe's law
mullet
2[H]4U
- Joined
- Aug 19, 2004
- Messages
- 3,363
I can't get to it now because of the dumb paywall but basically Amazon, was a huge proponent of green/clean energy and NYT was busting their balls for going against it with nat gas powerplants as the biggest polluter.
7.65 gigawatts is no joke and I aplaud Amazon if they build that many power plants. I hope they all get build before the bubble pops.
Don't worry, it's Canada. There is still plenty of time for nothing to get built.
There is significant overlap between the two groups. Not universal, but significant. Recall the number of corporations who made a big deal about green initiatives, including so many in tech. Fast forward to now, and even Greta isn't even talking about climate change anymore.
I haven't seen much pushback on AI from a climate change standpoint at all, myself. The pushback I generally see is focused on economic issues related to rising costs of energy, water usage, and NIMBYism. It's there, but nowhere near where it was the past few years on literally everything else.
We need all of these power plants to power the next strategic defense systems all throughout the US, big lasers. 
Joking aside, the do it all big AI systems seems to be the wrong target. Smaller, independent, locally run specialized systems that can link up as needed for a given problem seems like a better use of AI.
Joking aside, the do it all big AI systems seems to be the wrong target. Smaller, independent, locally run specialized systems that can link up as needed for a given problem seems like a better use of AI.You think Amazon was pro-environment at any point? They're like most corporations, they advertise what sells.
Corporations generally don't give a shit about the environment. They push for "green initiatives" when it helps them sell products. Water is, yes, the "environment" as is noise, as is digging up lots of wildlife. Pollution from the power has taken a backseat, yes, because people right now are more worried about drinking water and data centre noise than shit air quality. Oh, and Greta went fully on anti-Israel a while ago. Her silence on AI has nothing to do with the environment. She's on a whole other crusade right now.
- Joined
- Jul 17, 2006
- Messages
- 14,429
The datacenters will quickly realize they won't be generating their own power. The initial investment for hundreds of megawatts or multiple gigawatts is not economical or even feasible with a few dozen peaker style units. It would cost about $2B dollars, and their electricity price ammortized would be 3-4x more expensive than grid electricity.
That leaves them without enough baseload power from the grid to operate. Most of the datacenters being built will not come online. They aren't going to load shed neighborhoods.
Prince Valiant
Gawd
- Joined
- May 3, 2023
- Messages
- 812
Certainly not.
It is also about their own employees pressure and retention, they tend to attract employees that love those initiative, I think all those google of the world would be ready to pay a little bit more for Nuclear vs gas if it please them (the employees), a lot of their workforce will be uneasy working for what became big industrial-emission company.
Distribution is a giant part of the final electricity price, that why industrial can often get much cheaper deal with much simpler distrubution per watts infracstructure (in Canada industy pay cheaper rate and yet they are the one that subsidy residential rate which run at a loss being the norm), on location natural gaz that you can use a lot yourself must be compeittive with grids, specially for those with the ability to resell extra. On location tend to be the simplest distrubution, thus the cheapest way.
Right now, 1 gigawatt of cluster can easily generate 10-12 billion a year of compute revenu-value (google is paying 30 to 50 billion right now but that a bit of a desperate price), spending 1-2 billion upfront in comparison is not a problem at all, specially versus how much money you loose for the years of waiting for a connexion to the grid.
Regional natural gas price are good in that spot, xAI used mobile gas turbines with great success for Colossus, I am not sure how you come up with an 3-4x more expensive than grid rate here.
Last edited:
Yep, and I don't expect corporations to care about the environment. That's not why the exist, that's not who they're accountable to. They "cared" about it over the past few years insofar as they found it was an effective marketing campaign that helped them sell product. Having a "green" product was a feature and got a lot of positive press. No one cares about that anymore. Their responsibility is maximizing profits for shareholders, and that's it.
Likewise, I not convinced most environmentalist NGOs really care either. If they did, they wouldn't have executives travelling the world on private jets. The problem IS the product for them, and if the problem gets solved, their funding goes away. They don't want to solve the problem, they want to MANAGE the problem.
Datacenter opposition can have a variety of motivations. Now that costs of living have been revealed to be the primary concern amongst the general population, that's the focus, not the environment, not climate change. But opposition is profitable as well. They get donations from concerned citizens to fund their opposition campaign, and I have no doubt many of them are receiving funding from foreign actors who want to take the lead in the AI race that are not subject to the same restrictions as Western nations are. The most salient evidence for that is a lot of the discussion is not coming from a position of fact, but a position of hysteria, and critical thinking is seldom applied.
- Joined
- Aug 3, 2004
- Messages
- 24,217
- Joined
- Jul 17, 2006
- Messages
- 14,429
I work for a utility and grid supplier, so that's where I'm getting the numbers from since I see it every day. I can tell you that your assumptions are dead wrong.
DukenukemX
[H]F Junkie
- Joined
- Jan 30, 2005
- Messages
- 11,433
Cost of living should be the primary focus for anyone. What's sad is that everyone knows AI is flawed but they don't care. They want extreme surveillance and they want to use it to blackmail people. AI may create mistakes, but that's a risk they're will to take. I know you're referencing China, but it isn't China pushing for AI. It isn't China that's currently the most hated country in the world.
View: https://youtu.be/x7X_Z4Qq6Ik?si=s3ciQTnN8ijGWTPh
Can you walk the step, how a giant user making natural gaz electricity with a direct and simple-short line to themselve will cost them 3x-4x more than the giant user buying it from people that make natural gaz electricity, that has to do a complex transmission of it and a profit ?
Say the first 18 years, how does the average kwh you estimate will cost in both scenario, with a reasonable usage/selling rate of their production and interest cost on the upfront cost ?
If this do not work, how Colossus 1-2 made it work ? datacenter having much costlier natural gaz mobile generator instead of a nice full on installation is not uncommon.
Last edited:
China obviously did not made it illegal to replace human worker by AI... a court judged a work contract between an employee and a company was broken when they did it, if that where the meme come from:
https://www.staffingindustry.com/ne...er of the,to reductions in staffing due to AI.
talking and thinking in meme is a fun shortcut, but there is little reason to do it if one want a realistic view of the world and as the article say: as Chinese companies race to implement AI systems as part of a state-directed push to dominate the new technology.
with china working age population relative to the upcoming retiree population need:
An explosion in productivity is not that much of a choice but a necessity.
- Joined
- Jul 17, 2006
- Messages
- 14,429
Not sure we're in the same country so it may not be the same where you are. The smaller plants are not as efficient for one - mostly due to little or no cogen capability, the cost of construction is more - relative to output capacity, the maintenance costs are higher and downtime costs are detrimental since you have to pay millions per year in grid service fees regardless, they have to pay all the ecological and environmental fees and taxes that come along with power generation on site, etc. I could go on for pages long of why it costs more. I can't give exact figures of datacenters who already ran the numbers in my service zone, but they were far above (200%+) the yearly cost of what we could provide via the grid. We also have had municipalities believe they could provide power to their community of 20k-30k people for cheaper with better reliability, and they determined they could not after residents switched out of their contracts because they didn't like paying $0.25-0.35/kWh where our company was providing for $0.08-$0.12/kWh. Those plants are now abandoned with only 3-5 years of runtime and tens of millions of dollars of debt.
As far as Colossus, they are supplied by TVA, with peakers being used as supplemental power when needed so they don't violate their grid contract. They're also getting sued over it.
Last edited:
is a 2 billion level plant for that amount of gigawatt being mentionned, really a smaller plant too ?
Has for ecological and other fee that goes has well for the producer you will be buying from usually and do not forget the amount of money lost by year of waiting for a grid connexion in the math here, 10-12 billion of revenues a year per gigawatt, that ramp up fast.
- Joined
- Jul 17, 2006
- Messages
- 14,429
I don't quite understand your broken english, but yes, a few percent efficiency gain is what allows the electricity prices of baseload generating stations to charge far less compared to small units. They're moving too fast for the infrastructure which will hurt their bottom line. Coupled with the cost of the computer hardware itself, this is why they are unprofitable and most planned projects won't even come online.
Colossus have been quite profitable and they do tend to be (datacenter) quite profitable. Google, amazon, etc... they made fortune with datacenter for a decade
Not sure we are talking about a small unit here, (a gigawatt of power that what around what you use to power 750,000/800,000 houses) This is not a small town type of affair, closer to a good metro scale level. The whole max project if it ever get made is close to the biggest gas fired plant in the world here, power wise. Is a 1 gigawatt level of capacity pretty much already at the best sweat spot efficacy ? (how much difference versus the bigger power block you have in mind)
And you save on distribution... and on someelse profit, I would be curious for the exact math you have in mind here.
- Joined
- Jul 17, 2006
- Messages
- 14,429
I'm not talking about a general datacenter, I'm talking about AI compute datacenters for which the topic is about.
For power like 15MWe natural gas units you referenced, those are approximately $0.45-$0.55/kWh to run, and more if rented, sometimes up to $1.15/kWh. My company has many, and they're only used during high peak hours probably a few times a summer to sell power to other markets because otherwise they're not even profitable. These are not cost effective to run AI datacenters, or really much of anything. They're meant for emergency power, remote wilderness projects, etc.
For max efficiency, you can get that with well engineered baseload plant designs larger than 100MWe. However, smaller units are still more expensive to run since you're not maximizing your output compared to operating costs. A plant 750MWe in size or larger is much more justifiable but it will take several billion dollars in investment. You not only have to pay the price for construction, but also decomissioning up front in escrow, insurance, licensing, geological/ecological studies, etc. It isn't a fast do it now type of project.
For power like 15MWe natural gas units you referenced, those are approximately $0.45-$0.55/kWh to run, and more if rented, sometimes up to $1.15/kWh. My company has many, and they're only used during high peak hours probably a few times a summer to sell power to other markets because otherwise they're not even profitable. These are not cost effective to run AI datacenters, or really much of anything. They're meant for emergency power, remote wilderness projects, etc.
For max efficiency, you can get that with well engineered baseload plant designs larger than 100MWe. However, smaller units are still more expensive to run since you're not maximizing your output compared to operating costs. A plant 750MWe in size or larger is much more justifiable but it will take several billion dollars in investment. You not only have to pay the price for construction, but also decomissioning up front in escrow, insurance, licensing, geological/ecological studies, etc. It isn't a fast do it now type of project.
- Joined
- Jul 17, 2006
- Messages
- 14,429
But if you believe they're profiting from AI, they aren't. Stock prices aren't revenue from a product or service.
Already:
View: https://youtu.be/c7W8jpsjtCc
View: https://youtu.be/PtdI0KZRTLU (his testing is fun, involve making the ai agent make 3d model that he will print and test, making drivers for old hardware or that are just missing some functionnality for Linux, making executable to be burn on CD and to run on old XP machines and stuff like that that would have been hard to have been maxed by the labs on purpose)
The best in class compute show up in how aggressive they can go price wise.
Not ahead, but possibly the best quality/pricing at the high end/frontier of all possible option, cost about the same as using Kimi k3 on a provider, cheaper than full qwen 3.8
Yes colossus 1 has been widely profitable, they rent it for 1.25 billion a month, it did cost about 3-4 billion to build. We would need to go achieved to be build and been running for a while case by case, but they tend to have been doing really well.
if we look at google cloud revenue growth:
and their operating marging:
Why a megasite like an Amazon datacenter, that has the ability to built its capacity in advance and expected to scale up to a gigawatt+ over time go with 15MWE natural gas units ?, they need 200-400+MW just for the phase 1 and should get into gigawatt over time. Yes they would rent those for emergency extra, but not the type of solution to be built here.
Larger than 100MWe that exactly what we are tallking about for those megasite and yes they will use billions dollars in investment and they did get the approval to do it, it is not some theorical.
As for the numbers shown, google-meta revenues from AI since 2022 would be really hard to estimate, as all their ads business is an AI business, search quite a bit, everything they do tend to use AI, but they would be dramatically be bigger than this,
Last edited:
When you are looking at new demands that are 500% of your record peak demand and capacity, the fact that 75% of the 500% is probably bullshit doesn't make it workable. The real demands are stil asking to double your grid capacity. And that's assuming you are correct that 80% of them are CYA bullshit.
DukenukemX
[H]F Junkie
- Joined
- Jan 30, 2005
- Messages
- 11,433
Not to the point of breaking their economy. And yes, China's court rules that employers can't fire employees due to AI, which is something we do often here in USA to a fault. Meanwhile, USA has started to use AI to look through footage of people to issue arrests and tickets. Los Angeles Police Department (LAPD) audit did reveal a 32.3% false-positive rate specifically for stolen-vehicle alerts. Yet, we still use AI with that many false positives? AI will fail because it doesn't work well enough.
they cannot break an particulary contract they signed with a particular employee is not some large rule about firing employee due to AI, the precendant set here is thet cannot use simple LLM able to do what a worker could to use the force majeur-catasthropic unpredicted event close those contract often have.
and China use AI to score citizen and give tickets of course, using AI versus worker looking at picture of cars from photo radar is a small distinction. They are big on facial recognition, license plater reader with LLM analytics of them. I do not think it is the issuing of arrest and ticket part... that most people fear, if the plate is only read when you go to fast instead of paying a cop $60 an hour to look at the pictures, people would not necessarily mind, it is reading plates info outside the context of giving a ticket to its owner that start to be a big of an issue, would they read, look in a encrypted way if it match an open warrant and that info get deleted right away if not... that also be something else.
For example:
View: https://medium.com/syncedreview/ai-photographs-chinese-jaywalkers-shames-them-on-public-screens-ad0a301a46a6
AI Photographs Chinese Jaywalkers; Shames Them on Public Screens
- Joined
- Aug 18, 2006
- Messages
- 11,642