'Three New York Cities' Worth of Power: AI Is Stressing the Grid
wsj.com
wsj.com
These companies that have many big data centers are now big players. If they are stressing the grid they will need to be part of the solution in expanding the capacity and infrastructure, either by paying more to the electric companies or by including power infrastructure in their vertical integration. Microsoft has the right idea by investing in Three Mile Island, I think other big players will do similar things.
I'm only half joking as from what I've seen of those involved, I would not put this type of consideration past them.
there is a lot of money behind AI, but i have yet to see any of them actually get us a use case which work reliably.
Not to mention, AI is fun and all, but that energy could be far better spend making the economy less reliant on fossil fuel.
Classic HN. You personally can't see the value of X therefore the entire market for X is wrong.
And, if you want to pay, you can pay a premium to have more 'secure' power. This is an option today, already. No need for any discussion about markets and 'government is holding back the free market'. There is already a market to pay for premium power. And the utilities can charge AI customers more if they want.
a part of the energy grid cannot handle the load being generated by all the solar being deployed, and upgrading the grid takes years. In the meantime, bussinesses and even households simply are not getting a new power connection unless they are deemed of enough importance.
There have been many years of build up to solar power. If the people running the grid haven't done their job then replace them with people who can.
It's shared infrastructure, mainly paid for by public funds. If a government decides the priority should be the millions over the few, that's what happens in a functioning government.
Or they can pay for their own power stations and infrastructure to be built.
If the power grid regularly can't handle the customers it has them it isn't a functioning power grid.
there is not and should not be an "open market" for power. the sale of electricity from the power grid is highly regulated, and for good reason. delivery to residential customers and essential services should always be prioritized over industrial bulk purchasers.
bulk purchasers already have the opportunity to use power sources that are not connected to the grid, if they want to use grid power they should be prioritized according to their benefit to society, not their cash on hand.
And sure, tell them they get disconnected first when there's a shortage. But a hard block is worse than taking their money and using it for improvements.
Usually you do the opposite: In exchange for a lower rate you will be the first disconnected in an emergency.
The fact that they have the capability to pay for and consume this power is the strongest signal that they are indeed providing value to society. The money comes from somewhere. If that turns out not to be the case, they have the strongest incentive to stop.
But this sort of rationing at the hands of some regulatory body has historically proven to be a far more deadly cure than whatever the presumptive disease was.
money is not a good indicator of value provided to society. there is a ton of ways where it falls apart, and there are a lot of people devoted to finding the ways they can collect the most money while providing the least benefit to society.
Ok, it's the least worst system we have...unless you have an alternative. I'm all ears.
we don't need to invent a whole new economic system. we can just say "this is bad, and we should do less of it". and making sure that power goes to people's homes before it goes to datacentres is one of those things. and the best part is that's the way the world already works, so nothing even needs to change.
Public vs private isn't black and white, it's a spectrum. Energy providers sit on this spectrum. In some way, they're public infrastructure. In others, they're private companies. Slightly pushing them more towards "public" via regulation is what we HAVE been doing for a long time. This shouldn't be setting off your communism senses. If it is, you have a fundamental misunderstanding of how the economy works in regard to these kind-of-public-infrastructure goods and services.
Capping prices and participants is pure stupidity. Price controls and rations never work
In a Western society, everyone's needs are supposed to be equally important. We often don't even make majority decisions, because we assume that the constitutional rights of a minority are more important than what the majority finds reasonable.
What do you think prices do? What do you think income is reflective of?
Companies that make lots of money provide things that lots of people are willing to pay for. This is the reflection of what people need. Exxon makes money because people want to stay warm. Cargill makes money because people want to eat. Apple makes money because people want to be entertained. Eli Lilly makes money because people don't want to die.
What do you know that millions of people all acting in their own self interest don't know?
It isn't meant to be fair. Fairness blunts the signal.
In our current society, children can benefit from the wealth and social connections of their parents. That gives the needs of some people additional weight in the market without any actual merit. To get rid of that market distortion, children would have to be raised collectively without any contact to their parents.
There is a lot of wealth that has been created with business practices that later became illegal. Because we prefer not to punish people for actions that were legal at the time, we let people keep that wealth. But if we want the market to prioritize the needs of the current society, such wealth would have to be confiscated.
A society where the market actually reflects the needs of the society would be incredibly dystopian.
That education of the poor is important, perhaps?
In reality, I'm pretty sure they do know that. Which is why they vote for politicians ensure the poor are educated, regardless of the fact that that can't necessarily pay for it at the time. They don't rely purely on laissez-faire capitalism to determine all our social policies.
The Environmental Protection Agency (EPA) has many accomplishments, including: Reducing air pollution The EPA has reduced air pollution by regulating auto emissions, banning pesticides, and cracking down on factory and power plant emissions. New cars are now 99% cleaner than 1970 models, and the EPA's efforts have led to a decrease in smog and asthma cases.
Cleaning up toxic waste The EPA has cleaned up toxic waste through the Comprehensive Environmental Response, Compensation, and Liability Act (Superfund), which provides funding for cleaning up abandoned waste dumps. Protecting the ozone layer The EPA has protected the ozone layer by reauthorizing the Clean Air Act in 1990 to phase out chemicals that deplete it.
Increasing recycling The EPA has increased recycling. Revitalizing brownfields The EPA has revitalized brownfields, which has led to increased property values, reduced impervious surface expansion, and reduced residential VMT.
Reducing water use The EPA's WaterSense program has helped Americans save water, with a cumulative savings of 8.7 trillion gallons as of the end of 2023. Banning cancer-causing pesticides The EPA has banned or restricted many cancer-causing pesticides, including heptachlor, chlordane, EDB, daminozide, and chlorpyrifos.
Testing schools for asbestos The EPA required all primary and secondary schools to be tested for asbestos starting in 1982.
Rating energy efficiency of appliances The EPA introduced the Energy Star program in 1992 to rate the energy efficiency of household appliances. EPA History: Documents about Agency Accomplishments | US EPA Nov 17, 2023
Generative AI is experimental.
Gullible investors for now.
This is not how the grid works in the US. Industrial is very frequently critical and gets higher priority over residential. Outages at major industrial operations can cost millions in lost revenue and can cause direct damage to hardware.
The grid absolutely is a market and should be. Reliable power at fixed prices is a thing you can try to buy from it, but that’s by no means guaranteed.
...but generally not a free market in most jurisdictions. The Texas grid is an ongoing experiment, and yet even the Texas legislature had to step in to bail out[1] providers' massive losses from with very high energy spot prices - it turns out the demand for energy is pretty inelastic at below-freezing temperatures.
1. Via a surcharge on all future customers. That's not very free market of them.
In Texas it’s just a different pricing structure with instantaneous pricing. That’s not more or less free market.
The state government stepped in to socialize the losses incurred by private companies during the last snowstorm. Perhaps we can call it "Free markets with Texan characteristics"?
* Except big players will never go under, and substantial losses are socialized
In exchange they get a much better price on energy with the respective risks attached.
Part of the problem is the AI companies are unwilling to sign long-term power purchase agreements. That means the utility is on the hook for both the capital cost of the development and the risk of unloading a bunch of excess power if the tech company decides to relocate.
That may also be circumstantial evidence that they don't quite believe all their own hype.
No company, particularly one with negative cash flows, wants to saddle itself with long-term liabilities.
There might an opportunity for useful financial engineering: instead of a fixed power purchase agreement, one which floats and is married to a no-shop commitment.
For example: start-up agrees to purchase 10 units of power at a fixed price for 10 years. It has the option of only purchasing 9 units, but must pay a 20¢ fee on the unpurchased power. In this way, if the start-up hits a rough patch, the burden is shared in a way that is meaningful to the start-up's surival probability. In exchange, the start-up commits to not reducing its power purchase while ramping up usage with someone else, presumably at a lower rate, and, once lowered, commits to ramping up its usage with this provider before ramping it up with anyone else.
Now here is the critical part: the utilty doesn't hold that risk. Instead, it pays a fee to a hedge fund or endowment or whatever by which that party agrees to make up the 80¢ (or some fraction of it) on unpurchased power.
In summary, the utility winds up with a reliable cash stream in all worlds except the start-up failing. (One could mitigate that risk by requiring some pre-funding.) The start-up gets a long-term, guaranteed-price source of power. The hedge fund gets a weird financial thing to play with. (A natural provider of that guarantee would be one of the tech giants. I like the idea of it going into the market because one can see those contracts being aggregated and thus diversified.)
Google could entice someone to invest in generation capacity for a new AI datacenter by agreeing to buy their output for a few years, or building their own generation. Then if the AI thing doesn't work out, use that power for their existing data center uses, i.e. move the existing servers into the AI data center. That would be essentially shifting the risk to the power company in the location where the existing data center is, even if that's not the same one they have the power purchase agreement with.
We should work out a mutually beneficial agreement - I think it makes no sense to nationalize energy providers off the bat. But, part of that agreement should be ensuring energy for residential areas. And if that requires cutting off energy pigs in the meantime, so be it.
Others who are regular readers of IPCC publications might value the survival of big parts of earths habitable zones and the most densly populated costal areas higher. But priorities differ.
I happen to agree that Bitcoin has very dubious utility, but I suppose it has nailed the "store of value" use case, whereby you can send it anywhere without any friction, and also if you need to, say, escape the country during war or a natural disaster, it's much easier to remember 12 words than try to sneak gold through customs, while the cash you store in your house might be underwater. Bitcoin for many institutions is a digital commodity to diversify for tail risk etc.
Why would you do that when there is a strong likelihood that most of these LLM startups will be out of business within the next year or two.
In my thought process, an analogous statement might be, “if you ask for 10 kilowatts from the power company, that’s the power company’s problem. If you ask for 10 megawatts, that’s your problem. If you (an AI company) ask for a historic amount of power and your business fails if you don’t get it, then it’s your responsibility to ensure the grid can supply that much power.
The analogy with the bank just doesn't hold up.
the analogy sort of works if you flip a few things around, but it obscures more than it elucidate at this point.
And that single customer is paying more than likely millions of other customers.
Notably, if the grid has a good PR/marketing department, those customers are a lot more likely (individually) to take a black out ‘for the greater good’ than the data center.
See California’s rolling blackouts, etc. in the past.
Data centers/supercomputers/similar facilities have agreements to ensure uninterrupted power.
The grid would sooner cut off residential power than risk running foul of SLAs.
Most developing countries are very familiar with this, the technical term ‘load shedding’ having entered common parlance because of the frequency of power cuts faced by consumers.
> Every problem has an ‘obvious’ solution on the surface after thinking for ten seconds.
Is kind of a silly statement.
What’s the obvious way to automatically detect pictures of birds? Because diffusion and transformers are not it.
My cousin just (last year…) tried to do this as some extra credit high school thing.
The ‘obvious’ solution was detecting edges with a python library and manually writing up 4-5 functions to identify features.
It’s surprising that this even comes close to working.
Even then, grabbing an existing package is just offloading the solution discovery to someone else.
And offloading tricky logic to libraries is kind of how we build all software.
I’ve been programming since I was 9 and I don’t think that would’ve occurred to me when I was 17 lol
And yeah that is how we build all software, but just because someone else made a solution to something doesn’t mean that solution is obvious.
While your suggestion is the morally correct thing to do, you’re forgetting the other option makes more money and will always be taken.
In renewable energy it is actually common for the funding institutions to be paid a cut of the profits and to have no recourse to repossess assets. They take all the risk because they provide all the funding. The job of the actual developer is to provide a project worth funding.
To expand upon your bank metaphor, let’s say the bank are energy utility companies, and the customer that owes $2bn is Company A. Yes, this is the bank’s problem in a vacuum because it needs to remain solvent for its other customers (provide energy) while still collecting on its owed debt (the power bill from Company A). However, the bank also has an obligation to reduce the total amount of currency that changes hands in the first place (reducing energy use) or finding more efficient ways of exchanging that currency (renewable energy). With that broader scope in mind, Company A charging $2bn in debt isn’t just a problem for the bank, but a problem for that bank’s customers, and any customers of any other banks that do business with the affected bank.
In other words, the problem isn’t merely hyperscalers and generative models gobbling up vast amounts of power or purchasing nuclear power plant capacity, it’s that we sorely needed that capacity to wind down total fossil fuel usage and to focus on efficiency of resource usage until we addressed the climate crisis. That’s where I (at least) have a beef with the current state of datacenters, hyperscalers, and generative models. So long as compute remains insanely cheap and “infinite” in scale, we lack any incentive other than financials to reduce consumption or increase efficiency. And financial incentives only work as long as the penalty is significantly more expensive than the insurance.
If tech companies get into power and break down the regulatory capture of utilities, they can make money off of energy storage and virtual power plants while providing free power for people.
all communication is free
Edit: OP's comment is almost nonsensical I feel it might be AI. "Traditionally youtube would cost hundreds of dollars"??
For data, Compuserve charged $5/hour for access, regardless of whether you used 300bps or 28.8kbps. So a 1MB download could cost anywhere from 50 cents (28.8k at night) to $10 (2400bps during the day). And that was in addition to the cost of having a phone line in the first place, not to mention the hardware--a decently fast modem cost several hundred dollars, and needed to be replaced every couple of years as the tech was improving relentlessly.
A service like Youtube with on-demand streaming video of just about any content you can imagine would certainly be worth hundreds of dollars a month. Hell, cable television with 50 channels was $30-50/mo (or more depending on whether or not you splurged for HBO/etc), and that was whatever grab bag of content they decided to include.
Remember too that all these are 1990s prices, so you need to at least double them to account for inflation.
Yet we're in a conversation about data centers, and our concept of "unmetered communication" only refers to one end of the line.
What does it cost companies and orgs to communicate with us? What does it cost them to keep a server operational, functional and sending data to us? It's "free and unmetered" for you to watch YouTube, run yt-dlp, put up your ad blockers and PiHoles, but what cost for everyone but you?
I would also consider time/effort costs in communicating. Compare having a rotary-dial phone, the ultimate "thin client", with a general-purpose computer or smartphone. How much time and effort do we put into fiddling with them, upgrading them, fixing settings, troubleshooting whatever communication software is on top of them?
Newspapers cost us $0.25-$1.50, but that's a highly subsidized cost like riding public transit.
it is effectively unmetered
And that was just your access costs, you still needed to pay for the phone service. According to the FCC[1] you could expect to tack another $20 / month on to those costs, just for your local phone service. And all of that would have probably been enough to service one user online at a time. If you wanted multiple people in your household to be online at the same time, you at a minimum are going to want multiple phone lines, never mind the additional hourly access charges.
[1]: https://transition.fcc.gov/Bureaus/Common_Carrier/Reports/FC...
[2]: https://www.independent.co.uk/advisor/vpn/screen-time-statis...
I don't necessarily see anything wrong with that, and I can see the communication of demand changes back to generation being useful to those running the plant(s).
https://news.ycombinator.com/item?id=41601443
Microsoft's commitment allows Constellation Energy to restart the 819 MW nuclear fission reactor at Three Mile Island. This reactor operated without problems for approximately 45 years and was switched off in 2019.
The other reactor at Three Mile Island (880 MW) melted harmlessly in 1979 due to operator error. That reactor will never be restarted: https://en.wikipedia.org/wiki/Three_Mile_Island_accident
Clean baseload power is extremely valuable.
Nuclear fission is the only clean baseload power source mature enough to be widely adopted. Today there are 440 nuclear fission reactors operating in 32 countries. Having used fission reactors for 70 years, we know how to build them and operate them at 95%+ efficiency (https://www.energy.gov/ne/articles/what-generation-capacity). For example, Vogtle 3 and 4 have been operating at 100% of their capacity.
In contrast, solar and wind are not baseload power solutions. They are intermittent and unreliable, fundamentally weather dependent. If you have battery backup sufficient for time T and the weather doesn't cooperate for time T+1, your data center is offline.
Furthermore, even a day or two of battery backup eliminates the cost advantage of solar and wind. As a last resort, fossil fuel backup plants can be built, but building them is costly and using them is dirty. See Germany's power grid for an embarrassing real-world example.
Solar + wind + batteries is an intrinsically unreliable, unacceptable solution. Please see Yann LeCun's comments, reflecting Meta's view on this topic: https://news.ycombinator.com/item?id=41621097
Being 5x cheaper more than makes up for that though.
None of the viable modes of storage (pumped water, hydrogen, batteries) are expensive enough to make paying 5x per kilowatt hour worthwhile any more.
If you roundtrip solar and wind via hydrogen stored for months and back into electricity again (the most expensive mechanism), it's still cheaper than every single kilowatt hour of nuclear power delivered on the sunniest, windiest days.
It really is difficult to overstate just how mind bogglingly expensive nuclear power is. Unless it's used to cross-subsidize the military nuclear industry it just isnt worth it.
long transmission lines significantly reduce efficiency and add massive costs as well as creating easy targets for terrorism.
You don't store energy from solar on a sunny winter day to use at midnight when the wind is blowing hardest and consumption is lowest. Your store energy to use on those rarer hours when it is neither sunny nor windy and when consumption exceeds the production of both.
This can be done cost effectively and at scale in two ways - 1) by pushing water uphill for short term storage and 2) by manufacturing hydrogen and storing it underground for longer term storage.
So, you combine overproduced solar, wind, pumping water uphill and electrolyzing and storing hydrogen to match demand with supply.
Solar and wind and their transmission are distributed. This makes them poor terrorism targets. A nuclear plant, on the other hand, is a great target for terrorists.
I feel sure Ontario does not have nuclear weapons ambitions.
> If you round trip solar and wind via hydrogen
Do you have any examples of this at commercial scale (hundreds of MW) and their costings?
Canada actually used to be a nuclear power until 1984. Since then it has signed up to the NPT and disavowing weapons but maintaining a nuclear industry that would let it spin up a nuclear weapons program in a hurry should it deem it necessary (i.e. if the American nuclear umbrella faltered).
There are several countries around the world that follow a similar policy - Japan, South Korea, Sweden, Iran. There's value to belonging to the NPT but there's also value to being able to get a nuke in months if an existential crisis presented itself (as it has for Iran).
>Do you have any examples of this at commercial scale (hundreds of MW) and their costings?
Google for lazard LCOE.
No, it wasn't, except in the sense that Belgium, Italy, the Netherlands, Germany, and Turkey are nuclear powers today. It was a participant in a nuclear sharing agreements (under the auspices of NORAD, rather than NATO like the other countries) where US-owned and controlled weapons were deployed at Canadian bases, under control of the US military, with plans for transferring them for delivery by Canadian forces in the event of war. It never had its own nuclear weapons or, AFAIK, nuclear weapons program
I guess LCOH is a more relevant search term. https://www.lazard.com/media/2ozoovyg/lazards-lcoeplus-april...
Page 28 is informative. 220-235 USD/kWh for a 25% blend of unsubsidised PEM hydrogen, or 210-221 USD/kWh a 25% blend for unsubsidised alkaline hydrogen.
France would like your phone number.
It’s nuclear or fossil fuels until an order of magnitude change in chemical storage.
No one is going to be building enough gas peaker plants sitting idle 360 days a year to achieve the reliability of the grid most consumers and modern civilization require. So it’s burn more gas or coal on a regular basis or enjoy power outages if you want what we have today.
I love solar. I plan to have a week or so worth of battery at an off grid home someday soon. It’s both entirely financially impractical (yay disposable tech income) and I will still back it with an expensive propane generator since I live over here in reality.
I think there is a small matter of national/international politics (running nuclear as renewables backup), a lack of focus within EDF and perhaps a dash of French industrial action (at least a go slow on the refuelling outages).
Project Hercules was meant to solve the latter https://www.euractiv.com/section/energy/news/ceo-of-edf-defe... but this never happened.
On them positive side their nuclear generation is recovering a bit, albeit from a low base. https://www.edf.fr/en/the-edf-group/dedicated-sections/journ...
You nuke fans always conveniently leave out the costs of cleaning up the mess that the civilian nuclear generation program has already created. Last I heard it was about half a $T.
That’s a pretty damn low utilization rate. The cost per Mwh would make the Texas polar vortex incident look downright cheap. Or you’ll be having the government build and operate idled natural gas (or coal!) plants via taxpayer money since investors are generally not quite that poor at math. An idle plant still needs the same number of staffing and engineering as one operating at 90% capacity factors, before you even begin to look at capex.
Perhaps that will happen. Running your baseload generation balls to the wall at 100% seems to be a far better use of capital to me, and have your cheap intermittent power sources fill in for those peak loads that don’t require reliability.
I can come up with all sorts of great ideas on how to utilize spare solar capacity. None are financially practical in the scale of a human lifetime at a societal level, even with a magic wand to instantly enable the political will to do so.
Humans continue to amaze me with their inability to account for tail risk. The same political magic wand makes nuclear viable within a decade or two, accounting for said tail risk within it.
That said, the current unserious political climate gives me zero hope any solution will be found, so I’m simply preparing for semi regular power outages becoming a thing within my lifetime.
This is what the capacity market is for; ensure the plant needed to operate the grid can pay the bills. https://www.utilitydive.com/news/pjm-interconnection-capacit...
It does feel like an inefficient use of capital. My best explanations for intermittents are 1) they are fuel savers, making natural gas cleaner by having them run less; 2) they siphon money into the generator’s pockets (from taxes via the Inflation Reduction Act and other subsidies, from other electricity generators via renewable energy mandates, from the utilities via rooftop solar which reduces the consumption when energy is easier/less costly to produce).
ok so weather is predictable and highly accurate for at least 3 days in advance. You are still able to buy energy from markets. You can still load battery storage in case of bad weather forecasts.
I don't know what the problem is here? Why do you paint a scenario where data centers magically are now off the grid and have to be on 100% reliable energy sources?
Wow. Really chugging the coolade.
CEO of VerdictAI is quoted as saying “shit happens, if they have any value as humans they would have enough money to get out of this, so the AI must be right. One must not lose sight of the fact that we have freed Americans from millions of hours of manual labor with our InmatesAI volunteer workforce.”
I’m not saying there aren’t concerns, but risks can be mitigated and precautions put into place to prevent catastrophes. Stopping development because of valid concerns is not the right move. It’s not even the right move to stop developing weaponry which has a stated goal of causing direct harm to other humans.
There are people in the world who will attempt to do these things and much more dangerous things recklessly and with far worse intentions than making money, and if those people get technological and military superiority over The West, the world will look very different than it does today. There are a whole lot more ways the world could be worse than it could be better.
The current world order is not ideal, but think about what the actual long term outcome is if The West stops.
Nobody is saying “uh uh, no more problem solving”,drawing that conclusion would be uncharitable at best, and “wrong” at worst. The argument is “gosh, maybe we could have some innovation that’s actually useful, and not some polished-turd/enshittification”.
We’re still a democratic republic that can stop the enshittification of utilities (or whatever the problem de jour is). There are bad actors, but by and large, things are working and continue to be improving. Some things need to be bolstered, some things need to be replaced. Stagnation is not the answer.
What would you propose as the solution? All new ideas go through a centralized approval process? That experiment is already being run in other countries. If they start winning, the US will adapt or die.
You know, because of the global warming and such.
I think you were too busy chugging that contrarian kool aid to spend even one moment thinking about what you're arguing for.
Tech companies and their cults are perfectly represented by the South Park episode about Priuses
This is a "workable strategy" only if you can find a sucker to eat the losses.
Not sure where you are but where I am you still have to pay for cell phone data plans and internet to the house.
They're not going to build a second grid, or power plants for everyone else. That's an awful business. These companies are in good businesses. They don't want to get into an awful business.
There's so much wealth to be created given the multiplier these SMNR companies are likely to have. And the tech companies will invest in the SMNR companies they purchase power from. Some of them are already invested.
Hopefully, they'll have enough scale that whatever they need to do makes it cheaper for the grid at large.
Microsoft have already entered into an agreement with Constellation Energy which pays to restart one of the reactors at three mile island. This comes with a 20yr purchase of the power from that plant.
There aren't many usable and idle nuclear power plants sitting around. It's extremely likely, as you point out, that states will soon be competing for jobs by offering land for both data centers as well as new power plants.
The best estimates I've seen but new construction of nuclear plants at 5-7 years. If this could safely be sped up then I assume big tech is already working to make it happen.
Of note, there are already some third parties/regulators working to stop big tech from capturing large shares of existing power generation[1]
[0] https://www.spglobal.com/commodityinsights/en/market-insight... [1] https://spectrum.ieee.org/amazon-data-center-nuclear-power
If tech companies got into power, the inevitable result would be that they embrace and extend the regulatory capture of utilities, with the explicit goal of maximizing their profit at the expense of human needs. As well as of course the evisceration of as much pesky environmental regulation as their lobbyists can get their hands on, to the extent that it serves as an impediment to this paramount goal.
Their philosophical stance on these concerns and tradeoffs has already been made palpably clear, in any case: https://news.ycombinator.com/item?id=41552642
From everything we know so far, these people definitely will not do the right thing without the swift and exacting application of a very heavy stick.
In power generation there is little appetite for innovation from producers since they are either monopolies or government agencies. So instead all innovation has been going to local generation like solar and wind.
As if any tech company would build anything but solar. You don't need to bend the rules here and you don't need too much storage either. You just need cheap land, a fuckton of solar to get your peaks to your base consumption and you already do 60% of energy renewable. Anything above that is battery storage and more land.
It's such a no brainer in cost and from an operations PoV the only thing that is easy enough to go into power.
Internet data centers are fueling drive to old power source: Coal - https://www.washingtonpost.com/business/interactive/2024/dat...
From coal plant to data center: Old power stations are being repurposed - https://arstechnica.com/ai/2024/08/from-coal-plant-to-data-c...
Internet data centers are driving the resurgence of coal power - https://www.dailyclimate.org/internet-data-centers-are-drivi...
Meanwhile, both solar and wind also have their environmental externalities (from the mining of raw materials, to the sheer amount of land that needs to be put under shade for solar farms, and the noise pollution caused by wind farms), and correspondingly, tons and tons of rule for these companies to bend.
I don't think tech getting into power will make things better. Just look at what is happening in Texas and other parts of the world: grid operators have to pay bitcoin miners to turn off their farms when grid capacity is overwhelmed, which drives up cost for everyone. Depending on where data centers are located, construction of new transmission lines becomes necessary. All this increases the baseline operating cost of the grid. Personally, I don't want more AI generated content or timelines powered by AI driving toxic engagement because it gets more clicks. The "improvements" we're seeing now are beyond the point of diminishing returns. Tech will abuse any externalities that exist worth exploiting, and society will end up paying for it.
If you're "cord cutting" and not paying for cable TV anymore, then you likely still come out ahead even after folding a couple of streaming services into the mix.
If all you have are 1990 expectations, then you should be able to make do with something for about $50/mo, maybe less. Either a cheap phone plan, or entry level internet and VOIP. And it'll still be leagues more useful.
That private phone line you shared could only be used by the one computer in the household, if there even was a computer, which more often than not there wasn't, and when it was online, you couldn't take phone calls, which was more of a problem than now, because it was 1990.
And we don't live in fear of our telephone bills like we used to. We don't charge long distance, by the minute, or per text. Data caps are annoying, but they're not life-changingly expensive, and hey, if you're living it up like it's 1990, you'll never get anywhere close to exceeding them anyway.
I can very quickly generate a lot of revenue if my investors will let me buy eggs at $0.50 each and resell them for $0.05, but I don't think I'd get any bites unless I could explain how that enterprise will eventually turn a profit.
People underestimate how hard it is to sell things: you could have the cure to cancer in your hands today and still need to lose mllions to get anywhere with it.
Otherwise known as investing. Sometimes it works. Sometimes it doesn't. And the last 20 years has been chockfull of excesses absolutely. But Uber is profitable these days, and has genuinely transformed how people get around. It looked like a negative unit margin dog until it didn't.
> If there are no drivers, there are no customers. With no customers, who would drive?
No worries, Waymo is coming soon to kill Uber as well as everyone Uber took out of business.
So they are currently printing 1 dollar bills at a cost of $2.35 each.
Jerry: How is it a write-off?
Cosmo Kramer: They just write it off.
Jerry: Write it off what?
Cosmo Kramer: Jerry, all these big companies, they write off everything.
Jerry: You don't even know what a write-off is.
Cosmo Kramer: Do you?
Jerry: No, I don't.
Cosmo Kramer: But they do. And they're the ones writing it off.
Not a value statement on whether is this is right or wrong.
That is my point.
That's orders of magnitude less than the thousands of employees and billions of dollars being deployed with no clear business model or profitability in sight.
They're paying $2.35 in hopes that the $1/year customer sticks around for more than a few months.
If you're only referring to the big players creating foundation models, like OpenAI and Anthropic, it's mostly speculation of future improvements and revenue, because the amount of money they've raised pretty much requires massive returns. This is similar to how most venture capital works, e.g. Amazon was laughed at for being unprofitable for the first, what, 10 or 15 years of its life.
If you're referring to a broader set of AI companies, then most of the more "normal" AI companies have raised far less money, and are usually eyeing more clear use cases and are sometimes already generating real revenue. For them, I'd say it's not all speculation, a lot of it is productizing and monetizing this new technology that's come along, and is the obvious "low hanging fruit" that actually needs to get done.
Most other noticeable(drummed up in news media and various other prominent online media) are mostly wrappers around big players' APIs, more commonly OAI. I have seen some using Mistral(mostly nerds), but if I were to say that if OAI sank, a broader set of AI SME will sink with it, because they are symbiotic.
There are some covert "old fashioned" AI companies I know of, who are still stuck at previous era CV(majority) suddenly gained a lot of interest(+funding), but they are more well positioned to continue revenue generation compared to the more hyped smaller ones.
If we just sacrifice, 5 more New York’s worth of power, and the intellectual output of a generation, we will _totes_ get an AI that will finally tell us how to solve global warming!!!!!! (In a way that doesn’t hurt my feelings/money-machine like “use less fossil fuels” and “stop emitting carbon”) /s
... I already can sense somebody typing "but AI will help us design those better!!!1!!"
Yet here we are, pouring money, resources, expertise, and people hours into an endless pit of hopes and possibilities of a possibility.
Any day, now.
Which leads us to the GPUs. For a while everyone said ASICs (and the courageous ones, FPGAs) were the future of mining because they were inherently more efficient than GPU hardware could ever be. But it turns out that it's the other way around - mass-manufactured GPUs have access to more efficient silicon than any FPGA on the market could hope for. Similarly, AI chips that want to beat Nvidia on efficiency terms first have to climb a mountain of power usage that is inherent to their exponentially less-dense hardware.
With bitcoin its more arbitrary but you can see how they could hard-fork to a different hardware-favoring system if desired.
And the CPU/GPU/FPGA/ASIC thing is something that depends on the proof-of-work. Bitcoin went for the naive solution and it meant ASICs were the optimal solution once a certain amount of income was available. Later cryptocurrencies learned from this and specifically designed their proof of work systems to be optimal on a GPU, or even a CPU, because making ASICs the optimal solution is generally pretty bad if you want a diversity of miners.
Nuclear plants pair well with big water infrastrucure projects. Forced convection is just a really effective form of heat transfer
You need nuclear reactors or other base line power generation energy sources.
People understand the fact it costs money to dig out coal from the ground and bring it to a power station. Yet they don't understand it costs money to take power generated at noon and shift it around. This is never included in the price for solar or wind.
The people who make those predictions are never held to account and can't find their ass with a GPS.
One only need to look at the fact that the best predictor for energy generation prices increasing is how much renewables have entered the grid. Anything over 5% and you start seeing the costs rise upwards.
> We find that renewable energy is associated with a significant reduction in wholesale electricity prices in Europe, with an average impact of 0.6 percent for each 1 percentage points increase in renewable share. We also find evidence for a nonlinear effect—that is, higher the share of renewables, the greater its effect on electricity prices. On the other hand, while quantile estimation results are mixed with regards to the impact of renewables on the volatility of electricity prices, we obtain evidence that renewable energy has a negative effect on volatility at the highest quantiles.
Having people betting on energy markets sounds like a bad thing.
When the people betting (and apparently winning those bets) also have false beliefs about renewables raising wholesale energy prices that are contradicted by evidence?
Well, at the very minimum, it's not a good look. At worst it's Enron.
In Australia the introduction of grid batteries made its money back in slightly over a year not because of buying and selling electricity, but from undermining the price setting collusion of the gas operators.
If your bet was in the form of "look, the weather's bad and there's not enough grid storage", obviously so — that's why all the battery factories are getting built :P
The cost of those batteries is low enough to disregard your conclusions in the long term.
Blacksmiths could bet their horseshoe business against the automobile, until they couldn't.
The price is the cost of transmission and/or storage. Even if you're also ignoring wind because dunkelflaute.
PV with storage costs about the same as nuclear, with broad variation between specific projects that means each can be a close winner in different cases.
A global power grid could, in principle, be built for approximately $239 billion worth of aluminium (caveat: only China makes enough aluminium to do this), with 1 Ω electrical resistance going the long way around the planet.
That price tag is 5-9 Sizewell C reactors, based on current cost estimates. That reactor took 14 years to go from "government wants new reactors on sites of existing reactors" to "planning permission granted", and is planned to have merely 3.26 GWe nameplate capacity.
From a purely economic perspective, power grids are fantastic.
The logical thing is to place the solar panels where sun shines for more than 1/3 of the year. Additional considerations: DCs are never at 100% utilization,and the hyperscalers won't keep up with their current pace of investment in AI for the next 5-10 years - they will gladly push the power generation CapEx risk onto someone else (existing utilities), and let the taxpayers take the fall.
Now that's a whole bigger challenge than you started out with. The XLinks project [1] is a proposed 3.6GW 4,000 km undersea link from Morocco to the UK. The cost of the inter-connector alone is 11B GBP, so approximately 0.75M GBP per MW-km.
[1] https://en.wikipedia.org/wiki/Xlinks_Morocco–UK_Power_Projec...
"Why invest so much in a global communication network when regional servers can work cheaply, and still exchange data by snail mail?"
That's the system we have; CDNs, Google Global Cache [0], Netflix Open Connect [1] . Location matters, even for bits and bytes, because infrastructure is expensive.
Perhaps you are missing how large and expensive the cables are. See [2] for 2014 cable spec and costings.
----
[0] https://support.google.com/interconnect/answer/9058809?hl=en
[1] https://openconnect.netflix.com/en_gb/appliances/
[2] https://www.nationalgrid.com/sites/default/files/documents/3...
But AI fans will shrug and say “this has always been happening, AI is nothing new”. The usual refrain.
Also, AI is way more useful than blockchain etc etc.
AI intends to overthrow expensive people(costs) and make the bankers as well as investors plenty money. Climate change can wait, once AGI is achieved, it'll figure out how to solve all the climate issues and all will be well.
That being said, AI here is no different from crypto, only diff. is, AI has immense backing and lobbying. Even our EU climate friendly overlords are parroting that AI will improve everything and more adaptation of AI is needed in all businesses and investment must be injected into anything that has AI. So, all is well in the end.
How is this bad? Are the utilities just lazy?
The utilities are being asked to make massive speculative investments in transmission infrastructure, but are dubious that the demand will actually be there when those projects finish in half a decade, and they'll be left holding the bag.
Other forms of energy generation come and massive ecological costs. Are we really prepared to spend this much energy on a speculative technology? especially when it is destroying the planet while doing so?
this is a good way to ensure emissions become sustainable
I feel that energy generation (key component if we want to grow as a species) should be helped/sponsored by governments and not really expected to turn any profit.
The power is being used wastefully.
Sorry, I mean, "the market".
Is it annoying? Yes. Their trucks are heavy enough to shake the house when they go over a bump. They are loud as hell too. [1]
Do I want to shut the company down? Of course not. They are my neighbors, for better or for worse. I have to try to make the relationship work. They employ dozens of people and make it so other companies can build things easier in my city.
The instinct, however, is to think "I wish they would f** off with their big dumb trucks and their blocking traffic and their dangerous blahblahblah. I'm fine with industrial supply companies, but not in my backyard!". It takes an effort to stop yourself and remember to recognize their value. I could lose my job tomorrow, and maybe walk down the street and ask them if they need another set of hands moving gas cylinders around. I might one day need their services, or need the services of someone who needs theirs.
1. I wish the federal government would regulate the amount of noise pollution coming off some of these vehicles, yes, but that isn't a local issue.
And it's not clear to me at all why you want people who do not live in your city to force the people in the city to adopt rules they don't want.
Crypto still uses many multiples of the power consumed by AI, and unlike AI it can _never_ improve and provide more utility per watt since it is inefficient by design.
Energy will get so inexpensive in a few years (10?) from now that power consumption by AI will not be a problem.
Any evidence for that? I predict the exact opposite. The price of energy will rise.
Also, solar power is making a large dent in our power needs.
https://climate.benjames.io/solar-off-grid/
It's a start but give it time.
By what mechanism? Wishful thinking?
If so, to who, and why?
What's "bad-faith" about it?
https://www.amazon.com/Iron-Empires-Robber-Railroads-America...
it was hugely controversial in Congress.
I don't happen to know about roads, since that wasn't one single thing. But I doubt they happened because of blind faith in the future, either -- it was probably more businesses demanding roads NOW. So the argument that no one questioned them is dubious at best.
What I meant to say was that the AI bubble is feels different in that there seems to be relatively more skeptics versus railroad mania or the dot com bubble (but this could just be skewed by my own experiences).
Whether or not this suggests a flawed value prop for AI is another story.
> American Electric Power, which owns the line, has proposed a new, higher electricity rate for data centers and cryptocurrency miners that would lock them in as customers for a decade. The companies are balking.
Look at that! Net Neutrality comes to the electrical wire. They could have charged different rates based on industrial/residential or volume but that wouldn't be optimal rent extraction.
This is actually a really great article about the weird contract dynamics of large power customers and the chicken-egg problem of needing commitment for a certain amount of guaranteed spend to justify the cost of servicing them. It has nothing at all to do about AI.
But to directly answer your question, yes there is competition. In most states, Ohio being one of them, you can choose who provides your power generation. And since companies are shopping states for where to build data centers AEP has a lot of competition for delivery.
how is this stressing a grid that has to shed huge amounts of load to maintain availability off peak? instead of shedding the load, just price it effectively and the demand will even out and be a more efficient use of the fuel or generating station. the problem with our energy infra is a lack of consistent demand, and both AI and proof of work create the demand and a consumer of last resort that would make generation way more efficient.
afaik, the main compute load is done in training, not in the queries. this article sounds like astroturfing a pretext for regulation under the auspices of "climate."
also, the diminishing value of cash is accelerating so taking todays dollars and putting them into tomorrow's energy is probably the most fundamentally sound economic use of capital today.
power generators were loaded with debt and used as public sector slush funds (in canada where they were nominally publicly owned) and now the companies are not very productive because they have to pass on the debt service cost to consumers, and prevent anyone from entering the market to offer competitive service that would interfere with that debt service. if you build new energy infra, you need to protect it from managers piling it with debt to loot it, but if you solve that real problem, demand is not a "problem," it's the most valuable thing in the universe.