816
submitted 1 day ago* (last edited 1 day ago) by inari@piefed.zip to c/technology@lemmy.world
top 50 comments
sorted by: hot top controversial new old
[-] AlteredEgo@lemmy.ml 5 points 2 hours ago* (last edited 1 hour ago)

Obviously the AI rollout should be slowed down and AI use regulated, like stopping any replacement of workers by AI.

But China is set to win the AI race with their open weight models on par with the current frontier models. And they are building nuclear power and solar and wind like no tomorrow. And they just started to produce sodium ion batteries to solve the grid storage problem.

Spurned by the embargo they are developing more efficient algorithms and "Extreme UV lithography" to catch up to the embargoed chips. The neoliberal model of the west is now falling behind even the high tech sectors.

In case anyone is curious it takes 8 Nvidia B300 cards to run this Kima K3 frontier model yourself (2.8 trillion parameters), costs about half a million. For a big corporation or institutions even in the global south that is relatively cheap and that is the top tier option for running a frontier model. Including complete data sovereignty. Using AMD will be cheaper, and prices will go down once China catches up in like 2030. And you could run that on solar panels and batteries and regenerative water cooling. 8x 1.4kW x24 so maybe 300 kWh LiFePo4 battery so another 50k. ~~To power one complete AI server you'd need about 134 m² in solar panels which are basically free these days.~~

For solar panels you'd need about 270 / 6 * 24 = ~1000kW. With about 175€/kWp panels that is another 175 thousand plus installation. And something like 4000m². Plus chargers, and large public pool to cool those 270kW. But that is all the energy costs for 30 years up front. The bigger constraint is the land use then.

So basically less than ~~10%~~ 50% extra costs upfront to make AI data centers sustainable. This is a capitalism problem, both the hype / bubble / psychosis (b2b marketing induced) as well as the sustainability question.

[-] boonhet@sopuli.xyz 3 points 1 hour ago

Question is, how long are they going to keep making their new models open weight? GPT2 was open too, y'know. Hell, the company that made it is called "Open"AI. Money corrupts.

And how many concurrent users do you get with 8 B300s on the Kimi K3? I'm not 100% sure on the math, but apparently about <= 4 with full context size.

They can make these models open weight because basically nobody's gonna be able to run them and you can't just build a brand new model out of it like you can with regular open source software. That 8 nvidia B300 card minimum isn't enough to run it commercially, it's good for like a small 10-20 person company of fairly heavy users, or slightly more if not using heavily.

[-] patatahooligan@lemmy.world 1 points 7 minutes ago

Money corrupts.

You say it as if OpenAI started with good intentions and got corrupted by the money. But the most likely explanation is that it was a scam from the start.

[-] AlteredEgo@lemmy.ml 1 points 52 minutes ago

I couldn't find anything about concurrency. But I assumed it's more like ~1000 users? B300 has something like 100,000+ tok/s, not sure about context size. I also was wrong on my math for the solar panels, see above. It's more like +50% costs than 10%. Obviously all of this is napkin math, but even 50% isn't too bad. My point is that with a slightly saner rollout of AI there seems to be absolutely no problem to make it sustainable. And if you buy your tokens in China you can select a data center that gets it's power from nuclear or solar.

And yeah we don't know how long they will release open weight models. Once hardware to run such open weight models become available they might change tactics. But right now China releasing these open models has a serious impact on the stock price of these AI companies, which is a good thing. The sooner the AI bubble pops the better.

And like 7 of the top 10 suppliers of tokens right now is from China apparently. What Xi himself said is that the world needs a multilateral approach and AI should not become a monopoly. And while that is clearly propaganda aimed at the global south, China does have a different approach. They spend peanuts on their AI models compared to the west (and the algorithmic breakthroughs done by Chinese AI researches afaik have also been published).

But you actually can use an existing model to generate training data for a new model. Which is what the US corporations are currently accusing China of doing, which is of course hilarious. Of course China has generated huge amounts of training data themselves. So basically any industrial country could in the future fund a research effort to produce a new AI model and make it open weights. A monopoly situation would be another huge issue with AI.

Basically in regulation, sustainability and democratizing access China is doing very good. And in poking the AI bubble.

[-] Tollana1234567@lemmy.today 5 points 3 hours ago

and texas was going to add another 1800 datacenters, until they saw thier elections of the gop is jeopardy.

[-] ZILtoid1991@lemmy.world 8 points 3 hours ago

They will add them after the elections...

[-] gandalf_der_13te@feddit.org 14 points 7 hours ago* (last edited 7 hours ago)

it's very well comprehensible if you're good with numbers.

[-] TheVoiceOfRaison@thelemmy.club 15 points 11 hours ago

Usable quantum computing will annihilate the need for these data centres. They'll be defunct within a decade. The earth will be worse off and the money laundering that they provide will move elsewhere. Fuck this timeline.

[-] Psiczar@aussie.zone 18 points 10 hours ago

Sorry to burst your bubble but DC’s aren’t going anywhere. Even if we are in a position to make quantum computers commercially available in the next 10 years they’ll still need to be housed somewhere, powered, cooled and connected.

Besides, quantum computers are great at complex calculations and simulations but you wouldn’t use them to host a website or a VPS. They’ll be hideously expensive so will only be used for the specific workloads that they excel at and everything else will done with the same servers we have now.

[-] mofongo@discuss.tchncs.de 2 points 9 hours ago

AI data centers use complex calculations. Why bring up websites or VPS? these are data centers are no issue, their footprints are small and don't require much energy to work, unlike the AI data centers.

[-] beastrabban@lemmy.world 4 points 6 hours ago

Neural nets and LLMs are not doing complex math, they're doing very large amounts of simple math. Quantum computing is probably not going to accelerate our current NN tech.

This is why GPUs are excellent for AI applications, they are built around the goal of doing large amount of pixel shader computations, which is pretty simple math. You could do the math to shade a pixel by hand, for example.

[-] spicehoarder@lemmy.zip 1 points 2 hours ago

What are you talking about? Qubits are almost perfect for this sort of prediction algorithm.

[-] Psiczar@aussie.zone 3 points 8 hours ago

Fair enough, I didn’t realise they were so different from traditional DC’s.

[-] Einskjaldi@lemmy.world 1 points 8 hours ago

They're not servers. It's just gpus. It's not really a data center, basically the same as bitcoin mining.

[-] boonhet@sopuli.xyz 1 points 1 hour ago* (last edited 1 hour ago)

They're still servers, the focus is on the GPU side though. A single DGX B300 has 8 Blackwell Ultra GPUs and 2 Xeon 6776P CPUs. And networking, RAM, storage. Consumes 14 kilowatts and looks like this

You COULD host a website on one if you wanted to. You'd be stupid to do it, but you could.

[-] mofongo@discuss.tchncs.de 1 points 8 hours ago

They shouldn't be. One of the issue the sheer size and volume, they're being built based only on speculation not actual need (all the ssd/ram/gpu are sitting in warehouses waiting for these DC to be built — if they're fabricated at all). They're also being built where it's cheap, not where they would be needed (industrial and urban areas) just to save money, externalize costs, and avoid regulations.

[-] Psiczar@aussie.zone 7 points 10 hours ago* (last edited 10 hours ago)

I’ve never seen a data centre with a chimney stack. This must be an American thing.

[-] Tollana1234567@lemmy.today 1 points 2 hours ago

its the polluting gas turbine power plants that will power these AI DC.

[-] SonarTaxLaw@lemmy.world 16 points 10 hours ago

It's the "temporary" power plants that are stood up to feed the massive power demands. These are usually gas turbines that are not environmentally friendly at all.

[-] UsoSaito@feddit.uk 6 points 8 hours ago

One example would be the ai data centers Musk has that he illegally had gas generators installed at. The fine compared to not running them at all is small in comparison to keep them on. This is why it needs to be a percentage or increasing stack of cost not just a small fine every time.

[-] Tollana1234567@lemmy.today 1 points 2 hours ago

thats why he set them up in rural tennesee, actually most of the datacenters are in red states for those purpose.

[-] Psiczar@aussie.zone 2 points 10 hours ago

Well that’s one way to speed run environmental catastrophe.

[-] halezinflames@lemmus.org 32 points 15 hours ago

"AI is a tool stop being a luddite" mfs when being exposed to this reality

[-] AlteredEgo@lemmy.ml 1 points 2 hours ago

To make AI data centers sustainable with solar and batteries costs about 10% of the price of the base hardware. The AI hype is still insane but we can have a little AI.

[-] boonhet@sopuli.xyz 1 points 1 hour ago* (last edited 1 hour ago)

Man I want your supplier for batteries and panels tbh.

A single DGX B300 that could run Kimi for a few concurrent users, uses 14 kW. Costs about 500k. Assuming you live far enough from the equator to have seasons, you'll have about 8 hours of daytime at the worst points of the year (less at my latitude for an example, we get ~6 hour days in the dead of winter, with sun angles below 10 degrees and sunshine available... not every week)

Even in a slightly less extreme location, You'd be looking at roughly a 100 kW array to fill the battery and power that one server reliably like when there's clouds, about 300 kWh battery capacity to get through long nights. Easily 6 figures total cost, maybe 20-40% of the server itself. This assumes no electricity used for cooling, so it's very much back of the envelope math, reality could be much worse or a tiny bit better depending on your latitude.

[-] AlteredEgo@lemmy.ml 1 points 1 hour ago* (last edited 1 hour ago)

A single DGX B300 that could run Kimi for a few concurrent users, uses 14 kW.

Huh? 1400 TDP = 1.4 kW. This is official I believe.

And maybe my sources are off but: "As of July 2026, a single B300 GPU purchased outright runs about $53,000 (Spheron Network, July 5, 2026)" https://tech-insider.org/nvidia-blackwell-gpu-pricing/

Uh you're right, you need about 4x so 270 kWh over the day, which requires about 270 / 6 * 24 = ~1000kW. With about 175€/kWp panels that is another 175 thousand plus installation. And something like 4000m². But that is all the energy costs for 30 years up front. The bigger constraint is the land use then.

[-] AwesomeLowlander@sh.itjust.works 5 points 9 hours ago* (last edited 4 hours ago)

Looking at the actual numbers (less than 1%) other commentors are posting as opposed to clickbait headlines, I'm not seeing the issue.

[-] halezinflames@lemmus.org 5 points 8 hours ago

I'm exhausted of all this. The data centers have been so pervasive that some communities have been thinking about rationing electricity to get around the requirements. That's... a dystopia future if Ive ever seen one.

[-] boonhet@sopuli.xyz 1 points 1 hour ago

The politicians of those communities have failed their constituents.

When a data center permit is applied for, it should be weighed against the state of the grid. If there's not enough production, it should not be allowed unless they also build a bunch of electricity production capacity, which needs to come online first.

There are instances where this could actually be a net benefit for the grid. Right now in my country, power prices go negative at peak sunlight hours in the summer because we have nothing to do with all that electricity. If you produce solar, gotta disconnect from the grid at those points. A reliable consumer working 24/7 would help even things out for the consumption side of the grid math.

Not that I want them in my country, we have a very dirty grid (shale plants, yay) and live far enough north that solar only works for about 7-8 months a year. But somewhere more moderate in terms of solar availability, it would be fine. Here we need more consumption in the summer and less in the winter.

[-] AwesomeLowlander@sh.itjust.works 2 points 6 hours ago

Hyperbole is hyperbole. I refuse to fall for bullshit just because it's aligned with my views and opinions.

[-] halezinflames@lemmus.org 3 points 6 hours ago

I personally think it's bullshit that these things are happening to begin with, not that people are ringing the alarm to our new reality. The latter is just us doing our part to fix this.

[-] kromem@lemmy.world 121 points 20 hours ago

Cornell researchers found that at the current rate of AI growth, the burgeoning industry could represent 24 to 44 million metric tons of carbon dioxide emissions by 2030

The United States emitted 4.9 billion tonnes of CO₂ in 2024.

So by 2030 the AI industry CO2 release might be 0.9% of total US emissions.

That 'almost' in "Almost Incomprehensible" is doing a lot of work there.

[-] FlashMobOfOne@lemmy.world 58 points 20 hours ago

I've said it before and I'll say it again.

I love Lemmy math nerds and honestly appreciate how there's always at least one to put things in perspective for those of us who worked really fucking hard to get a C in college algebra. Cheers.

[-] HieroProtagonist@feddit.org 1 points 1 hour ago

But...but...but... the RAGE! With a measured approach to topics like that, HOW AM I SUPPOSED TO FUEL THE RAGE!?!?!?!11

load more comments (20 replies)
[-] mrgoosmoos@lemmy.ca 16 points 15 hours ago

to be fair, the scale of pretty much any global industry is incomprehensible to most people

[-] CCMan1701A@startrek.website 10 points 12 hours ago* (last edited 12 hours ago)

It's true, I can not understand how all grocery stores constantly have fruit and vegetables available.

[-] FatVegan@leminal.space 3 points 2 hours ago

That always blows my mind, still. If you ever grew cherry tomatoes, you know how long it takes and how much it yields. Even the smalltest corner super market has cherry tomatoes, every day of the year. It's not even a necessity, it's just there.

[-] leanleft@lemmy.ml 9 points 15 hours ago

i wish we could focus on making small LLMs better.
some companies are doing this. some definitely aren't

[-] beastrabban@lemmy.world 4 points 6 hours ago

I'm keeping an eye on Hailo for this purpose. I think the future is small embedded AI systems, not massive data centers.

load more comments (3 replies)
load more comments
view more: next ›
this post was submitted on 17 Aug 2026
816 points (100.0% liked)

Technology

87280 readers
2947 users here now

This is a most excellent place for technology news and articles.


Our Rules


  1. Follow the lemmy.world rules.
  2. Only tech related news or articles.
  3. Be excellent to each other!
  4. Mod approved content bots can post up to 10 articles per day.
  5. Threads asking for personal tech support may be deleted.
  6. Politics threads may be removed.
  7. No memes allowed as posts, OK to post as comments.
  8. Only approved bots from the list below, this includes using AI responses and summaries. To ask if your bot can be added please contact a mod.
  9. Check for duplicates before posting, duplicates may be removed
  10. Accounts 7 days and younger will have their posts automatically removed.

Approved Bots


founded 3 years ago
MODERATORS