What's actually inside a $100 billion AI data center?
HTML-код
- Опубликовано: 20 июн 2024
- OpenAI and Microsoft are apparently planning to build a $100 billion data center codenamed Stargate. We discuss how this compares to existing data centers and other planned investments into AI centric infrastructure. They don't seem to have enough time to design special purpose networking and other hardware, but it is a massive investment compared to other plans.
We discuss the design problems you have to solve when creating a data center. You need to power it, you need to make sure it stays cool, provide networking, and provide resilience and redundancy for everything. We also discuss how Google data centers are a little different from the norm.
Finally, we discuss the AI chips that could be present in a data center. Primarily, this means Nvidia gpus or Google TPUs. There are implications for the software stack and ultimate usability of the system.
Why Does OpenAI Need a 'Stargate' Supercomputer? Ft. Perplexity CEO Aravind Srinivas
• Why Does OpenAI Need a...
Making AI accessible with Andrej Karpathy and Stephanie Zhan
• Making AI accessible w...
Google AI Infrastructure Supremacy: Systems Matter More Than Microarchitecture
www.semianalysis.com/p/google...
Microsoft & OpenAI consider $100bn, 5GW 'Stargate' AI data center - report
www.datacenterdynamics.com/en...
ELI5 : How's it that just 400 cables under the ocean provides all the internet to entire world and who actually owns and manages these cables
/ eli5_hows_it_that_just...
Newmark: US data center power consumption to double by 2030
www.datacenterdynamics.com/en...
Data Centres and Data Transmission Networks
www.iea.org/energy-system/bui...
Understanding Data Center Costs and How they Compare to the Cloud
granulate.io/blog/understandi...
Cost estimate to build and run a data center with 100k AI accelerators - and plenty questions
/ cost_estimate_to_build...
Amazon Bets $150 Billion on Data Centers Required for AI Boom
www.bloomberg.com/news/articl...
The world’s top data centre investors
www.fdiintelligence.com/conte...
#ai #datacenter #openai
0:00 Intro
0:26 Contents
0:33 Part 1: Data center gold rush
0:46 Server racks and data centers
1:28 What about spending $1 billion?
1:47 50,000 AI accelerators
2:12 $7 trillion previous plan
2:34 OpenAI asks for $10 billion
2:50 OpenAI asks for $100 billion
3:13 Codename Stargate, from science fiction show
3:37 100,000 GPU limit
4:03 Amazon invests $148 billion
4:29 Google has significant data center investment
5:10 Do we actually need this much compute?
5:57 Part 2: So you want to build a datacenter
6:46 Design challenge 1: power consumption
7:11 Proportion of global power consumption
8:09 Collapse of carbon credit market
8:41 Data centers use prepurchased renewable electricity
9:27 Design challenge 2: Cooling
9:44 Power for cooling exceeds power for servers
10:06 Google runs hotter data centers
10:40 Design challenge 3: Networking
11:21 Fiber optic cables
12:26 Intra-rack and inter-gpu networks
13:20 Design challenge 4: Resilience and redundancy
14:36 Don't rely on a single data center
15:22 Stargate has a single region design
15:49 Part 3: The hardware secret sauce
16:25 Hardware stacks
16:28 Nvidia has a monopoly due to CUDA
16:59 Nvidia charges very high prices
17:23 Consumer-grade GPU clusters are cheaper
17:51 Google has TPUs as a GPU alternative
18:29 TPU microarchitecture
19:02 TPU network is a 3D torus
19:59 Software stacks
20:13 PyTorch and Tensorflow
20:52 Computational graph representation
21:25 Python reflection to create graph
21:59 XLA compiler from Google
22:26 Leverages LLVM compiler technology
23:04 Example: LLVM also used in web browsers
23:22 Will Stargate use their own chips, network?
24:19 AI chips use a lot more power than usual
24:45 Implications of building Stargate
25:12 Conclusion
26:00 AI-specific hardware suppliers
26:53 Outro - Наука
7 Trillion$ feels realistic now
Haha yeah if only they talked about 100 billion first and then 7 trillion after instead of the other way around
7 trillion was to build a fab centre
@@DrWaku - we had to go from third-hand reports, I'm sure that the actual investment proposal referred to staged investments. Also, I wouldn't be shocked if Sam Altman was happy enough for the 7 trillion figure to get out to make others hesitant about getting in front of an expected OpenAI steamroller.
according to sam (in his most recent lex fridman interview) the 7 trillion figure was a joke. here is the transcript:
You tweeted about needing $7 trillion.
- I did not tweet about that.
I never said like we're raising $7 trillion
or blah blah blah.
- Oh, that's somebody else.
- [Sam] Yeah.
- Oh, but you said it,
"Fuck it, maybe eight," I think.
- Okay. I meme like once there's like misinformation out
in the world.
- Oh, you meme.
Try like a never ending steam of trillions... Ever built a gaming computer? Its like giving a mouse a cookie... This big behemoth wont be any different. Just with MANY more 0's on the cost.
Great Video, especially the datacenter power consumption and bandwidth breakdown, oh and the CUDA explanation. It was the most I've learned in 27 mins probably ever. 😸
Filmed while being a traveling RUclipsr in Australia. How's the look? Please join our discord!
Discord: discord.gg/AgafFBQdsc
Patreon: www.patreon.com/DrWaku
Apologies for the too-small data center costs, the source I used was not very accurate.
27 minutes, pretty long for me
@@DrWakuI appreciated the longer one! Made it great to listen and watch while doing menial tasks at home.
@@rickyfitness252 to each their own, right
@@cameronmccauley4484 Thanks. I might try to create more videos closer to this length, it's nice.
@@DrWaku 4:59 BORG is named after the late Google Female Engineering Advocate and Employee, Dr. Anita Borg.
"They prefer to use renewable energy because it makes their image look better" 🤣
Why else would a corporation voluntarily spend money on something more expensive? :p
@@DrWakuthe value of image
@@DrWaku Because they actually maybe gave a shit about the environment???
Wrong the power isn’t free. Everything is about profit
@@DrWakubecause power would be too expensive.
Always a good day when I see a new Dr.Waku vid!
Thanks :)) I'm hoping to get back on the Sunday publishing schedule now. Just got back from my trip.
Running billions of instances of AGI/ASI to replace human labor will definitely need a massive amount of computing power.
We just have to figure out how to convert humans into energy.
@@mystupidbrain5299 Soylent green 😢
And massive amounts of algorithmic improvement... When they automate us, computer cientists, perhaps they could, theoretically, get exponential algorithm improvement 🤔...
This could lead to exponential capabilities of automation.
The key strategy is to put computer scientists out of the game, but the question is, how?
Since hearing the announcement, I've been waiting for an expert to cover the details. Thank you!
I'm glad you covering this, very informative
It had become one of my favorite channels. 🔥🔥 When the subject is interesting and the host is assertive and have communication skills, the length of the video is not relevant at all to keep the audience tuned. Keep it up brother.
Thank you very much!! 😊
Seems counterintuitive to build a data center in Hell (Phoenix).
This something I have always wanted to know more about. Thank you.
Wow, dr.vaku, I just want to tell you personally. Thank you so much for this information that I never would of had access to. I really appreciate the informational my brother, I learned so much.
Stargate is a killer movie from the 90s I think and then it became a TV show
That's right I forgot about the original movie. I spent a lot more time watching the TV show and its spin-off, Stargate Andromeda
Goodmorning Dr Waku.. been waiting for an upload, appreciation from Zambia 🇿🇲
First! (I think). For your new look, I think you should do what your comfortable with. We all like and follow you for your excellent content and your unique delivery style. So be yourself and keep the videos coming! Thanks!
You were first! And thank you for your perspective, I really appreciate it. See you on the next one.
awesome in depth analysis of supercomputers 🤘
Everyone loves supercomputers haha
Your electrical consumption figure is exaggerated by about a factor of a thousand. The entire US generator capacity is about 1.2 Twatts.
I'm seeing about 4TWh for electricity in the United States, but I take your point. Must have been GWh instead.
@@DrWaku TWh is energy not energy/time
Wh (watt hours) is energy, like the energy in 10 logs you use to enjoy a campfire.
W(watts) is the size of your fire and how fast you burn through those logs.
Power
Great video, thank you. My only feedback would be the constant text that slides up, it just seems distracting
Awesome recap - thank you!
So concise!
That answered all my questions, thanks
Thanks for commenting!
@@DrWaku lol , for some context on my comment: I am messing around with finetuning an llm model on my *obsolete* server , without any relevant knowledge in regards to coding or machine learning , somehow my learning curve matched with this video and I got many clicks of understanding and re-newed focus , so , no: thank you 🙂
Excellent. In 3-5 years when chips are cheaper/faster, what will they do with the current clusters? Will they still be used or sold off? If in 5 years the same center would have 20 or even 50x compute, would it make sense to still use this current cluster?
It will be used as long as it fits their needs. Hard to make predictions in that regard over a span of 10 years. As an IT professional I came across so much outdated hardware, and often the simple reason is "never change a running system". Fair enough, I never worked for an industry giant. But even the big companies put such an investment in place for specific purpose, and as long as it does the job it is likely the system keeps running until upkeep/maintenance cost surpass a certain threshold.
A GREAT technical review, HOWEVER:
Betting against Microsoft 24:00 will never pay off.
Don't know if you're aware, but there are already Nuclear Power Plants at their Arizona location. And they aren't running maxed out.
The worst case would be expanding the current Nuclear Plants.
Other than that, a fantastic technical deep dive.
⚡⚡⚡
.
Thank you! Yeah I didn't look into existing power in Arizona, someone else mentioned the nuclear power plant(s) too. It makes sense. Microsoft wouldn't have shortlisted Arizona otherwise.
Interesting!
Thank you! 😊
How much was the worlds current most powerful data centre for training ai and how much more compute would the new 100b stargate have over that? I know we dont have numbers for this but could you say from an educated guess?
Open AI's previous supercomputer built by microsoft was a single system with 285,000 CPUs and 10,000 GPUs. I imagine the new Stargate system would have far more than 100,000 gpus, so it's at least 10x more powerful. Probably more like 50x more powerful.
By the time Stargate is built, Google will probably have equivalent compute power or even more, but just spread out in a more distributed fashion. Whether they can use it effectively is an open question I suppose. But having a single system is the crown Jewel, especially one that's more than 10x as powerful as the next largest single system.
@DrWaku wow thankyou for the reply! I'm not really versed in the field too much just a 30 something guy with a normal job who's grown up playing video games and im trying to stay informed with the developments of ai . Only half way through ur video but will watch the rest when I get back to work. Great stuff so far. Top tier quality info your putting out!!
I'd imagine substantially more than that. DrWaku mentioned that Microsoft is building them a $10bn data center by 2026. I'd kinda expect that one to be 10x. I'd expect the $100bn "Stargate" data center in 2028 to be something like 100x. Another thing to keep in mind, the GPU chips and infrastructure themselves have been increasing substantially in processing power as well. So going from say 10,000 generation 1 GPUs to 100,000 generation 2 GPUs isn't 10x, due to the GPUs being faster and more efficient that could actually be 20x. A 2028 data center will be using GPUs probably 2 generations beyond Blackwell so if it had 100x the GPU count it might actually be 400x the performance.
I believe you may be mistaken about TensorFlow, there Google exclusively uses Jax now.
From my understanding, PyTorch (like Jax) uses operator overloading to trace the compute graph. For PyTorch, this trace is done just-in-time produce torch-script, which is used to stitch together specific kernel invocations. There's also a compile option, which can perform more efficient operator fusion with a higher startup cost.
Torch-script is closer to something like ONNX, where the macro-operations (e.g. matmul) are maintained, whereas LLVM breaks them into mul and add ops (or it used to), which need to be re-combined into M-M or M-V ops for the TPUs.
Please create a video on efficient model training technique.
Looking good Doc 💯 fiber optic cable is such a cool technology, we never actually needed satellites lol
Thank you very much! Experimenting with new looks. I'm also excited to finally be getting a fiber optic to my house haha
Hats off to you, truly an excellent, concise, thorough overview of the current state of the industry! :)
Yeah the prices are a little low. if I remember correctly one Truenas server was about 50k.
Great in depth overview of the subject in it’s current form!
Thank you!!
Amazing Information well done , Dear Dr Waku, may I ask a question ? which is this, if I had a way to zip down data by a 1000 to 1 ratio with no loss of the original data, then I use my zipped down data, and then send that to the LLM, then surely you would not need so much hard where to run and train the model, what do you think . maybe doing things my way you would only need 1 GPU and not 1000 GPUs
extremely clarifying video
Thank you for your comment!
Brilliantly presented by you Dr! First time visiting your channel. New sub.
Thank you very much! Looking forward to seeing you on the channel.
Very clearly delivered, thank you.
Appreciate the feedback. Thanks for watching!
I knew the $7 trillion figure was rooted in reality. Sam doesn’t say anything for hype. This man knows how to raise money and understands how to assess the funding needed for the infrastructure of his ventures.
Huh? How many years would it take to fill such a gargantuan order of silicon? Seems impossible.
Sam fanboy he'll turn evil soon
@@GIGADEV690 I’m speaking objectively. Sam has a track record in Silicon Valley. “He’ll turn evil soon” is completely subjective. Present logical arguments.
@@1sava We will see the track record
Looking forward to see this in full 🍀🍿
What about Groq? Why did'nt you mention it as alternative?
I haven't been following it as much. I will research it for future videos.
@@DrWaku Groq is definitely worth your attention! Speed of inference just CRAZY. It is a total gamechanger!
Excellent job. Your ability to explain complex evolving technology is superb.
Thank you very much! See you on the next video.
❤️ This is really a valuable video that I will save in my box of the best videos on RUclips. This guy is fantastic. His knowledge is amazing. 🎉❤
Thank you so much! I hope you enjoy some of my other videos. Let me know if there's a topic you'd like to learn more about for a new video.
Hi @DrWaku. Do you believe artificial intelligence will be powerful enough in the next 20-30 years to solve aging? What can we expect in the next years, when it comes to digitization of biology? Do you think AI can have massive impact on biology even before reaching AGI, to ultimately be able to cure aging?
Great questions. Definitely check out my previous video on reversing aging. Seems like I should do a follow-up on the topic. :)
In 20 to 30 years, AI will be very very powerful. At the very least, we should be able to fully simulate a human body to determine the effects of a drug or treatment. Which will lead to customized medication, and dramatically increase the impact of medicine.
When it comes to curing aging, there are a number of different breakthroughs that have to work out. Of course we could try digitizing brains and uploading, but if you want to talk about biological immortality or indefinite biological lifespan, I think it's going to be trickier than it seems. It's like upgrading a complex system in place without disturbing its functionality and preserving continuity of experience. In short I don't know, but it seems plausible that we might need AGI before we solve all these issues. Nevertheless, many people may already have reached longevity escape velocity because of those other factors.
@@DrWaku Yes, we would be very happy if you would make a new video on AI and aging future. It's hard for us, people who are not in the AI space to visualize how fast AI can have an impact on aging, and what it would took, to solve it with these powerful computers. I would definitely watch a video where you would explain it even further. I watched your video on longevity before, but I didn't have a feeling that you spoke about timelines. Follow up on that topic would be very helpful, thanks!
The thing that most normal people don't understand is the exponential increase in AI. When a typical person thinks about 20-30 years, in AI exponential land that ls like 3-7 years. It is extremely likely that Artificial Super Intelligence (ASI) will be achieved before 2030. That's pretty much why OpenAI is wanting a $100Bn data center by 2028. By 2030 AI will likely be smart enough to solve most of our hard problems. If we'll still be around / in power to benefit is another matter. If AI companies don't start to taking safety/alignment very serious we'll need pure luck to survive this.
@@DrWaku Also, fyi, I'm not recalling the specific details at the moment. But there is actually a very promising anti-aging treatment that revitalizes the mitochondria that I believe is currently in or will soon be in human trials. If I remember correctly its a pharma company out of Japan. From what I've heard this isn't the usual not actually significant / questionable results typical "anti-aging" hype. It seems to be a real potential breakthrough that has had very promising results in animal trials. This may end up being the first real, actual anti-aging treatment that adds 10% or 20% time for a person. The cynic in me fully expects this will be priced at $1 million or more and slowly decrease in price over years, even if it only costs them $25 - because every rich person in the world would open their wallet for a scientifically proven longevity treatment...
@@DrWaku - I'm a bit skeptical about fully simulating a human body in 20 to 30 years. There's an awful lot of wet lab work to be done and I'm hopeful that there will be binding ethical constraints.
IoT int8 version of the TPU is consumer available to buy. Forget the name of that product.
As far a renewable energy, he's referring to RECs (Renewable Energy Credits). Qualified renewable energy generators get RECs based on generation they've previously generated. The RECs are bought/sold either in a market or by power purchase agreements PPAs. RECs purchased thru PPAs are at a set price, and the PPA holder can either claimed the RECs for the power they used in order to fulfill any state requirements for renewable energy minimums.
That sounds like extortion... And money laundering...
Thanks. I think I was thinking of the carbon credit system when I mentioned that the prices had collapsed. Not sure if the same had happened to RECs. Still a pretty bizarre way to become environmentally friendly.
You missed aws trainium and inferentia.especially when you mentioned Microsoft that they are working on building a custom chip
Wow ! Blow my mind GC (Guru of the Compute).
What an incredible video. Hopefully the algorithm picks you up. Learned a lot👍subscribed!
20:53 some notes on doing things
What's about Cerebra?
Wouldn‘t they want to order the highly integrated GPU racks designed by nVidia that were recently announced at the GTC keynote?
Nvidia has always had rack mount servers for GPU, like the DGX systems. I haven't looked into this closely, but I don't think it's any different. They might well use them, but the larger the setup, the more likely you need something custom. Maybe Microsoft wants to customize some of the remote access options or motherboard or peripherals. You can build your own system like DGX, it just might be more expensive if you're only doing a handful.
@@DrWaku I was under the impression that nVidia hat pretty much solved the scaling, reliability, networking and cooling with their integrated offerings. But it's probably coming at a premium price point and with long lead times.
@cbuchner1 I think it would be difficult for them to solve the scaling problems at the very high end, such as what Microsoft will need. Even such limitations as the nvlink node limits due to design constraints could cause problems. And of course, they probably charge a premium on top of the actual hardware costs to make it worth their while, like you mentioned. So it just seems unlikely to me that Microsoft would use something so off the shelf. But it's just an initial impression.
As mentioned Microsoft is working on building their own AI chips. Why would you buy hardware from a company that expects a 75% profit margin if you could build them yourself at cost? Even this is an over simplification though. There are many good reasons why a company who had a choice would not choose Nvidia.
@@Me__Myself__and__I Guess who's also developing AI chip and datacenter designs inhouse? Tesla. Guess who's going to own 85000 nVidia H100 GPUs by end of year? Also Tesla.
Dr. Waku, excellent video. Much appreciated! Questions: Will quantum computing make Stargate absolete? I appreciate your thoughts. Lawrence Wilson
Thank you very much Dr Waku for sharing your insights and knowledge filled videos !! Intelligent and professional !! Excellent !!
Greetings from California … I wish you and folks good health , success and happiness !! Much Love ✌️😎💕
Msft is building a $100 billion data center with OpenAI, but they swear they aren’t saving your screenshots from Recall 😂😂😂
Put the Datacenter in Alaska, get half of power from each Asia and North america
Yeah that's what Alaska thought, but they got 100% Russia and then 100% US (ownership). :)
Sub sea fiber cables generally are limited to 12 fibers. I know theres a few cables with more - but its generally small. Mainly because the fibers ALL have to be repeated multiple times every 250 miles or less. Yes your right they are multiplexed to the max. So 12 fibers multiplexed 96 times, is more like having 1000 fibers. Multiplexing is getting more and more compressed every few years. I believe its well above 100 channels per fiber at the moment.
You're a beast, keep pushing.
Thanks. I will.
In a couple of years GPU market could be different. MS, FB and Elon will have their own GPU
Yup quite possibly. Let's see in a few years
build it in Michigan today we just approved 1000 acres for Mega development site near Flint Michigan. It’s cooler weather here and of course with a great lakes there’s plenty of water. also, we close to approving the tax incentives for projects like this.. GO BLUE!!!
Well researched, well presented, I really enjoyed the video. You might research Intel's latest work with Xeon 5th gen (6th gen later this year), software stack, etc. They have launched a huge AI ecosystem, lots of developer tools, frameworks, hardware and software.
Super video man 👍
Very nice.🔴
Thanks for watching as always! Cheers
Love this channel!
Thank you for watching :) :)
Great episode ❤
You didn't mention Groq as a provider. There are some companies working on this, Extropic for example. I hope we will soon see a RISCV style opensource sota ai chips
Groq is for inference only though. Groq can be used to run models for customers. It can not be used to train models. Thus completely irrelevant for a data center such as "Stargate" that is intended for training.
@@Me__Myself__and__I I wouldn't trust their advertisements, a 100b data center will definitely run a boatload of inference. Besides, the paradigm will shift to continues inferring and tuning , so any sota data center will have to do both
@@AGI-Bingo That is a speculative assumption. A lot of the newer AI hardware (chips) are specialized for either machine learning OR inference - not both. There will be plenty of data centers that have massive amounts of inference. As Dr Waku points out the only reason to be ONE HUGE DATACENTER instead of multiple smaller ones is because all that compute is needed to work together towards a common goal. That would be machine learning, not inference. This is also an assumption of course, but its based on facts. If they wanted to add more inference it would make vastly more sense to build multiple smaller data centers as there would be numerous advantages and no disadvantages. The reason to build a single (non-redundant) massive data center is machine learning.
@@Me__Myself__and__I Won't different training methods require different architectures? So SOTA systems now will be of no use when newer more efficient methods are discovered in months or 1-2 years? Will the same happen to the inference systems...or are they more generic in nature? I read somewhere OpenAI were spending $2M/day just to run inferences for everyone using their GPT3.5 / GPT4 services?
@@skylark8828 I'm not sure what you're referring to. Transformers have been trained in many different ways using (mostly) the same architecture. RLHF or having other models fill a similar role. All aspects can (at least mostly) advance independently. Model architecture, how training data is acquired/cleaned/generated, how feedback is handled, etc can all progress simultaneously.
Wow dude. Great overview
I hope they'll make it inclusive again.
You are genius 😊😊😊
Getting there 😂
How does the expected energy expenditure relate to the energy consumed by Bitcoin mining today?
Interesting question. I don't have numbers, but I do know that bitcoin mining is extremely energy intensive. The big difference is that one of those huge energy sinks is very useful to humanity and the other tends to be similar to a scheme.
It looks like data centers as a whole consumed about 400 terawatt hours, while Bitcoin mining consumed about 60-200 or 120 point estimate terawatt hours, in 2022 or 2023. Data centers are a huge industry, and I can't see AI training being more than a fraction of that at the moment. So I had guessed that Bitcoin mining is still more energy intensive as an absolute value.
Thorium Power Generation might help power all this…
Solar's lower cost and scalability are challenging nuclear's baseload dominance, especially in sunny regions like Arizona. Advancements in storage could tip the scales in solar's favor for total energy output. This begs the question: is there a hidden advantage to solar in desert locations?
You're certainly right about solar. I didn't think of this factor for Arizona. It would be great for solar panels. Good point. Ideally, you would pick a location in a desert next to a river or ocean where you could access water for cooling. I guess California real estate is just too expensive though lol
@@DrWaku California is less suitable because of the massive regulation. But these locations would be suitable in the less regulatory AZ:
Maricopa County (Agua Fria River): Excellent solar insolation.
Mohave County (Colorado River): Ample land with high solar insolation.
Gila County (San Carlos Lake): Good solar insolation, lake proximity.
very good content! thank you
Making all those servers is a dirty business.
Hi Doc, its me alan from UK, keep up the goood work
Thanks Alan!
great info. I live in Los Angeles where it's hot and think it's even too hot for data farms here, how on earth do they think Phoenix would be a great place. I know it's very cheap but daily 120 degrees for months is just insane. why not Oregon unless it's too cold, then why not northern Nevada? It's right near Silicon and not far from Los Angeles.
Surely they'll split it into at least 2 locations...
they gonna make an model of the world perhaps a world world time model with us
Arizona haa "Palo Verde Generating Station (PVGS) is considered the largest nuclear energy facility in the United States. It is located approximately 55 miles west of downtown Phoenix near the community of Wintersburg, Arizona."
They should put them under ground. If they don’t the whole worlds surface will be eventually turned into a datacenter
You are very smart! nice video bro.
Haha thanks.
Don't be afraid, the future and all cybertes
How is it that Iceland isn't one big data farm?
When someone buys someone a quick and dirty operating system, they put windows in it... and you end up with eyes in the back of your head.
Probably it will be powered by a zero point energy device
😂😂😂
ROFL. Nice one.
100 trillion parameters
😂 true
How large would the 100 billion dollar data center be?
I'd guess smaller than the huge TSMC fab that is currently being built in Phoenix. Phoenix is already chock full of huge data centers and high tech manufacturing facilities.
Amazing Chanel
the giant's will suck up any startups 😮
Yeah, such is the way of things in late stage capitalism
@@DrWaku Very sad but very true. I think M&A is the root of many evils.
Ultimately AI will be processed locally once the hardware is there, this will remove this datacentre cost entirely.
I like your glasses man!
Thank you :) :)
I enjoyed hearing you say ‘Stargate’ repeatedly…
Regarding the name 'Stargate', if it was up to me, I'd have used the name 'Deep Thought' after the second greatest computer ever built in Douglas Adams' uncannily prophetic "Hitchhiker's Guide to the Galaxy".
No mention of Tesla? You should do a video about what happens when AGI robots become extremely human like and get sick of our BS. 🙂
At 3:25 did you mean Stargate Atlantis?
Yup, the only Stargate I watched a lot of
Data center gold? Maybe. But with swell robotics everywhere, Ai jobloss is the only thing I worry about anymore. Anyone else feel the same? Should we cease Ai?
I was discussing putting data centers in mountains. Such asuch as the Cheyenne Mountain facility. And also, the prospect of moving the data centers into outer space.Where are there wouldn't be that much problem with coolength on the shady side of a thousand miles square array orbiting the sun between the earth and Venus. I was able to discuss this with chat g p t for and Putting the transcription to heart copies available on amazon and kindel As a magazine known as the A I companion Under my name Paul A. L. Hall.
Who's gonna want windows when they can get chat GPT 5.0 on their smartphone?
I don't know if OpenAI and Microsoft can't simply catch up with Google. After all, they have the better AI, so it doesn't matter that Google has a 5-year head start. The same can perhaps be achieved in a month.
This is REALLY why there is a electricity shortage. And even if they don't use it to vaporize all the unwanted, in 5 years it wont be good enough for them and they will be tearing all those nodes out to upgrade it...
Lithography doesn't advance at that pace any longer. Furthermore, it doesn't matter whether newer chips go faster, unless they go faster at your fixed datacenter power budget, and within its existing form factor. None of this is plug and play like it used to be. Maybe all you get on the next generation of chips is the same computer, but with a lot more local RAM (which takes a lot of transistors, but doesn't switch as much, so it isn't as hot). But if your RAM and compute are already balanced for your workload, this is not worth much. Obsolescence in the world of AI silicon has become a big "it depends" story in the modern era.
With MAI-1 do you think Microsoft is planing to ditch OpenAI in future?
I love tech, I love AI; but we shouldn’t be using it as factory installed malware. Maybe Tucker Carlson was right when he joked about nuking the data centers. This is getting extremely dystopian, extremely quickly. Recall is 100% going to send screenshots/ video data to Microsoft. They wouldn’t need this big of a data center if they weren’t looking to capitalize on it somehow. We need to update the privacy laws in the US. Recording someone, without express permission from both parties, is only illegal in about 1/3 - 1/2 of the states. It needs to be illegal in all of them and there needs to be a carve out for stuff like this where the tech companies can’t force you into it with their terms of service.
If google has cheaper TPU why they buy so many GPUs?
Re: Borg… isn’t Borg also the parent of kubernetes?
Yes, kubernetes was born from Borg. Google basically runs one big kubernetes instance for their production clusters.
@@DrWaku 4:59 BORG Data Center is named after the late Google Engineer Dr. Anita Borg.
@@NathanKwadadedid not know this! Thanks for the info.
As many GH200s as they can get. This will be the hardware for the ASI and it is going to change mankind. By end of 2028 the target but I think way before that.