Hundreds of billions being spent. When and how will they get any return?
Good question
I have long wondered what is the business case
It just looks like their competition spending is billions on AI so we must do the same to avoid being left behind which is a poor decision process
Well, technology is moving and evolving so fast now, it’s come to the point of take the risk and get in head first. If a company tries to get in later in the game, they are too far behind to catch up. It costs them more in terms of ‘perception’, client attraction and retention, etc, all leading to revenue loss.
A prime example is when UBS tried to get into internet banking back in the early 2000’s. It was way too late. Most other banks already had proper and functional on-line and internet banking. Because of late decision on UBS, a lot of clients (revenue) was ‘missed’.
Many times, it’s a case of one head-honcho bragging ‘i have this’, say at a golf game, and wanting to see the puckered face of his peer. The puckered face guy comes down with mandates the following morning, and it’s a scramble to ‘outdo’. Have seen it happen often. Also have fond memories of those ‘following Monday outdo scrambles’. … LOL
I’m sure Larry has a ‘recover in case of loss’ strategy. In the worst case, may be he will proceed further with his cloud offering. Notable mention is that he also has large involvement in ‘US Tik-Tok’. For that, he needs DCs - space, storage, etc. plugging in AI to data that huge is a trmendous value/revenue generation opportunity on it’s own. Data that can be gathered from AI mining can be sold off for sooooooo many uses generating additional business opportunities.
IT = Information Technology = Data manipulation and Data management. Information (Data) is power and $$$
He is in a rather unique place, I think …
Well xAI colossus was estimated to cost around 27B to build. And now they are renting out part of it for 27B per year, so seems pretty profitable to me!
It’s like renting out a house for 2’500 per month after paying 30k to build it.
That pushes the business case down the road.
What are the costs and revenue of the people renting xAI.
That’ll depend on how much it costs to run. With an estimated 280 MW, Colossus I consumes 200GWh a month.
Normally, the tenant pays for electricity.
Does Anthropic rent Colossus then? That’s what a tenant does.
Spacex’ S-1 speaks of “cloud services agreement” with Anthropic rather than some kind of hardware or data center rental. IMO it wouldn’t make sense to exclude the electricity, otherwise the renter risks being unable to run and use the hardware. Colossus I is estimated to consume 280MW, the supplier can’t and won’t deliver such an amount of electricity without due preparation.
With that said, and only indirectly related to the question above, on top of the direct contractual costs Anthropic states that “we recently made a commitment to cover any consumer electricity price increases caused by our data centers in the US”, addressing fears and complaints that AI’s power hunger drives up electricity prices for John Doe across the board.
The sister of my driver in Spain got two fines of 600 Euros each the first two days the new system was installed:
The village did mount 150 cameras that control the complete traffic and probably some more. An AI checks license plates, probably faces too, and sends out fines for holding a phone, speeding and parking. I think in Spain the max speed in villages is 30 and nobody did do that. Now they will have to.
It is OK for me, I never use a car since I got my nice 1000W e-bike. Wonder when they start issuing fines for those too. I live 4 km outside the village and have no business there, did not even visit once the 80 days I was in Spain.
What a nice new world. 
Now this is absolutely amazing!
We got another bunch of H200 cards and can now run GLM 5.2 (with reduced precision,not the full version because you can only link four H200 cards with some sort of bus-bar called NVLink but not eight, which would satisfy the VRAM requirements for running it at full precision…).
It’s very good but also rather slow, especially if the input is rather complex.
I do wonder how this is going to work at scale for a company where say 2000 people use it at once. 2x4 H200 certainly isn’t enough for that.
At least, I don’t have to worry about running out of free tokens with the commercial LLMs. Though the token memory (or what it’s called) is just around 200k tokens it seems. If the input and the “thinking” together takes more, it will stop and complain…
Also, I don’t have to sanitize the input too much either.
Well, until they start metering and charging at work, too. For now, it’s just a playground with “unlimited refill” 
One card is apparently 20k, so eight of them is around 160k plus the server itself probably 20k, give or take.
At least, we’re not spending billions…
THe building of AI data centres in Thailand is destroying peoples lives.
Add it to the list of things that is destroying people’s lives.
This one’s hilarious from the SVP. Traditional Swiss family in Bavarian clothing.
And where are grandma and mom? Let me guess - stuck back in the kitchen where women belong. 
Was going to happen sooner or later methinks.
How to use a little less energy when consulting bots:
https://knowablemagazine.org/content/article/food-environment/2026/four-ways-to-reduce-energy-use-with-ai-chatbots
Whats this with AI companies buying up masses of books, then destructivly scanning them to “train” the AI ?