Amazon can't build AI capacity fast enough, throws another $200B at the problem

AWS has an open cash spigot for AI infrastructure, with Amazon CEO Andy Jassy telling investors the company has been monetizing compute capacity as fast as it brings it online and it plans to double capacity by the end of 2027. "We’re growing at really an unprecedented rate. Yet, I think every provider would tell you, …

[Michael Hoffmann]

AMZN

Wall Street, at least, was not happy...

[MachDiamond]

Re: AMZN

"Wall Street, at least, was not happy..."

Well, if the company brings in $80bn a year in net income and they announce plans to spend $200bn, financial people somehow, despite an MBA, will spot the problem.

Will that capex pay for itself?

I can buy a $14,000 camera (if I had the cash) but I can't charge clients any more for jobs and they'd likely not spot any differences either. I was looking at a new Rigol O-scope the other day (may have been some drooling), but I have a perfectly serviceable older Tek scope that does the job. While I'd like to have the Rigol, it's a business decision that doesn't add up. Customers don't care and what tools I use doesn't matter as long as I deliver what they need. They will spend much more time grinding me down on price.

[MonkeyJuice]

Re: AMZN

The difference is really that at least in some cases (I'm not much of a camera guy, but know several professional photographers)- upgrading your kit can improve _your_ quality of life, whereas attempting to get AI to do anything other than source port something I could do in an hour (in two hours) just lowers it. I don't understand why 'paying more' helps people as they claim- it just reduces your MTTF. It is bad enough wrangling LLMs at the default rate they produce crap, wading through even more effluent per hour neither improves the quality of your output, but sure gives you a lot more time to just find yourself begging them to please, PLEASE just edit the function, and don't delete everything else in the file.

Hey I know you made a change and then the test broke, but did you just 'fix' the test rather than the thing you just broke? AGAIN?

We've all gotten out of dysfunctional work relationships in the past, there is he need to pay to enter a new one.

[MachDiamond]

Re: AMZN

"upgrading your kit can improve _your_ quality of life"

The last camera body I bought for work did improve my quality of life. It had a couple of useful features that sped up my workflow. That camera is 3-5 gens back (Canon 5DmkIII). Lots of bang for the buck. There was no benefit to getting the mkIV or the 5DS or moving into mirror-less and needing to start buying more lenses. The newer features are irrelevant for what I do.

I've stepped back in time for my personal camera and have an Omega-View 45D film camera that takes forever to make a photo. The prints are absolutely gorgeous with such a big sheet of film (4x5 inches). I'll stay away from 8x10 and leave that for the real geeks. Still on the fence for 6:17. I've got a long way to go with film before I should buy anything else. I could see if my Kodak Brownie and Hawkeye work as well as being cool decorative pieces.

[Charlie Clark]

Re: AMZN

As with Google, I think there is a certain amount of beggar thy neighbour in this because both companies can probably cover most of these costs from cashflow, especially because they are using so many of their own chips. Compare that with OpenAI's burn rate or Oracle's bond yields and you can see there an inversion of network effects producing last man standing effects: those companies with the strongest balance sheets will survive.

[Porco Rosso]

Re: AMZN

The real elephant in the room is Power. The U.S. grid is already strained and nowhere near ready to deliver the multi‑gigawatt demand coming from massive AI data‑center expansions by Amazon, Google, Microsoft, Meta, and Oracle.

[DS999]

There is a flawed premise with AI

Companies spending like mad are operating under the assumption that it will be like PCs or smartphones, where there are at most two winners. So they are spending everything they can because they want to be one of those two winners.

But those markets ended up with only two dominant players due to network effects. Despite stuff like Java we still operate in a market where software is written to run on a particular operating system. Yes it can be ported to others but that's not free, nor is the support. So it is really really hard for a new PC or smartphone operating system to break in to the market and establish anything beyond a niche (like Chromebooks in schools)

I don't see much in the way of network effects for AI. I can use OpenAI and if it doesn't give me what I'm looking for I can try Claude. If I can't find an app I'm looking for on my iPhone, if that app exists on Android that doesn't help me at all. Now sure there will be software layers added to specific ones (like Claude Code) but there's nothing stopping you from having Claude Code write you some software, then later have ChatGPT modify it. The output of AI is not tied to a particular API, it is stuff like English words, C code, or whatever. It is fungible, so the only reason one or two AIs might "win" is because they are better (like Google Search was 20 years ago) but even then you will probably end up using Copilot on Windows because Microsoft is pushing it, you'll use Siri on iPhone because that's built in, etc. You might go looking for a "better" AI for certain tasks. A corporation might sign a deal with a particular AI but they aren't likely to stop employees from using other ones that fulfill certain roles better.

I think all this massive investment to "win" is based on a fundamentally flawed notion, that AI will be like smartphones and there is only room for one iPhone and one Android and everyone else will be the Windows Phone or Blackberry of AI.

[MachDiamond]

Re: There is a flawed premise with AI

"but there's nothing stopping you from having Claude Code write you some software, then later have ChatGPT modify it. "

Some of the code I've seen come out of AI is very opaque so I'm not sure if the argument would hold up that it could be modified at all, in some cases, without starting over. The next AI used would have to analyze what the software is doing to make a model and then take into account the modifications you want and build something new from scratch. Human software can be less efficient, but be far more understandable to humans down the road.

[Winkypop]

If you build it

They might come.

Probably not.

[Pickle Rick]

*Pop* Go The Weasels

>> Andy Jassy: We’re growing at really an unprecedented negative rate.

FTFY. The same goes for nVidia, M$, AMD and most of the LLM Cabal.

Yet all the spiel is "We're doing great, and we could do even better!" All this "move fast and break things": build fast; change stock market rules; ignore laws; burn energy; fill the sky with junk...

No one's buying it, literally or figuratively. The markets don't want it. The punters don't want it. For the few that do, you don't need to spav money on this scale. Prove the tech first FFS. I don't want a massive crash, everyone suffers. Hopefully the markets will continue to plummet as an indicator to these morons that this shit will end them if they don't stop. Don't take the rest of us with you.

You're not geniuses. You're. Just. Humans. And dirty ones at that.

/rant

[Pascal Monett]

Re: *Pop* Go The Weasels

Oh stop talking sense. We're way past that.

[Pickle Rick]

Re: *Pop* Go The Weasels

Thanks Pascal. Your words, and this coffee, have helped. Phew! That was a close one.

[Anonymous Coward]

Re: *Pop* Go The Weasels

Sadly the users buy it. We did bullshit carbon training at work. How we can do better for the environment, how we, ourselves can help. I said ban all LLMs. I got booed because everyone loves CoPilot.

So we're all fucked.

[Doctor Syntax]

Re: *Pop* Go The Weasels

You should have reminded them that if it was really that good some of them would lose their jobs due to the improved productivity and as manglement would look at it as part of employment expense the survivors would be paying for it by reduced salaries.

[Craig 2]

Re: *Pop* Go The Weasels

The thing is they are growing because they're selling AI to their customers who are then putting it on the front end of their e-commerce etc sites where it drivels over everything.

It's gonna take time for it to filter through that the end-users (ie. paying customers) don't like it and try to avoid using it. Then it's all going to be cancelled because it's a massive waste of money.

Hopefully (for them) they're not all in Oracle-style contracts that even Houdini would have a hard time getting out of...

[Scene it all]

Re: *Pop* Go The Weasels

AWS will continue to make money until their customers realize that THEY are not making any money from this and cancel their accounts, leaving AWS with a whole lot of idle hardware.

[Ken Hagan]

Re: *Pop* Go The Weasels

Exactly. This sounds like Amazon are just a little downstream of nVidia. They are spending squillions but as long as there is someone else downstream of them buying capacity, Amazon aren't (necessarily*) the mugs here.

(* Of course, these huge companies have divisions and it is possible, even likely, that one of the mugs is a different division of Amazon.)

[Anonymous Coward]

Re: *Pop* Go The Weasels

I hate generative AI so much that I use an enterprise policies.json added to my /user/lib/firefox/distribution directory:

{

"policies": {

"GenerativeAI": {

"Enabled": false,

"Chatbot": false,

"LinkPreviews": false,

"TabGroups": false,

"Locked": true

}

}

}

[Pascal Monett]

So, another $200 billion

Plus Alphabet's $180 billion.

Does anyone really think that they're really going to be able to avoid this deluge of pseudo-AI (at least, on Windows) ?

Thank $Deity for Firefox and NoScript+uBlock Origin . . .

[Anonymous Coward]

This is essentially the IT example of a drained sweaty man in a casino at 3am betting increasing sums on roulette as he’ll lose the house if he doesn’t eventually win.

[alain williams]

This the day after Bezos guts Washington Post

Bezos owns the Washington post and cut approximately a third of employees. I suppose that good journalism costs and that they can save much money by pumping out Trump supporting fake news - which will result in Trump awarding AWS many lucrative contracts.

[MonkeyJuice]

Re: This the day after Bezos guts Washington Post

And nice contracts to Theil to do transparent 'endpoint kernel monitoring' on your EC2 nodes. For 'security'. Gotta keep Satan off your VPC.

[Scene it all]

Re: This the day after Bezos guts Washington Post

Wapo can use AI on AWS to write copy for them. Readers will spot the difference and think "why am I paying for this slop?", cancel their subscriptions, and the dominoes start to fall.

[alain williams]

Re: This the day after Bezos guts Washington Post

It seems that Bezos is being well rewarded: Amazon ran a $1.2 billion tax bill last year, down from $9 billion the previous year, and even as its profits jumped by 45 percent to nearly $90 billion - a tax rate of 1.3% -- I wish that I could get that !

[PCScreenOnly]

Aws, F1, overtake

Will this improve the aws ai predictor in F1

Always made me laugh at how bad it is

[Doctor Syntax]

If by monetising it, he means placing utterly irrelevant sponsored ads in Amazon product searches then the publishers paying for sponsorship are getting a raw deal. The other day, out of 22 results in the first page after following an author link 6, including the first 2, were sponsored links to different authors and one was a bad search result with 3 authors, the first having the same forename and the 3rd the same surname but with an added 's'. All errors were fiction and the author was a prolific non-fiction writer.

(Fair enough, I wasn't looking to buy, just get the correct bibliographical details and I eventually realised it was a journal article and not a book, but even so, 1/3 errors in the first page.

[elsergiovolador]

Tax

It's a fault of idle tax inspectors. If they combed their books, we probably would not have that AI spending craze, but functioning services.

All that reminds me of VAT carousel, just on steroids.

[Sam not the Viking]

Is the return worth it?

Adding 3.9 GW in energy.... About $5 billion per year extra costs then.

If it could be expressed as an increase in overall efficiency, thus resulting in a nett energy reduction, this might be impressive. But to produce more 'monetizing'.... For AI?

The world is going mad. As it's Friday, I suggest this solution ---->

[dippy1]

Re: Is the return worth it?

So how do we equate energy consumption to processing power?

Am I getting too old in my thought that is it processing capability that should be the important metric not so much energy consumption?

or is energy consumption the new metric for measuring datacentre "performance"?

I guess we are at the point of "My datacentre uses more energy than your datacentre hence mine is bigger and better".

[MonkeyJuice]

Re: Is the return worth it?

Number must go up. In this new post-rtfm world, bragging that just one of your datacentres consumes more energy than the entire steel industry of several countries is a bragging right, not a line item that needs to be burned to the ground.

[JacobZ]

Re: Is the return worth it?

"So how do we equate energy consumption to processing power?"

<Sensible>

There is actually a sensible answer to this.

During the lifetime of a datacenter (say, 15 years for a regular non-AI datacenter?), the processors are going to be replaced several times, (usually) increasing the processor power. So while processing power is a good measure of the initial capacity of a datacenter, it is also a constantly-changing number, and doesn't tell you anything useful about the capacity, say, five years from now.

The maximum power capacity, however, is a much more stable number, as it's a major engineering exercise to increase it. It's also often the limiting factor, especially in the current madness where the demand for power for datacenters far outstrips the ability to build out more (which is why you have, for example, Microsoft bringing Three Mile Island.) And third, it's hard to measure the processing power of modern datacenters. If you have a mix of CPUs (some ARM, some x86), GPUs, TPUs, etc., how do you add all those up to a sensible number that is comparable across different datacenters?

</Sensible>

We now return you to your regular El Reg comment thread in progress

[MachDiamond]

Re: Is the return worth it?

"So while processing power is a good measure of the initial capacity of a datacenter, it is also a constantly-changing number, and doesn't tell you anything useful about the capacity, say, five years from now."

I'd still say the metric is useless and dependent on how efficient the software is. Nobody seems to have a "standard" question that can be run on multiple systems with measurements on how much power was consumed to present an answer. It could vary widely so maybe nobody wants to know and it may affect investments. That could be why they stick with processor count and energy consumption.

[Doctor Syntax]

Re: Is the return worth it?

"So how do we equate energy consumption to processing power?"

If you can't report quantifiable useful outputs you measure inputs and quote that instead. Any government questioned about poor service won't produce figures to show the service is actually not bad, they'll tell you how much more they're spending to improve it. Manglers who can't measure how well their department performs due to WFH will want to see bums on seats to reassure themselves by seeing the input (oddly the BoS metric isn't a useful one when it comes to headcount reduction despite the fact that heads and bums have a 1:1 ratio).

[MachDiamond]

Re: Is the return worth it?

"or is energy consumption the new metric for measuring datacentre "performance"?"

It's the same with comparing loudspeakers. Mo watts = mo betta? My friend, with some help from me, built a sound system for a boutique cinema that was all of 200W of amplifier power and caused their neighbors to complain. The low end was 2 15" speakers on massive horns for some amazing efficiency. Each "woofer" was ~60W max power (Altec 515). Many car stereo woofers are a brick with a voice coil attached that take enormous amounts of power to get audible. It's a trade off for working in a very small enclosure since cars don't offer much space.

I have agree that comments from Elon are complete nonsense when he talks about 88jigawatts of "compute". "Compute" isn't a defined metric and power usage doesn't correlate to useful work. The crawler that moves the SLS rocket from the barn to the launch pad uses liters of diesel per meter. That's not a good argument for a vehicle you want to commute to work and back. Left out is how heavy that vehicle and cargo are so a proper accounting would be ton/km/liter or some such. It could turn out that a Clio is less efficient in those terms, but more efficient in £/mile. An Electron rocket is more money per kg than a Falcon 9 and less expensive per launch if you aren't using the full capacity of the F9.

[Winkypop]

Feed the worlds poor, heal the sick, protect our environment

Nah.

Flog more tat!

[Anonymous Coward]

Oracles investment plans do not seam to have done its share proce much good

https://finance.yahoo.com/quote/ORCL/

$332 in October and $135 today

So they are actually lower now then they were before they announced the all in with AI. They lost all their gains from Ai, so it will be interesting to see how low they go once the bubble bursts.

[Charlie Clark]

Re: Oracles investment plans do not seam to have done its share proce much good

And things will probably get worse from here as it's competing for CAPEX with companies with better financials…

[Ossi]

All driven by FOMO. One tech giant is sitting out the CAPEX binge - Apple. Its shares are doing very nicely, thank you. Wait for the dust to settle, the market to mature, and the hype to deflate, then pick up the useful pieces at bargain prices - or just avoid the whole thing altogether, depending on the situation. Not a terrible strategy at all.

[Excused Boots]

You will be downvoted by the Apple haters on here, which is fine and expected.

But, you do make a good point, they do seem to doing quite well, without spending stupid sums on, well, what looks very much like the 17th Century tulip mania.

They do say that 'history repeats itself’, and ’those who fail to learn from the lesson of history are doomed to repeat it’’ Do they still teach history in Universities?

[druck]

This is not financial advice

I closed my position on tech stocks last month - don't say I didn't tell you so.

[Boolian]

Scraping the Barrel

Are companies really queing up to host all their data in a new AI cloud? I mean isn't it explicitly stated Amazon will train their AI on all that glorious IP.

Is that not "Here are the keys to the company, lock up on your way out thanks"

Am I missing something here?