In the Loop AI

On Tuesday, voters pushed back on AI data centers. Democrats won the House and the Senate, flipped governorships in Texas, Ohio and Georgia, and data-center restrictions passed in towns across the country. Goldman Sachs warns that 45% of planned US data-center capacity could be affected, and Texas alone is about a quarter of it. Ejaaz breaks down why an AI bull thinks the slowdown might actually be good for AI: no overbuild, data centers that pay their own way, and a race that rewards whoever uses compute most efficiently.

We also cover why voters turned on data centers (power, land and water), Gina Hinojosa's conditions in Texas, the water myth, the market's split reaction (stocks up, AI stocks down, Elon's companies hit hardest), the 10 to 15 gigawatts that may not come online, the new stakeholder in the AI money loop, and what it means for the race with China.

==
FOLLOW EJAAZ ON TWITTER:
https://x.com/cryptopunk7213

==
Timestamps
00:00 America Hits Pause on AI
01:34 Power, Land and Water
02:38 Why Voters Turned
04:18 What the Midterms Changed
05:08 Texas Wants Data Centers to Pay
07:07 A Bipartisan Backlash
07:29 The China Problem
08:34 The Water Myth
09:55 Who Pays for Power
10:39 The Market Reaction
12:06 A New Stakeholder
13:07 Good News for the Bears
14:19 Stocks From Here
15:34 Efficiency Wins
17:24 AI Gets More Expensive
18:03 US vs China
19:51 Final Thoughts

#ai #datacenters #midterms #elections #elonmusk

==
AlphaSpace by Yahoo Finance: A Professional-Grade Investment Platform Built for Everyday Investors.
👉 https://finance.yahoo.com/about/promos/gold/alphaspace/?ncid=100003571

Creators and Guests

Host
Ejaaz Ahamadeen

What is In the Loop AI?

Exploring the frontiers of Technology and AI

On Tuesday, one of the most pivotal decisions in tech history was made,

and it's not looking great for AI at all.

Almost 45% of planned data center capacity, data centers being the bedrock of

this entire AI economy, is planned to be paused, slowed down,

or at worst, shut down completely.

We're talking about $400 to $750 billion worth of AI capex, the same capex that

the likes of Anthropic and OpenAI are projecting all their revenue on for their

massive IPOs, all slowing down in a halt.

And the reason behind this is because of a small group of people,

oh yeah, that's right, the entire American population disagreeing and disliking

AI, to the extent where 67% of them think that AI is going to do more harm than actual good.

But in a crazy turn of events, this might actually end up being net-net really

good for the AI industry because it slows everything down, so there's no danger

of actually overbuilding.

Anyway, I want to unpack this because even though it's all to do with politics

and stuff like that, there are massive ramifications for how this affects the

AI industry build-out, the biggest build-out that we've seen in tech history.

Now, some obvious disclaimers, I am neither American nor am I a politician.

And to be honest, I think I'll do a terrible job at both of those things.

But the point around this episode is I have a lot of thoughts and I have a lot

of opinions as to how this is going to affect the AI build-out.

Stocks, the entire market, and the economy, which, better or worse resides a

lot on the AI CapEx build-out.

So let's get into it on today's episode brought to you by Qualcomm.

Now, as you already know, data centers need power, land, and water,

which basically form the kind of three issues around this whole thing.

If you need all of that stuff, you're probably not going to set one up in New York or LA.

You're going to go to a place where there's a lot of land, preferably cheap,

plentiful supply of water and electricity, and most importantly, space.

So the likes of Texas, Ohio, and Georgia come to mind, which are the three states

in the US, which have garnered a

lot of investment from the biggest hyperscalers and AI labs in the world.

When I talk about Elon Musk's Colossus 1, 2, and 3 data centers,

which are housing up to, I think, 2 to 3 million NVIDIA GPUs,

and then you've got Amazons and Meta's one, it's all happening in these types of states.

So all those hundreds of billions to trillions of dollars worth of CapEx that

I keep mentioning on every single episode are

It's going to those places. And you might think, well, that's fantastic for

them. The economies of these states basically just keep going up.

But the resources there are scarce and most importantly, built to serve other

humans, not millions and millions of machines that run 24-7.

So obviously what we've seen is increasing prices of electricity costs,

increasing prices of land. People can't like buy a home for the same type of price anymore.

And then water becomes a scarce supply as well. These are effectively the claims

or issues or concerns that a lot of the US population has had over the last.

I'd say, one to two years.

And as you can imagine, as the AI boom has just exploded and accelerated faster

than any of us could have presumed, these issues each get a lot worse.

And so the negative sentiment around AI, specifically data centers,

but AI in general, has become really bad.

It's easy for me to sit here in my AI bubble day to day and forget or ignore

what's going on for 99% of the world who doesn't actually interact with air.

Remember, air penetration is still so small right now. Majority of the world

barely use an LLM beyond just the average Google search. So when you have the

likes of Dario Mode or the Anthropic and Open Area Alignment researchers saying.

Yeah, there's a chance humanity might die over the next 10 years.

There's a 10% chance of that.

You kind of look at that, you look at your lifestyle, you look at your electricity

and water bills rising, and you're like, I don't like this thing.

I'm going to vote against this thing. I don't want the data centers in my town.

And so that's effectively what was voted on on Tuesday.

And it particularly pertains to three main states, Texas, which accounts for

25% of all the compute commitments and investments that were made for next year,

Georgia, and I think Ohio.

So collectively within these states. And it's funny that it's not actually like

we're not talking about nations here.

We're not even talking about the entire U.S. We're talking about these particular

states that have attracted the most amount of investment.

These are the places that we want to focus on. Now, this isn't a lecture on

politics. Like I said, I know nothing about this stuff. But the relative way

to think about this is since Trump got elected, the Republicans pretty much

had an overwhelming majority of power.

And since then, we have the midterm elections two years later,

where we're talking about not a change in the president, but the changes for

all these different townships, states, senates, etc.

So that's what happened on Tuesday. And it was overwhelmingly leaning to the

Democratic side of things.

So they effectively seized back more power than was initially sprung on to the

Republican Party when Trump took over.

And so now there's kind of roughly like a 50-50 kind of say of things of what happens in AI.

And this has a major impact as to what kinds of data centers get set up and

what kind of legislation gets passed to, you know, fund the entire AI CapEx

build out in the first place or enable it.

So the point is there was an overwhelming support for anti-data center movements,

and they ended up getting what they want.

Now, let's focus on Texas, where a Democratic candidate won for the first time

since 1990. And her policies are very simple on data centers.

One, no more subsidies. If you want to pay your fair way, you can,

but we're not going to give you any tax incentives to do so. Number two,

We want proof that the electricity and water bills for the township,

for the community, for the people that live in the area does not go up.

And if it does, you will cover those costs or even at best subsidize if you have surplus energy.

Number three, that there's some form of union labor formed so that employees

aren't taking advantage of.

Now, my initial reaction to this...

Isn't negative. And I think it's how these data centers should be constructed.

Now, to give you a bit of context here, what it looked like prior to this election

and all of this happening was governors or state leaders will basically say,

hey Elon, hey Meta, hey Microsoft, come and build your data centers in our states.

We'll give you a bunch of tax incentives, credits, so that you don't have to

pay as much, so that you don't have to pay the fair market price.

And the idea behind this is that they can attract more investment in that recent area.

It'll result in a lot more employment jobs going up, employment opportunities.

And naturally, the economy of those towns, of those cities, of the states will

ascend and increase and enjoy from the spoils of labor or wealth that AI data

centers generate. So that was their kind of thinking back then.

But the overwhelming sentiment is that one, majority of the population don't really want this thing.

And then two, if you're going to do it, you have to pay your fair way.

And when I look at this, I think that's pretty good.

And that was pretty much the same story for all the other major states in Ohio

and Georgia, where people have been very kind of standoffish with the data center thing.

And two things I want to point out here. Number one, I think it's practical

to kind of slow down the build out of these data centers to make sure that they're

done right. Emphasis on doing it right so that people pay their fair way.

But number two, it's interesting that AI in general has become very much a bipartisan issue.

It's not just one particular party that was advocating for all of this.

Before this election, I started to see a bunch of the opposing party,

the Republicans, basically ask for the same thing because the constituents,

the people of America, were asking to slow things down.

Now, my initial thought about this as someone who is incredibly bullish about

AI is, God, this is gonna be a bad thing. We're gonna slow down progress, which means that.

Major competition with the likes of adversaries like China is going to beat

us, especially when it comes to power and electrical energy.

I mean, think about it, right? US alone has limited scarce amount of electricity

or abundant electricity compared to China, which I think has on order of magnitude

three to five times more electricity and abundance energy to fund their data centers.

Now, they have the kind of opposite end of the problem, which is they don't

have enough high quality GPUs.

So they either need to buy or smuggle them in from NVIDIA or build their own

cells, which is what we're seeing with the likes of Huawei. And they're doing

a lot of investment there. On the US side, we have all the sexy machinery,

but we need the energy to actually turn them on.

And recently, we've seen a bunch of these GPUs pile up and start to collect

us to the tune of hundreds of millions of dollars worth of these GPUs.

And that alone was a problem and we had to solve it. We had to turn on power,

generate more electricity and get it to these machines.

That was already a problem before these elections and now after these elections

it's become a bigger problem but i do want to address one other important issue

which i don't think is true.

Water thing. I've mentioned this on a previous episode, goodness,

it was like last year at this point, that the sentiment around AI data centers

consuming a heck ton of water is just not true at all.

In fact, the viral piece of literature, I forgot the name of the book,

that actually spread the rumor that AI data centers were using so much water

was actually undoubtedly false.

In fact, the author actually corrected herself. But of course, that didn't go viral.

The seed of the thought or the myth had already kind of been spread.

Now, to give you context here, if you look at Semi-Analysis's take or review

of this entire thing, they broke it down very simply.

If you compare the largest AI data center, I think at the time they used Elon

Musk's Colossus 2, which had around, I don't know, 300,000 to 500,000 GPUs,

and you compared the water consumption of that data center to

one or even two In-N-Out burger restaurants uses more water than that massive data center.

And of course, the reason behind this is in the olden days, the data centers

used to use a lot of water and they used it really ineffectively,

really inefficiently. But nowadays, we just recycle the same supply of water

over and over and over again.

So we don't actually end up using a lot of water and the water bills don't actually

go up. That's just kind of like a rumor at this point.

But the electricity thing is a very valid concern.

In the defense of hyperscalers and all these AI labs, they've made several moves

to resolve this thing. In fact, give you the benefit of subsidies.

So in many cases, Microsoft and I think Meta has said for all their data centers,

listen, if we see that the electricity bills are going up, we will not only

cover it, but we will subsidize the local community's electricity bill so that

you end up paying less on average.

I'll believe it when I see it, but the point is we're making steps in the right direction.

And I don't think we're going to see pushback from any of these hyperscalers

and data centers who are supposedly meant to be generating hundreds of billions

of dollars worth of revenue from all of this, and they already are,

why on earth would they care about paying the cost? They'll just simply pay

the cost. And that's my prediction that's going to happen.

Which leads me to the stock market reaction to this entire news.

And it might surprise you, it certainly surprised me, because it was very mixed.

The stock market or the stock index in general is now only 3% off its all-time

highs. The markets actually responded really positively.

But where it responded negatively was the AI sector specifically,

which has seen broad drawdowns across any and all semiconductor and AI stocks.

So you can see the public reaction is, this is great for the economy.

We're going to be okay. We're going to do well. Let's buy all the stocks that aren't AI adjacent.

And for all the AI bulls out there, you're probably not having the best of days.

But to be honest with you, I think a lot of the negative sentiment is overdrawn.

This is by no means a moratorium on AI data centers.

Things are just going to have to slow down. And that's not necessarily a bad thing.

By slowdown, I mean all the projections of having gigawatts worth of compute go live next year.

I think the estimate from semi-analysis was 47 gigawatts, which is a lot more

than all the gigawatts that America has ever turned on for AI data centers.

So a big ambitious goal already.

That's going to be cut down by around 10 to 15 gigawatts by estimates if you

follow what Brad Gerstner said from ultimate to capital. So the idea here is,

okay, we're going to have to lower down our estimates of how much compute that

we can deliver. The demand is still going to be there.

And we're still going to be able to stand up the majority of the compute.

We just need to comply to these new rules and pay a fair price.

I think that's a good outcome. So I think the market is overreacting slightly.

But where there is now a new interesting, I guess, stakeholder in this whole

AI circular loop is typically you would have the AI labs, open air and anthropic

paying the hyperscalers or NeoCloud's access to GPUs, right?

Once you get access to these GPUs, you can then kind of like train your models

and inference your models.

And then these hyperscalers would buy their GPUs from the likes of NVIDIA overwhelmingly,

and then maybe some from AMD and Intel.

And so that was kind of like the circular loop, right?

Now you have this new stakeholder, which is the governors or electives of all

these different towns and states,

all the senates that basically now have decision-making power as to whether

you even get to build your fancy little data center toy shop in their backyard.

There's rules and you're going to have to comply. Otherwise, you will fall behind.

So there's this new vector that you need to consider in this entire AI build-out

when you map into kind of like the projections of whether Anthropics IPO is

going to be worth that much or whether they're going to be allowed to earn as

much revenue or whether they even have the demand to sustain it.

Now, for all the bears out there, listen, I was arguing this bear case a few

episodes ago as well. When I think about.

The concern around building or investing too much into AI data centers or buying

GPUs to stand it up because we don't know if the demand is going to be there,

man. We don't know if we're going to have ROI on AI.

This is probably a really good thing because it just slows everything down,

which means that even if we want to spend all that money, we can't exactly spend

all that money because legally, we're slowing down.

Now, you might then be worried about, well, what about that $900 billion CapEx

commitment that Anthropic alone is committed to over the next couple of years,

what happens to that? Aren't they on the line to pay a lot of that?

Well, I think they're technically on line to pay around 40 to 50% of that.

And so yeah, there's definitely a huge risk and concern there.

But one, they have a huge cash balance sheet. And two, I believe a lot of that,

or the majority of that is cancelable contracts and dependent on certain conditions.

So it's not as highly levered as the dotcom bubble or the 2008 type credit default

swap type stuff. That's not what we're dealing with here. So with all of that

said, where do we go from here?

I'll get into it on today's episode brought to you by Qualcomm.

First, stocks. Personally, I don't think they're going to be affected that badly at all.

Now, today, or at least at the time of recording, seems like one specific person's

company stocks have taken a beating.

Elon Musk, Tesla and SpaceX, I think, are down on average 7% to 10% as a reaction

to this because it's like, OK, they're making one of the biggest bets.

Elon Musk has purchased the most amount of GPUs. I think he's the number one

buyer of GPUs from Jensen Huang from NVIDIA.

So he's obviously going to suffer a lot when he's invested so much money into

these different states to build his data centers. That's obviously going to

take, you know, a backseat for a bit. But the rest of the market,

even though AI stocks have drawn down, I think are going to do very well.

If you look at the historical context of the stock market six months from every

midterm elections over the last, I think it's like 50 years.

It's always gone up on average between, I think it's like 7 or 20%.

So if you want to look at historical data, I think we're going to be okay.

And as I mentioned earlier, AI at data centers, whatever way you want to look

at it, is a bipartisan issue.

So whether you're on one party or the other, you're going to recognize that

this is an important thing that generates a lot of wealth and jobs that could

increase the improvement of lifestyles for these different cities and towns.

And so it should at least be considered. And people, I think, will come along to that.

Now, two, if you are anthropic or open AI, the game so far has been spend as

much money on compute as you can and you can build a better model.

So basically, if you have more money to spend, you'll have a better AI product

and therefore you can keep raising more money and making more money, etc, etc.

That game changes completely now because you have limited compute and you'll

probably have limited compute for the next couple of years whilst we figure

out the legislation and laws.

So that means the winning AI lab is not only going to be one that spends the

most, it's going to be the one that uses the compute most effectively.

So the most efficient AI model that uses the fewest tokens to give out the highest intelligent answer.

Can end up winning. And why I find this so interesting is

whilst Claude and ChatGPT have demonstrated a really good capability of being

able to do this, especially with the newer versions,

some of the open source or cheaper models have actually done a really good job

of doing so by side effects or by a consequence of not being able to access

the amount of compute that Anthropoc and OpenAI have been outwardly buying up.

So they've had to kind of adapt.

We've seen this most prevalent in China with the open source labs kind of running

these experiments because they haven't been able to get their hands on the bleeding

edge GPUs that Jensel Huang and NVIDIA are selling.

So they've had to get creative about how they build these models.

Those models are some of the most efficient out there. So it's interesting to

see in one way, the playing field level out a bit over the next couple of years.

And the other point is those who can adapt the most. And what I mean by that is.

Kind of get around the gray area, red tape, and actually set up power and deliver

power to GPUs in however way possible, will end up winning.

And there's one man, whether you like it or not, that has shown his experience

in being able to do this, and that is Elon Musk.

He's been the fastest to scale these GPUs. Jensen Wang himself has said that.

Now, the bad thing here is AI on the whole is probably going to get a little

more expensive. Remember, we presumed that there would be an abundance of compute

that these labs could use, that the cost of compute was going to go down.

If you add a lot more employees, if you add a lot more laws,

if you add a lot more taxes and remove the subsidies, naturally the cost of

AI or the margins are going to compress.

Now it's going to be up to the likes of Anthropic and OpenArea whether they

want to kind of pass on those costs to the customer and raise the cost of their

subscriptions, or if they just eat the short-term cost whilst betting that the

ROI is going to increase in the future.

If I was a betting man, I'd probably say that they would go with the latter.

Now the The other most obvious thing is the battle between the U.S.

And China. So far, the U.S. has held the lead.

Centralized Labs, particularly Anthropic and OpenAI, have had the best AI models.

And even when they've kind of slowed down the new model releases,

we've seen that the open source competitors from China haven't really been able

to catch up. When I look at the most latest open source model, I think it was Kimi Kus.

K 3.7 that released this week in fact the weights went live this week you should

check that out if you're a open source developer

uh it's actually pretty good isn't as good as the frontier llms and so it will

result in a few things which is number one these labs will continually keep

these super smart genius asi level models.

Inside and not release it to the public because it's too costly to serve but

it will not stop them from training the ai model so if you're listening to this

and you're listening to all the election stuff and you're thinking,

huh, slowing down AI data centers is actually going to stop the build out of AGI.

No, that's not going to happen. It's just going to result in a more insular

build out of these AI models that will be continually restricted or privatized

so that majority of the public don't get access to it.

I don't believe that's a good world. And so I would like to see more abundance with this thing.

And if you peek over the fence at China, they have the opposite problem to us,

which is they have free legislation to spin up as many data centers and they

have an abundance of energy, what they're lacking is the infrastructure side

of things and they're plowing in billions of dollars to try and figure that thing out.

If it gives them some time to catch up, that could end up being a potentially

contentious issue in the future.

And I'm not exaggerating. If you look at the recent couple of weeks of cyber

attacks, imagine if those models were created and in the hands of adversaries

that could use that against the Western frontier or whatever that might look like.

You could imagine there being some dramatic scenarios that could happen in the future.

So my general thoughts on this is, one, don't freak out from a market's perspective

and a slowdown perspective because for the markets, I think they'll do fine.

People are still very bullish on AI and it's going to be a very necessary technology

on both sides of any political party. So it doesn't really matter from that sense.

But two, the slowdown might actually end up being a really good thing.

And I look forward to that being the case.

And three, this is a really good opportunity for some of the smaller players

in AI to potentially catch up.

And three, see some really important, I guess, optimization efforts,

like models that don't just waste tokens that actually end up providing the

value that you want to see. Like, it's been a very...

Common take, I think, at least within my circles, that AI models today are extremely

competent of doing really magical things.

We're seeing that with internal models solving Navier-Stokes mathematical problems

and making huge advancements in physics, chemistry, and medicine.

And now if things start to slow down, we really get to focus on the technology

that we already have at hand and figure out how to diffuse that to the rest of the world.

And I'm a big proponent of this, because when I talk to the likes of my family,

friends who maybe don't use AI models and agents and have a bunch of things

running 24-7 whilst I sleep, I kind of think, huh, maybe you guys are missing out on some stuff.

I wonder what would make AI important for you.

And maybe that generally changes sentiment going forwards.

So yeah, I guess that's the episode. I tried my best to stay away from this

being a political thing.

This is not a politics show by any means, and I'm way out of my remit to do that

but i still thought it was an important technology i guess story and analysis

to tell and hopefully you found something valuable from this so

if you enjoyed this episode and you listen to this on youtube spotify apple

music wherever you listen to this please give us a thumbs up subscribe to us

turn on notifications that helps us out so much

also give us a rating and leave me a comment i love all the comments that i've

been receiving send me a dm my x details are in the descriptions below and i'm

posting my new thoughts as they come out so you don't have to wait for the podcast, I guess.

And yeah, if you have a friend or colleague that might be interested in listening

to this, send it to them. They'll find it interesting, hopefully.

And if not, maybe I'll get some good criticism. So thank you so much for listening.

And I'll see you on the next episode of In The Loop brought to you by Qualcomm.