In the Loop AI

Over the weekend, Elon Musk told a Grok Bot user that "the cost of serving Grok Bot was unsustainably high relative to revenue," and usage limits tightened for hundreds of thousands of paying users. Three days earlier he'd pointed out that Grok Bot has no free tier, and two weeks before that he said he'd never seen anything grow as fast. Ejaaz breaks down what's actually going on: SpaceX can rent the same GPUs to Anthropic, which pays $1.25 billion a month today and about $3.1 billion a month from December, while Grok Bot's hardest tasks run on Anthropic's own Claude Opus 5.5.

We also cover Grok Bot's rise to 600,000 daily users, the delays on Grok 4.8, the $60 billion Cursor acquisition, the trade-off between renting compute and running his own AI agent, the bear and bull case, and whether this is the first sign that AI is too expensive to sell at today's prices.

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

==
Timestamps
00:00 Elon's Grok Bot Problem
01:51 What Grok Bot Is
03:48 How Grok Bot Took Off
04:47 Grok's Model Problem
06:18 The Contradiction
06:58 The Landlord and the Tenant
08:35 The $60 Billion Cursor Bet
09:53 Rent the Chips or Run the Bot?
10:50 Why AI Agents Matter
12:46 Limits, Not a Shutdown
14:01 The Bear Case
17:46 The Bull Case
19:15 Final Thoughts

#elonmusk #grok #spacex #anthropic #ai

==
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

Investor & Host of the Limitless AI Show. Former Product at Coinbase and ConsenSys

What is In the Loop AI?

Exploring the frontiers of Technology and AI

Guys, I've been trying to make a bullish episode on AI for so long,

but it keeps getting harder and harder because stories like this keep falling in front of me.

And this time it's from one of my favorite advocates in the business,

Elon Musk, who announced to hundreds of thousands of people over the weekend

that the cost of serving Grokbot, their flagship AI product,

has become too costly relative to revenue.

Now, this would be all well and good if not three days earlier,

he said that it is actually one of their fastest growing products ever and emphasizing

the fact that there is no free tier.

So people were paying money up to $300 a month to get access to this thing and

he still couldn't generate a profit.

And this is weird because it's coming from the guy who's arguably invested the most in AI

to the point of tens of billions of dollars a month to stand up GPUs to become

the number one neocloud in the world, but also simultaneously calling potentially

the top of the AI bubble and saying that hundreds of other neoclouds are going to crash

and the fact that it's really hard to fund the capex build out in AI.

On one hand, he's super bullish about the long-term implications of AI and the

money it's going to generate.

And on the other hand, he's unable to do so on his flagship product and his

Grok AI models, the fundamental bedrock of the entire thing,

also hasn't been able to catch up.

So something's not quite adding up and I have to just call it out when I see

it. And this is coming from one of the biggest Elon supporters out there and

one of the biggest AI advocates out there. I promise you, deep in my bones,

I feel extremely bullish and optimistic about the future.

But I have to look around me and see that market stock prices are crumbling.

Funding is getting really hard. IPO valuations for Anthropic and OpenAI are

rumored to be slashed. And I guess we've got to call a spade a spade.

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

Now, first, some quick context. GrokBot is basically SpaceX AI's flagship AI

product. It's their AI agent, which competes with the likes of Meta Muse,

which is their own personal super assistant.

OpenAI's Dots, which is a similar product as well. And then you have Claude, what's it called? Duet.

And so the idea behind these AI agents and what makes it different to a model

is that they're basically your shepherd or effective channel into the rest of the digital world.

So the idea is you can interact with these agents and they'll spin up multiple

versions of themselves and go and do a bunch of cool stuff for you,

like book tickets for you, book reservations, go sort out and manage a friend's

trip, that kind of stuff.

But also equally on your career side of things, you can go and automate a bunch of your work.

And remember, this is the whole core thesis of Elon's long-term AI vision,

an idea of a human emulator, AI that can basically automate a bunch of software

interactions that humans do today.

And since launch, GrokBot has seen a pretty meteoric rise. Honestly,

in my opinion, I rank it as either one or two of personal agents out there.

And it's really effective at doing work. Specifically, it's been up hundreds

of versions of itself, and it can go and effectively take over things like slack

or jira or salesforce or whatever that might be

and actually effectively and competently do that work and it's scaled to hundreds

of thousands of users i think the most recent data point was 600 000 concurrent

daily active users that's a lot of people that are reaping value from this and

that's a very important point to make which is there is no free tier version

of this you can get it access to it

for as cheap as i think 30 a month

but up to grok heavy subscription, which is, I think, $300 a month, which is.

Up there with one of the most expensive subscriptions out there.

I think it's only second to OpenAI's $500 a month thing.

If you have that many users using a paid product, the idea is they're reaping

enough value from this. So this is probably a product that you want to double

down on. So why then has Elon decided to pull back on investing into the flagship product?

Well, let's look at the timeline. When GrokBot eventually launched,

it had a meteoric rise, at least initially.

But over the next two weeks, it then decided to flatline. And Elon Musk went

back to his team and was like.

Hey, what's going on here? And essentially, the product team told him that the

product is actually quite good, but the underlying models, which they were using

Grok, which is SpaceX's flagship model,

wasn't really working for the product. It wasn't really sufficient enough.

So Elon's response to that was to go model agnostic, which is say,

hey, use whatever model you want for GrokBot. The most important part is that

the product is actually good and serves the user.

So they had the bright idea of being like, okay, well, let's use Claude Opus

5.5 or any anthropic model.

And so they did that. And the product, again, achieved virality to the point

of where they had 600,000 concurrent users.

So that's all well and good, except I have to put myself in Elon's shoes and

think that must be pretty expensive.

Think about it. His whole thesis for Grok the AI model over the last six to

nine months has not just been, hey, I'll achieve the frontier LLM and be able

to compete with the likes of Fable and ChatGPT6.

Astra, but I'll also be able to be the most cost efficient, Pareto efficient

model out there, which means that you can use our models for anything between

low level intelligence tasks, as well as high level intelligence tasks.

Okay, that's a good and fair enough thesis. So Grok as an AI model definitely

serves that bottom low end task.

But of course, people want to be able to do longer, complicated,

harder tasks. So you need higher intelligence.

And to be honest, SpaceX AI has sort of fallen behind. They haven't quite fallen

behind to the last place, but they've struggled to scale up models.

And that's not a secret at this

point. We've seen multiple delays on delivering very effective models.

And even when they have, it has been to a lackluster audience,

a reception where people have basically said.

This can get beaten by GPT-6 Luna or one of the other cheaper models from one

of the flagship AI labs. And that's another thing worth pointing out.

The flagship AI labs typically could only put out expensive models.

And then they thought, why don't we just distill some of the more expensive

models to create cheaper versions so that everyone stays within our ecosystem?

And that's the competition that Elon Musk has come up against.

And to be honest, I have to give Elon credit where he's frequently said he doesn't

care about optimizing models for benchmarks, which is what a lot of other AI labs do,

he wants it to actually be meaningful when his users actually end up using the product.

So he's taking his time and I can't fault that.

But I have to acknowledge the growing bits of evidence between Elon's recent

pullback on GrokBot and other bits of evidence behind his acquisition of Cursor

and the way that he's spending money on compute that kind of tell another story

that is contrary to the one that he's telling us.

The first is, on one hand, Elon is saying that compute is the ultimate resource for AI.

And I agree with him. Those who acquire the most compute could effectively be

the most likely to train the best models or at least serve the best models.

And so Elon Musk has been the best at that. He's been able to acquire more GPUs

and turn them online than any other company.

To the point where SpaceX AI has become the number one neocloud,

which means that they're effectively serving the who's who of AI Labs,

the main one being Anthropic, with which they're earning around $3.1 billion per month.

So you could argue that he's making enough money from that to fund something

like GrokBot, of which is a very important market share to kind of own.

So it's worth kind of sinking the costs and taking the costs in order to win

this market. But he has a problem, which is his Grok AI models themselves aren't

necessarily good enough.

And so he's had to resort to leasing out or accessing the models from the very

AR lab that he's selling compute to.

And that definitely eats into his margins pretty aggressively.

So if we look at their recent quarterly earnings, which they made a ton of money,

the first profitable quarter, they're making a bunch of revenue.

It's looking very bullish.

Remember, we spoke about that a few episodes ago, where like the projections

for next year is something on order of magnitude of $500 billion ARR to a trillion

dollars ARR. Yes, that is a real figure that came out of Elon's mouth during

the quarterly earnings report.

Well, why can't we use some of that money to fund the GrokBot acquisition?

It's because margins are being eaten into.

Well, it seems like that revenue has a lot of costs incurred to it if they have

to rely on expensive Claude models. And that's what we're seeing with GrokBot.

So there's this really interesting flywheel of Elon being the landlord supplying

Anthropic with the compute and.

But then using the tenant's AI model to make his product functional,

which seems a little bit circular.

And it's not just Claude Opus that he is renting out.

He's also using the likes of Suno, which is a music model, as well as Mid Journey,

which is an image and video generation model.

So he's taking a very agnostic approach, but all of this incurs a lot of costs.

The other elephant in the room is the hefty $60 billion.

Cursor acquisition, which was one of the biggest acquisitions at the time and

the whole premise behind this was

we're acquiring the cursor team for two reasons one their talent who will help

us build a much better model specifically better at coding ai and number two

we're acquiring their agent harness which is the data or the harness that they

wrap around ai models which will end up allowing them to make a better ai model

so there's a software stack that they're acquiring and there's the human density talent.

We haven't quite seen the ROI of that. Yes, we've seen cheaper,

more effective models that cost fractions of what it is to do a same particular

task, but you still also need the frontier level models.

And if we think back to cursor pre-acquisition, they weren't necessarily able

to build a frontier LLM. And the idea was they were going to combine forces

with SpaceX AI who had an AI lab at the time to be able to do so.

We currently haven't seen the fruition of those results. And so what that's

done is it's eaten into the quality and cost of the actual flagship product.

And so Elon has to go to a much more expensive vendor like Claude in this instance

and say, hey, I need you to plug into my backend and be able to kind of deliver

a better superior performance.

But it comes at the cost of making money.

The other juggling act that Elon is doing here is there's a trade-off between

where he allocates his compute.

Now, let's say he serves all 600,000 users, and they're willing to pay him 200

bucks a month, which is like they're just below most expensive tier.

That will result in around, I think, $120 million per month,

which is around, let's call it $1.5 billion per year.

He makes double that from one single vendor, Anthropic, in a single month.

That's like 100x higher return on compute.

So the logical business sense is to just keep selling compute to Anthropic and

all these other guys at a premium to make more money and allow SpaceX to be profitable.

But that's at the sacrifice of SpaceX no longer being a frontier rocket company or AI company.

It becomes a cloud service or neocloud company, which isn't necessarily a bad

thing, but is very differentiated from the thesis that Elon has pitched for the company.

Now, to kind of step back for a second, I think about short-term and long-term goals here.

So if you're Elon Musk, you want to be able to own whatever the panacea looks like 10 years from now.

And in my opinion, AI agents or personal assistants are going to be not only abundant...

Arguably the most important and valuable layer to anyone's interaction with AI.

What do I mean by that? Well, right now we have to pull up our phone or look

at a computer screen. We need to type buttons, scroll, search for apps,

navigate, look for which button to click, look for the information,

ingest it, think about it, and then kind of imbue it onto the rest of the world.

That is incredibly archaic when you have a personal assistant that can kind

of do a bunch of that with you and for you and make you incredibly more competent

so you have more time to go do other more productive stuff.

So I think that we're going to see the world shifting towards using these personal

agents as the way to interact with the rest of the digital world or the internet.

And indeed, we're starting to see that more now.

If you look at Google's recent quarterly earnings, you'll see that their search

revenue, although it's increased, has started to slow down because a lot of

people are getting their information directly from chatbot LLMs like ChatGPT and clawed.

So if I were Elon and I was observing this, I would do a balancing act,

which is I need to make enough money by selling cloud compute to the likes of

Anthropic and stuff and make a bag, a huge amount of money.

But I also want to spend some of that money to keep funding the costs of AI

agents, in this case, GrokBot, which many hundreds of thousands of people seem

to be enjoying and paying for,

because it means I will capture the market of a much more important or future valuable market.

Than anything else that exists right now.

So with that said, it wouldn't surprise me if in the next quarterly results,

we'll see that Elon has done kind of like a flip-flop act and he keeps continuing

to do that whilst he's able to kind of like fund this build-out.

It's important to note that he hasn't killed GrokBot, he's just reduced the

usage limits and it might be a sign of either restricted compute or increasing

costs. The other interesting part about this is I guess compute isn't as abundant as Elon Musk.

Elon or rather industry experts have been making it out to be.

Now, listen, this isn't a surprise, right? We have all the GPUs,

but it's getting harder and harder to actually power them on.

There's hundreds of millions of dollars worth of GPUs owned by Amazon that are

sitting in a warehouse collecting dust. So I get it. It's hard to spin these things up.

But at the same time, Elon has been the number one competent individual and

company that has been able to do this.

And so the idea here is, and the idea behind this whole thesis is,

we'll be able to spin up so much abundant compute that it won't even matter

over the next couple of years.

If that's the case, then maybe we could read this restriction limit as something

that is short-term and temporary, or maybe it's part of a larger problem that

might extend to longer-term stuff.

Remember, the entire thesis of SpaceX's compute hasn't got anything to do with data centers on Earth.

It's to do with data centers out in space that can harness the energy of the

sun 24-7 and end up becoming the cheapest source of compute,

which could be used to train and inference any of model he'll have so much compute

that you can sell it to whoever use it for his own ai models

which leads me to a few conclusions which includes a bear case a bull case and

my overall general take which is the thing i actually believe in on today's

episode brought to you by qualcomm.

So number one is there's been this huge debate as to whether ai can actually

pay for itself you've had a few reports come out recently which has said,

Aside from the top seven companies in the world, we haven't seen ROI come out

from most of the companies that are using AI.

Then you look at something like Anthropics S1, and you see that they're obviously

making a heck ton of money if they're able to pay 900% more than they eventually

started at the start of the year to Anthropics Claude model.

So there's kind of like this mismatch.

Now, if you look at the Elon case of restricting GrokBot, it shows you that

he's probably not making as much money as he would hope, and it's not as profitable,

therefore, as an investment in compute or capex

and listen if you're looking at the one to two trillion dollars that are going

to be invested over the rest of this year and throughout 2027 in capex from

seven to ten companies including spacex ai and the likes of google amazon you

could start getting a little hot under the collar and i wouldn't blame you for

it because if you don't see necessarily roi how can you justify the spend

then you look at the likes of open models where you see a lot of companies sort

of shifting their use of primary frontier LLMs from the likes of ChatGPT and

Claude to these more open models to defend them against cybersecurity risks, where we've seen.

Numerous hacks surface over the last couple of weeks, as well as just cheapen

down your costs when you're spending hundreds of millions of dollars versus

tens of millions of dollars versus hundreds of thousands of dollars to do the same thing.

It's obvious what trade-off you should make as a business, but also there's

the point of owning your data and owning your model so that you don't leak your

product to the likes of Anthropic or OpenAI who could eventually copy or change your product.

So the point here is that there's multiple things coalescing together,

which would imply that people aren't spending as much on AI models as they used

to, or if they are, they're spending it on different types of models and different types of use cases.

And the number one thing you're going to need to be able to provide if you want

to compete in that market is a competent model.

Which brings me to my second thing, Elon Musk and Sabasex AI so far has been

unable to put out a frontier number one model. If I think back to the end of

last year, that was at 2025, just as Claude Code was getting up off the ground now,

Grok had a really competent model. But since then, they've been falling behind.

And listen, other companies have as well. It's a very tricky business.

Google, even with all their money, has been unable to do so.

And I think it's a mixture of governance, like allocating resources within your

company to be able to fixate and actually build a really good model.

I think number two thing is data, being able to acquire enough valuable data to train a good model.

And then number three is acquiring the talent. And when I look at SpaceX AI

and I look at Elon Musk, he has certainly been the most.

Betting man to try and do this but it's been tough and listen we've seen this

being pivoted and turned successfully with other companies the most obvious

one being meta and their product muse who

indirectly or directly competes with grokbot now grokbot is more oriented around

work stuff muse is more of a personal assistant that kind of like does casual

leisure type tasks for you but the point is they functionally are the same thing

they both run Similarly, on the back end, they have their own virtual computer

and they can go off and do multiple things whilst you're asleep.

And so the idea here is it is a decent one-to-one comparison and Meta,

who had been spending so much money, arguably even more than Elon Musk has,

to acquire, compute and build good products, has been able to pull that off.

So I think there's a bit of a talent delta between the two companies.

Now, if I were to give you the flip side, the bullish side of the argument,

Elon has a shortage of compute right now.

And it's a good thing because the compute that he's selling to other companies

is making the company the most money and arguably one of the most profitable

companies next year of all time if he's able to continue doing so. So he would fund the Grok.

Build out, but he currently has limited compute supply.

And that is a good thing because it shows that there's overwhelming amounts of demand.

And remember, even if GrokBot isn't being highly desired by individuals that

are willing to pay for it, although $600,000 is quite a lot.

In my opinion, Anthropic obviously does.

That's why they're renting so much compute. So the demand, however you want

to look at it, for these AI models is definitely out there and you have people

paying to access these things.

The other thing is, to Elon's credit, Grok 4.8 still hasn't come out yet.

And maybe when it comes out later this year, or by the end of the year,

I think he mentioned that GROC 5 is going to come out, and that's going to be

like a 10 to 15 trillion parameter model, which will put it on the same level

or playing field as some of these larger astrotype models.

Maybe we'll see a flip of sentiment because the models are just good enough

to do the thing. Remember, if I use MetaMuse as an example, the model that's

powering it on the back end isn't clawed, and it isn't even a front LLM.

It's Muse Spark 1.3, which is a Meta model, which isn't the best model.

The idea here is I think if Elon is able to figure out training the model using

very proprietary or good data, which to be honest, you can just get from X social

media platform or any other one of his companies, he'll be able to create a really good model.

So just to leave you with my final thoughts, my collective take is,

listen, the market is in a bit of a bearish sentiment right now.

And there are some good reasons for that.

There's been the largest amount of money, hundreds of billions of dollars invested

over the last three months alone, three to six months, in building out CapEx

GPUs data centers to fund this AI build out.

And the very valid question that's being asked is, where's the money coming from?

And then when you see a headline like this, where it's like,

GrokBot is being restricted because it's not making money, but three weeks later,

you see it's the fastest growing product ever, you can kind of guess that there's

some kind of PR thing going on. And I wouldn't blame you.

But I think what's really going on behind the.

A lot more nuance, which is the compute is there, but there's not enough of

it. The demand exceeds the amount of compute that is available.

And Elon, as a CEO of one of the big frontier labs, needs to make a decision.

Do I sell compute in the short term to make enough money to fund out the build

out of all my other AI products and Grok models?

Or do I just go all in on the Grok models now at the sacrifice of having bad financials?

And I see that it's a, listen, I'm not the CEO of a multi-trillion dollar company,

but it's a hard decision to make.

The other thing I'll say is I do believe AI agents in the long term aren't going

anywhere. If anything, they're going to proliferate. And that is a bet I'll

make hands down for the next couple of months, for the next couple of years.

Because the point is, people don't want just general LLMs, they want personal LLMs.

I did a review episode, I think two weeks ago, covering all of these different

things and why one is good over the

other, but why we should generally be interacting more with these things.

If you are someone that last used an AI model or uses it casually as a Google

search replacement, I implore you to actually play around with these agents

because they're a lot more competent than you'd expect. And the last thing I'll say is

Yes, I think a lot of this seems very circular right now. And in many cases, it might be.

But if you look at this specific example and any of the more recent examples

with the Anthropic S1, etc., you'll notice that a few things are very familiar.

Number one, there's outsized demand for available compute. Now,

whether that continues, I don't know. I'm not an expert. We'll see how it goes

over the next couple of months.

But if that demand is maintained, then we need to be able to effectively allocate

compute to either train or build models or to fund and sell compute to other

model providers, which in Elon's case is exactly what he's doing.

So I hope to see the return of high usage limits on GrokBot.

I wanted to use it as a headline to kind of like discuss this wider topic around,

you know, are we in a bubble? Is this thing deflating in any other way?

And just give my overall thoughts because I promise you I'm trying to give a

more bullish take in general for AI. but I look at the markets crashing around

and I ask myself, why is that happening?

And I kind of test my own sentiment and assumptions to try and figure out what

is right and what is actually wrong.

And hey, I might be wrong in the future, but long-term I'm bullish on this entire

industry. So I hope you enjoyed this episode.

If you're listening to this on YouTube, Spotify, Apple Music,

please give us a rating, leave us a comment. I love reading your comments, it's amazing.

More episodes

Chapters

Creators and Guests

What is In the Loop AI?