Economics Nerd Central

[How bad boy]

You're some poster HBB, nice wan boy.

Too kind. It's only a shame it has absolutely nothing to do with my day job.

That said, if my day job was doing this sort of thing, I'd probably be on here making detailed posts about AI and <stuff>

 

[How bad boy]

In a break from some of the standard macro news around things like growth an unemployment, wanted to do some thinking on the AI boom.

I'm starting from the position that, despite being neck deep in AI professionally, I think we're in a massive boom on the infrastructure side that's going to pop pretty soon.


NVIDIA just topped $5 trillion market capitalisation. That's more than the GDP of every country in the world besides the US and China. Which is a lot.

This is a graph of its value, it was worth only $400 billion this time 3 years ago:



That should set off alarm bells. In that time their revenue has gone from just under $30bn in 22 to $130bn. Which is a pretty impressive improvement.

There's a bit of a circular economy growing up around NVIDIA that I think is fucking with their valuations:



That's not easy to understand, but as an example, NVIDIA are investing $100bn in OpenAI. OpenAI are then going to pay NVIDIA for chips. The way data centres are specced are by power, they're commiting to 10GW of chips, which in today's prices would be about $200 billion in spend.

Combined with other deals like that with Microsoft, Oracle, Coreweave and AWS, they're committing to over half a trillion dollars in spend.
They've done another deal with AMD, similar in nature.

OpenAI's current revenue is reported to be about $13bn a year, but they're losing a lot of money. They're aiming at $100 billion by 2027.

Which is a lot. However an 8x increase in revenue in 2 years is a lot.

Meanwhile, a lot of companies are struggling to get a positive return on investment on AI, an MIT study has only 5% of enterprises integrating AI tools into production at scale (I don't agree with the headline, no production deployment is not zero ROI):

Generative AI does nothing for 95 percent of companies

: MIT NANDA study finds only 5 percent of organizations using AI tools in production at scale

www.theregister.com

Bringing it back to GDP, according to some analysis, you were to take AI related investment out of the US economy, the total US GDP growth in the first half of 2025 would have been 0.1%.

Without data centers, GDP growth was 0.1% in the first half of 2025, Harvard economist says | Fortune

Is a U.S. without data centers a country without GDP growth?

fortune.com

It all looks very dotcom boom. When will it pop? Dunno, maybe never. Or maybe tomorrow. I suspect it'll need something like an interest rate hike or some absolutely disastrous results announcements to trigger it.

 

[Handy Run]

It all looks very dotcom boom. When will it pop? Dunno, maybe never. Or maybe tomorrow. I suspect it'll need something like an interest rate hike or some absolutely disastrous results announcements to trigger it.

How deep is the contamination is the big question

 

[How bad boy]

How deep is the contamination is the big question

Well, if over 90% of US GDP growth is based on AI data centres, I'd say pretty deep.


Europe isn't immune, there's a hell of a lot of money being spent on it on this side of the pond too, mostly linked to "Sovereign AI" efforts or the hyperscalers trying to get more eco friendly electricity. It's why the nordics are experiencing a boom, you're seeing things like district heating from data centres and salad growing facilities using the waste heat, all powered by expanded hydro power.

One of the worst things about all of this is that the kit has quite a short shelf life.
As an example, the xAI cluster that has enabled Grok to catch up to some of the leading models (and significantly surpass them when it comes to racism), was built on 100k NVIDIA H100 servers, distributed over 1,500 racks and theoretically consuming 70MW of power.

H100s came out first in October 2022.

The latest generation, Blackwell B200 based, NVL72 racks came out earlier this year, now you only need about 500 racks to replicate the performance and I know from experience that the H100s have trouble running the very biggest models efficiently. It'd theoretically consume 66MW of power, but in reality would operate significantly more efficiently than the H100 based system. As an example, if you're trying to run the biggest Llama model, then you can't run that on a single H100, you need to split it and distribute it, but the NVL72 works as a single system.

As a result they claim that it'd run 30 times faster in the newer GPUs. Which makes an enormous difference when it comes to how much work you get done per watt.


In a few years time, those H100s are effectively stranded assets, too old and power hungry to be cost effective against the latest generation.



The AI companies need a pretty sharpish return on investment and they're going to start getting desperate exactly as enterprises are really questioning their spend on AI.

 

[Bennyton]

In a break from some of the standard macro news around things like growth an unemployment, wanted to do some thinking on the AI boom.

I'm starting from the position that, despite being neck deep in AI professionally, I think we're in a massive boom on the infrastructure side that's going to pop pretty soon.


NVIDIA just topped $5 trillion market capitalisation. That's more than the GDP of every country in the world besides the US and China. Which is a lot.

This is a graph of its value, it was worth only $400 billion this time 3 years ago:
View attachment 45551


That should set off alarm bells. In that time their revenue has gone from just under $30bn in 22 to $130bn. Which is a pretty impressive improvement.

There's a bit of a circular economy growing up around NVIDIA that I think is fucking with their valuations:

View attachment 45552

That's not easy to understand, but as an example, NVIDIA are investing $100bn in OpenAI. OpenAI are then going to pay NVIDIA for chips. The way data centres are specced are by power, they're commiting to 10GW of chips, which in today's prices would be about $200 billion in spend.

Combined with other deals like that with Microsoft, Oracle, Coreweave and AWS, they're committing to over half a trillion dollars in spend.
They've done another deal with AMD, similar in nature.

OpenAI's current revenue is reported to be about $13bn a year, but they're losing a lot of money. They're aiming at $100 billion by 2027.

Which is a lot. However an 8x increase in revenue in 2 years is a lot.

Meanwhile, a lot of companies are struggling to get a positive return on investment on AI, an MIT study has only 5% of enterprises integrating AI tools into production at scale (I don't agree with the headline, no production deployment is not zero ROI):

Generative AI does nothing for 95 percent of companies

: MIT NANDA study finds only 5 percent of organizations using AI tools in production at scale

www.theregister.com

Bringing it back to GDP, according to some analysis, you were to take AI related investment out of the US economy, the total US GDP growth in the first half of 2025 would have been 0.1%.

Without data centers, GDP growth was 0.1% in the first half of 2025, Harvard economist says | Fortune

Is a U.S. without data centers a country without GDP growth?

fortune.com

It all looks very dotcom boom. When will it pop? Dunno, maybe never. Or maybe tomorrow. I suspect it'll need something like an interest rate hike or some absolutely disastrous results announcements to trigger it.

We're already onto the next thing in the VC world m8, too much vapourware.

The money will dry up, and anyone with cash in the likes of Nvidia is going to take a hit, temporarily at least.

Still and all, Jensen has been buying everything that sounds like it might be a new industry in 20 years, so Nvidia will stay up there in and around the other FAANG crowd.

 

[How bad boy]

We're already onto the next thing in the VC world m8, too much vapourware.

The money will dry up, and anyone with cash in the likes of Nvidia is going to take a hit, temporarily at least.

Still and all, Jensen has been buying everything that sounds like it might be a new industry in 20 years, so Nvidia will stay up there in and around the other FAANG crowd.

Agreed, they're just the Cisco of the AI boom, Cisco was worth half a trillion back then, It's still worth $280bn, so still a significant player.

Over the long run, AI will be hugely transformative and NVIDIA will stay as the key player, so I don't see them going too far down.

Currently on the lookout for the next big thing. It's probably robotics, both humanoid and industrial, with the former being an interesting curiosity that gets better but slower than people think, and the latter taking off due to manufacturing being brought back to the west.

 

[Bwana_Dik]

AFAIK Nvidia announced a rake of robotics based stuff at their jamboree last week, both hardware and supporting software - Blackwell related and the infrastructure around it.
Also autonomous vehicle stuff.
Hope it's not vaporware, I invested in a wee lock of stock at the beginning of the year, and added a few more lately . Similar to the 3:30 at Kempton Park, I hope the going suits my punt and it gets the distance.

 

[Bennyton]

Agreed, they're just the Cisco of the AI boom, Cisco was worth half a trillion back then, It's still worth $280bn, so still a significant player.

Over the long run, AI will be hugely transformative and NVIDIA will stay as the key player, so I don't see them going too far down.

Currently on the lookout for the next big thing. It's probably robotics, both humanoid and industrial, with the former being an interesting curiosity that gets better but slower than people think, and the latter taking off due to manufacturing being brought back to the west.

We're investing most of the next fund in Robotics, all in EU/UK startup's, mostly on the software side actually, as Asian software is still shit.

 

[How bad boy]

AFAIK Nvidia announced a rake of robotics based stuff at their jamboree last week, both hardware and supporting software - Blackwell related and the infrastructure around it.
Also autonomous vehicle stuff.
Hope it's not vaporware, I invested in a wee lock of stock at the beginning of the year, and added a few more lately . Similar to the 3:30 at Kempton Park, I hope the going suits my punt and it gets the distance.

I work with folks who are doing some of the robotics stuff.

The progress being made is impressive, but there's a long way to go still.

It's like the excitement about Generative AI and Large Language Models. They can get 80-90% of the way there on an impressive array of tasks, but the challenge is that any decent production system have a number of them involved, doing things like breaking down tasks, etc...

And if you have one model or agent that's accurate 95% of the time with another equally accurate model behind it and another one behind that, then you're suddenly only 85% accurate.
Which might be fine, but it sure as hell isn't ok for any industrial applications and if you're relying on the system to do something, but it gets it wrong for a multitude of different reasons 1 in 7 times, then you're going to have problems relying on it in production. It's generally ok-ish for customer services (albeit at the cost of customers who do not like AI bots), but beyond that, it's why they're struggling to get them effective in production.

It's easy to create a great demo, entirely different to make it work when there's serious cold hard cash involved.

There's two ways to think about it at scale. Six Sigma as a manufacturing quality technique (yeah, I know, bit old school...) allows for 3.4 failures per million components. Your typical smart phone might have 1000 components, so even that super impressive failure rate results in a 0.34% failure rate in production due to crappy components. If you sell 10 million units, that's 340,000 failures due to crappy components. Which is a lot and there are many, many other ways phones can fail.

That's with a process accuracy of 99.99966%. Which is really quite high.

Cloud providers generally provide 99.95% reliability, which sounds impressive but it involves allowing your service to be entirely unavailable for up to 4 hrs, 22 mins a year. Which can feel like a lot if you're dependent on it. Once you start building cloud services out of loads of individual services each with that downtime SLA, then the chance of your service being offline for a day or two a year because one of those critical services goes offline goes up rapidly.

People have hyped GenAI too much. It's amazing, I use it dozens of times a day, help customers use it and even build services on it, but it is overhyped today.

Now keep that in mind and watch this video on one of the most hyped humanoid robots:

It's cool, but it's certainly nowhere close to ready for industrial operations. It's a decent distance from ready for commercial or home operations. There absolutely will be applications where it's useful and the pace of improvement is very impressive, but it isn't the panacea that's being hyped


A way to think of it was Pets.com, which post dotcom crash, everyone laughed at, "haha, imagine just sending out a van to deliver a single package of kitty litter or dogfood, what were those morons thinking??"

I can now order a bag of cat litter to be delivered before 1pm tomorrow morning for 20% less than I can buy it in my local Sainsburys shop.


We've gone way beyond the vision that was laughed at post crash, but it was still hubris at the time.

 

[How bad boy]

We're investing most of the next fund in Robotics, all in EU/UK startup's, mostly on the software side actually, as Asian software is still shit.

As I said in my long comment above, be careful with humanoid robotics.

Robotics overall, for sure.