Comment by belter | Hacker News Reader

Comment by belter | Hacker News Reader

Parent Full thread

I run the numbers on hyperscaler AI capex and the math is not going to work out.

With these assumptions:

– Big 4 keep spending at current pace for 3 more years

– Returns only start showing after aprox 2 years

– Heavy competition with around 20% operating margin on AI and Cloud

– Use of 9% cost of capital

This is the current reality:

AWS aprox $142B/yr

Azure aprox $132B/yr

Google Cloud around $71B/yr

Combined its about $330B to $340B annual cloud revenue today

And lets says Global public cloud market of $700B total today.

To justify the current capex trajectory under those assumptions, by year 3 the big hyperscalers would need roughly $800B to $900B in new annual revenue just to earn a normal return on the capital being deployed.

That implies combined hyperscaler cloud and AI revenue going from: $330B today to $1.2T within 3 years :-))

In other words...Cloud would need to roughly do 4× in a very short window, and the incremental revenue alone would exceed the entire current global cloud market.

So for the investment wave to make financial sense, at least one of these must be true:

1 Cloud/AI spending globally explodes far beyond all prior forecasts

2 AI massively increases revenue/profit in ads, software, commerce and not just cloud

3 A winner takes all outcome where only 1 or 2 players earn real returns

4 Or a large share of this capex never earns an economic return and is defensive

People keep modeling this like normal cloud growth. But what we have is insanity

Azure revenue is growing at 39% year over year. If Microsoft can sustain this growth, in four years Azure will be ~3.73x its current size. This is of course very difficult, but you really don’t need a deus ex machina to hit 4x growth under your assumptions.

The issue in the late-90s was all the investment created a lot of real revenue for telecoms and other companies. Even though there were a lot of shenanigans with revenue, a lot of real money was spent on fiber and tech generally.

But the real money was investment that didn’t see a return for the investor. The investments needed to have higher final consumption (such as through better productivity or through displacing other costs) to pay back the investment.

> That implies combined hyperscaler cloud and AI revenue going from: $330B today to $1.2T within 3 years :-))

You’re ignoring the fact that gaming is going to the cloud.

That industry is bigger than Hollywood.

Desktop computers will invariably follow.

The RAM shortage will drive the transition.

For instance, my wife uses her personal laptop about four days a year.

People like that won’t be buying personal desktops or laptops, five years from now. The RAM shortage will drive a transition into thin clients.

I already see it with our kids. They use an iPhone, unless they need to type. Then they use an iPad with a BT keyboard.

Even if gaming goes to the cloud, how are they going to run the massive existing library of video games on the dedicated AI inference hardware that everyone is buying right now? Seems like that pivot would require even more spending.

And how are they going to get sub-5ms round trip latency into the average consumer’s home to avoid people continuing to see cloud gaming as a janky gimmick that feels bad to use?

What amount of the gaming industry do you think will go to AI providers and not game developers?

You think we'll replace gaming and desktop computers into the cloud in the timeline of the poster above (2-4 years?)

Just not realistic.

Cloud gaming is crap and any actual gamer will tell you that. The niche of gamers casual enough to not care about playing over network latency but serious enough to pay real money for cloud gaming is microscopic.

It's not 2023 anymore. Have you tried cloud gaming in 2026? I can barely tell it's connected to the cloud.

Yes, it's amazing because it's streaming directly from a computer in the room behind me. :)

And the increases in network speed are one of the last bastions of Moores Law.

> nd the increases in network speed are one of the last bastions of Moores Law.

Throughput has increased but latency hasn’t changed much

Latency hasn’t decreased substantially since the late 90s when I remember getting sub 50 ms ping in Quake III from my dorm room in college

I'm not surprised; you need a lot more servers and even so, there are a lot of places where something low ping times is difficult. While there is a lot of room for latency to go down, 1 lightmillisecond is ~300 km (~186 mi). This means that if a computer is 150 km away, 1 ms is the minimum ping allowed by physics, if I am talking directly to it.

By that yardstick, we've actually done very well in a lot of cases. :)

Speed of light doesn't adhere to Moore's law :) and it's made worse by the fact most everyone connects via WiFi these days and it alone adds a few ms more.

>The niche of gamers casual enough to not care about playing over network latency

In the saddest way possible, the niche of gamers are people playing on desktops with ethernet connections.

The majority of gamers are buying booster packs on mobile games.

Yes, but that majority doesn't need cloud gaming precisely because those games run just fine on their phone - there's no benefit in putting them in the cloud, that was supposed to be for fancy stuff where you need a beefy GPU for the eye candy.

The RAM shortage is extremely temporary. It’ll last as long as it takes for new capacity to come online. RAM shortages and price spikes have happened many times before.

Eventually China will catch up in EUV fabrication and flood the market with cheap silicon. When that happens a terabyte of RAM will cost what 128gb costs now.

That may be true, but all of this can be done today without the massive capex and without “AI”.

Reply on news.ycombinator.com