Great breakdown of the macro numbers, but the model misses a critical physical and accounting reality: the operational lag between capex deployment and placed-in-service status. Having worked on server room fit-outs, capital allocation does not equal an operational room overnight. Racking, cabling, testing, and grid interconnection bottlenecks add major delays before compute generates revenue. Under GAAP, depreciation does not start until the asset is placed in service. If fit-out or utility delays hold up a cluster for 12 months, you place hardware into service that is already a generation behind, burning part of its 5-year useful life (should be 3 years or less but that is an argument for another day) before making a dollar. That idle capital materially drags down the real IRR.
Honestly, there is no set timeline. It depends entirely on whether it is a greenfield build or expanding an existing powered shell, plus whatever local utility delays you hit. But even when the building and power are already there, racking, cabling, testing, and commissioning take real time before a single workload runs. It is rarely month zero, and that lag creates a capital drag the model misses.
Really interesting way to move the AI-bubble conversation from headlines to unit economics. It seems like the whole thesis ultimately rests on three things:
- Compute staying highly utilized
- Rental prices declining gradually
- Cloud revenue catching up with capex
If any one breaks badly, the returns change quickly. But at least this gives us the right variables to watch.
Have you tried modeling what if new contracts and new gw get repriced and sold at current Blackwell contracted rates and for a bull case repriced at Vera Rubin levels. My model shows when using these spot rates the hyperscalers can add around 900-1 trillion in cloud OCF. Now adding on non cloud OCF by 2030 the hyperscalers will have 2.3-2.5 trillion in OCF which will fund almost entirely the ai buildout.
The key question is what will happen to the demand for training capacity in the event of problems if the two largest laboratories decide or are forced to reduce training costs. In my opinion, we need to consider separately the capacity utilization for training and for inference
That's a wonderful article on a question widely asked at the moment. Are hyperscalers overspending on AI? Some figures start to show signs that AI spending was the right decision (MSFT and AMZN). I am not as deep into the exact split as you are but I believe ROIC or RONIC are pretty good metrics to measure success. That way, investors can get an idea of where the journey is going.
RONIC and recovered FCF. Both lagging indicators, but critical IMHO. We saw how Alphabet got punished for dipping into negative FCF, they recovered but the initial reaction was harsh. I saw that price divergence as a nice buy opportunity that has re-converged with my original thesis and KPIs.
What is a leading indicator to spot this accordingly? From a valuation perspective it is understandable that a stock breaks when FCF turns negative. Predictability is no longer a strength and uncertainty about FCF recovery sets in.
I'm still working on fleshing those out for my analysis... Something like forward demand coverage ratio. Committed demand requiring capacity over capacity coming online. These largely have to be inferred by analyzing backlog reports and then comparing to the CAPEX spend. Something like $100B of committed backlog with $100B in CAPEX spend. Totally fictional numbers, but that wouldn't be the best ratio... I would want 1.5, 2.0 or higher, just thinking this through right as I'm typing. I mean GOOG got spanked for -5.8B FCF, but it was foolish and short-sighted. That $5.8B last quarter is likely going to produce several times that in incremental revenue when the capacity it provides comes online. To my way of thinking if you have a current committed backlog of >$500B, and an incremental increase of $50B in just the last quarter, you would be more than justified going into the red by a measley $5.8B to fuel that future contracted growth. I've begun to outline my perspectives at : https://wintermute69.substack.com/ where I'll eventually be covering the whole AI stack top to bottom going forward.
I get where you are going and it makes sense, many use order backlog as a leading indicator but I think the main weakness of leading indicators is that they are not transparent. We don't know how that future demand converts into cash flow given the company structure their spending habits and cost structure. It is a challenge for sure...
Hi, thanks for the comment. Happy to see that someone actually went through our models.
For our IRR modeling purpose we have added back the retained DC value back to the company. Hence for the sake of simplicity in Revenue / Capex we have shown what it would take for company to generate a 21% IRR (implicitly it assumes all DC value is monetized to cash).
But you are right and a more detailed 30 year IRR model would show a different dynamics.
Great breakdown of the macro numbers, but the model misses a critical physical and accounting reality: the operational lag between capex deployment and placed-in-service status. Having worked on server room fit-outs, capital allocation does not equal an operational room overnight. Racking, cabling, testing, and grid interconnection bottlenecks add major delays before compute generates revenue. Under GAAP, depreciation does not start until the asset is placed in service. If fit-out or utility delays hold up a cluster for 12 months, you place hardware into service that is already a generation behind, burning part of its 5-year useful life (should be 3 years or less but that is an argument for another day) before making a dollar. That idle capital materially drags down the real IRR.
Very interesting point that I didn't model in. Do you have any sources (not discounting your expertise in anyways) which gives a timeline for this?
Honestly, there is no set timeline. It depends entirely on whether it is a greenfield build or expanding an existing powered shell, plus whatever local utility delays you hit. But even when the building and power are already there, racking, cabling, testing, and commissioning take real time before a single workload runs. It is rarely month zero, and that lag creates a capital drag the model misses.
Really interesting way to move the AI-bubble conversation from headlines to unit economics. It seems like the whole thesis ultimately rests on three things:
- Compute staying highly utilized
- Rental prices declining gradually
- Cloud revenue catching up with capex
If any one breaks badly, the returns change quickly. But at least this gives us the right variables to watch.
Have you tried modeling what if new contracts and new gw get repriced and sold at current Blackwell contracted rates and for a bull case repriced at Vera Rubin levels. My model shows when using these spot rates the hyperscalers can add around 900-1 trillion in cloud OCF. Now adding on non cloud OCF by 2030 the hyperscalers will have 2.3-2.5 trillion in OCF which will fund almost entirely the ai buildout.
Interesting. Will check it out.
We wanted to understand the base case scenario and why all hyperscalers are rushing to build as much capacity as possible.
Our guess is that they have done this math and figured it out already.
Oh, for sure they have done those math's. Those calculations are being done up and down the entire value chain.
Why isn't this going viral like the Citrini article?
Great report! Thanks for keeping it free.
Thank you :)
One day hopefully!
Good article with an interesting analysis. Thanks
Thank you.
The key question is what will happen to the demand for training capacity in the event of problems if the two largest laboratories decide or are forced to reduce training costs. In my opinion, we need to consider separately the capacity utilization for training and for inference
That's a wonderful article on a question widely asked at the moment. Are hyperscalers overspending on AI? Some figures start to show signs that AI spending was the right decision (MSFT and AMZN). I am not as deep into the exact split as you are but I believe ROIC or RONIC are pretty good metrics to measure success. That way, investors can get an idea of where the journey is going.
RONIC and recovered FCF. Both lagging indicators, but critical IMHO. We saw how Alphabet got punished for dipping into negative FCF, they recovered but the initial reaction was harsh. I saw that price divergence as a nice buy opportunity that has re-converged with my original thesis and KPIs.
What is a leading indicator to spot this accordingly? From a valuation perspective it is understandable that a stock breaks when FCF turns negative. Predictability is no longer a strength and uncertainty about FCF recovery sets in.
I'm still working on fleshing those out for my analysis... Something like forward demand coverage ratio. Committed demand requiring capacity over capacity coming online. These largely have to be inferred by analyzing backlog reports and then comparing to the CAPEX spend. Something like $100B of committed backlog with $100B in CAPEX spend. Totally fictional numbers, but that wouldn't be the best ratio... I would want 1.5, 2.0 or higher, just thinking this through right as I'm typing. I mean GOOG got spanked for -5.8B FCF, but it was foolish and short-sighted. That $5.8B last quarter is likely going to produce several times that in incremental revenue when the capacity it provides comes online. To my way of thinking if you have a current committed backlog of >$500B, and an incremental increase of $50B in just the last quarter, you would be more than justified going into the red by a measley $5.8B to fuel that future contracted growth. I've begun to outline my perspectives at : https://wintermute69.substack.com/ where I'll eventually be covering the whole AI stack top to bottom going forward.
I get where you are going and it makes sense, many use order backlog as a leading indicator but I think the main weakness of leading indicators is that they are not transparent. We don't know how that future demand converts into cash flow given the company structure their spending habits and cost structure. It is a challenge for sure...
Very interesting article!
I think there could be something wrong in the math of 'The Model'
"Bringing it all together, we can now model the cash generation of a $100B datacenter.
• At a 2:1 semis-to-shell ratio, $67B goes to chips and $33B to the shell.
Now assume $100B is invested in datacenters every year, and let’s see what steady-state revenue it will generate.
From year five onward, one cohort retires as a new one comes online, so revenue holds at ~$160B while capex stays at $100B."
After 5 years only the semis will retire not the shell, so I think Capex should slow down to $67B instead of stay at $100B
Hi, thanks for the comment. Happy to see that someone actually went through our models.
For our IRR modeling purpose we have added back the retained DC value back to the company. Hence for the sake of simplicity in Revenue / Capex we have shown what it would take for company to generate a 21% IRR (implicitly it assumes all DC value is monetized to cash).
But you are right and a more detailed 30 year IRR model would show a different dynamics.
Great article. Any comments on competition from Chinese AI labs and effect on compute demand? Curious to know about data centre economics in China
Hi Siew, we just posted an article on Open source. That should answer your question Chinese AI labs. Thanks