Monetizing AI search, continued innovation and expansion of Gemini, rising depreciation but management pointing to expanding operating margin nonetheless.
This analisys of Alphabet's earnings really made me think about the future of AI infrastructure. Given the massive Capex step-up for 2026, what are your thoughts on the most significant long-term ROI drivers for that investment?
Further to below - you also have to factor in the related cloud revenue that comes with Ai workloads. Some estimates show up to $2-$3 of additional software and cloud services (storage, data transmission, applications etc) for every $1 of compute revenue. That further boosts returns for the overall data center and helps the hyperscaler make a high return on the 40%-50% of DC capex thats not silicon related.
I'm actually working on a research piece about that now. The interesting thing is that I think we're starting to see data points that are going to convince investors that the returns are there to be value accretive. Those deploying ther own custom silicon at scale (Google and AMZN ) are those that are showing EBIT margin increases in their cloud divisions. Those like Microsoft and Meta who are (or have been) traditionally locked into Nvidia (premium silicon) are the ones seeing expense growth exceed revenue growth.
The deployment of your own customer silicon in training data centers for example costs about $4k-$6k per chip (internal manufacturing cost for ASICs) vs $28k for Nvidia so its about 80% + cheaper. However AWS for example charges about $2.50 per hour for GPU rental of its trainium chips vs $4.10 for Nvidia . So they still get 60% of the revenue off 20% or less of the compute capex. That massively boost margins and ROIC.
Returns differ a lot in inference workload facilities vs training because of lower ultilization. Early days on calcs yet but I'm seeing north of 20% ROIC on model training facilities vs maybe 15% on inference. Those like Google who are more vertically integrated are likely getting higher ROICs.
Appreciate the detailed breakdown! The depreciation spike warnings are worrying tho - feels like classic mgmt trying to manage expectations downward while actual operations r crushing it. I picked up some shares after reading ur structural cost advantage piece and honestly its been a great entry point.
I see it a bit like a warehouse logistics model. Even with strong demand it takes time to fill a new facility after construction. That means rapid expansion (when you accelerate the rate of new DCs opening) comes at an initial margin dilution but each individual facility (given comments of demand > supply) fills quickly and margins are likely 35%-40% at the EBIT level (eg AMZNs AWS margin expansion this quarter). The hyperscalers are really stepping up the rate of investment so that makes margins either reduce temporarily or flatline (or plummet in the case of META). Once that plateaus we'll see margins rebound.
Related to this I think theres data points building supporting the ROIC on DC investments. I'm researching a piece now on this. Of course, those who have their own custom silicon and offer full stack cloud services will achieve much stronger returns than those building DC assets for their own AI alone workload who are still overly reliant on Nvidia.
The Cloud backlog jump from $106B to $240B is staggering, especially when paired with that 78% drop in Gemini serving costs. Management's supply constraint framing through 2026 despite doubling capex is interesting though. Either demand is even more explosive than anyone thought, or there's inefficiency in deployment timing that hasn't been addressed yet.
This analisys of Alphabet's earnings really made me think about the future of AI infrastructure. Given the massive Capex step-up for 2026, what are your thoughts on the most significant long-term ROI drivers for that investment?
Further to below - you also have to factor in the related cloud revenue that comes with Ai workloads. Some estimates show up to $2-$3 of additional software and cloud services (storage, data transmission, applications etc) for every $1 of compute revenue. That further boosts returns for the overall data center and helps the hyperscaler make a high return on the 40%-50% of DC capex thats not silicon related.
I'm actually working on a research piece about that now. The interesting thing is that I think we're starting to see data points that are going to convince investors that the returns are there to be value accretive. Those deploying ther own custom silicon at scale (Google and AMZN ) are those that are showing EBIT margin increases in their cloud divisions. Those like Microsoft and Meta who are (or have been) traditionally locked into Nvidia (premium silicon) are the ones seeing expense growth exceed revenue growth.
The deployment of your own customer silicon in training data centers for example costs about $4k-$6k per chip (internal manufacturing cost for ASICs) vs $28k for Nvidia so its about 80% + cheaper. However AWS for example charges about $2.50 per hour for GPU rental of its trainium chips vs $4.10 for Nvidia . So they still get 60% of the revenue off 20% or less of the compute capex. That massively boost margins and ROIC.
Returns differ a lot in inference workload facilities vs training because of lower ultilization. Early days on calcs yet but I'm seeing north of 20% ROIC on model training facilities vs maybe 15% on inference. Those like Google who are more vertically integrated are likely getting higher ROICs.
Appreciate the detailed breakdown! The depreciation spike warnings are worrying tho - feels like classic mgmt trying to manage expectations downward while actual operations r crushing it. I picked up some shares after reading ur structural cost advantage piece and honestly its been a great entry point.
I see it a bit like a warehouse logistics model. Even with strong demand it takes time to fill a new facility after construction. That means rapid expansion (when you accelerate the rate of new DCs opening) comes at an initial margin dilution but each individual facility (given comments of demand > supply) fills quickly and margins are likely 35%-40% at the EBIT level (eg AMZNs AWS margin expansion this quarter). The hyperscalers are really stepping up the rate of investment so that makes margins either reduce temporarily or flatline (or plummet in the case of META). Once that plateaus we'll see margins rebound.
Related to this I think theres data points building supporting the ROIC on DC investments. I'm researching a piece now on this. Of course, those who have their own custom silicon and offer full stack cloud services will achieve much stronger returns than those building DC assets for their own AI alone workload who are still overly reliant on Nvidia.
The Cloud backlog jump from $106B to $240B is staggering, especially when paired with that 78% drop in Gemini serving costs. Management's supply constraint framing through 2026 despite doubling capex is interesting though. Either demand is even more explosive than anyone thought, or there's inefficiency in deployment timing that hasn't been addressed yet.