Extraordinary CAPEX spending has become a defining feature of the hyperscalers, but treating them as a homogeneous group misses the point. Not all capital is being deployed with the same discipline, focus, or likelihood of generating acceptable returns.
Meta is the clearest example of capital misallocation. Over recent years it has burned tens of billions annually on moonshot projects, most notably the Metaverse, a strategy so central it prompted a corporate rebrand. Framed generously, this spending looks speculative; more realistically, it shows little prospect of producing meaningful returns any time soon, if ever. From a shareholder perspective, it is difficult to justify this scale of investment relative to the visibility of outcomes.
Amazon’s CAPEX profile looks fundamentally different. Its spending is targeted and diversified across a collection of businesses with exceptional long-term potential. AWS remains the market leader in cloud infrastructure and is effectively a toll booth on the AI revolution. It is now generating roughly $170bn in annual revenue and growing at around 24%, implying a doubling in under three years and a potential quadrupling within five. Importantly, AWS operates at materially higher margins than Amazon’s retail business, driving steady expansion in group profitability.
As Andy Jassy noted on the most recent earnings call, demand is so strong that the primary constraint is not customers, but the speed at which Amazon can bring new capacity online. He frames this as investing ahead of what the filings show as a strong, contracted and pipeline demand environment, rather than a speculative spend. He highlights that AWS can still run at very high operating margins (around the mid‑30s) even while they are ramping this capex, implying that incremental scale on this base should be extremely profitable over time.
Beyond AWS, Amazon is building a portfolio of high-optionality businesses. Project Kuiper aims to compete with SpaceX in low-earth orbit satellites. Zoox is developing purpose-built autonomous vehicles rather than retrofitting consumer cars, as seen with Tesla and Waymo. Advertising continues to scale rapidly with very attractive margins. Amazon’s in-house semiconductor business is producing chips that management claims are around 40% more price-performant than NVIDIA’s offerings, and these chips are already being used by Anthropic to train the Claude large language model. Viewed this way, Amazon is less a single company and more a federation of businesses operating under one capital allocation framework.
Seen through that lens, Amazon’s projected $200bn CAPEX spend in 2026 appears far less alarming. Spread across multiple profitable and fast-growing subsidiaries, with clear demand signals and immediate utilization, the odds of achieving attractive incremental returns look materially higher than headline numbers alone would suggest.
A further distortion emerges when comparing CAPEX across large tech companies without adjusting for stock-based compensation. Many firms are effectively spending enormous sums each year simply to offset dilution. This is often buried in footnotes rather than highlighted in GAAP summaries. Alphabet’s most recent earnings report shows nearly $48bn spent on offsetting stock-based compensation and paying associated taxes upon vesting (don't look at the SBC numbers in the cash flow from operations section - that is a fiction - a market value at the time of issuance, not the real value at the time of vesting). The capital used for this purpose produces no return for outside shareholders; it is a direct transfer of wealth to insiders and employees.
Amazon stands out here as well. It does not repurchase shares to mask dilution, so its cost of issuing stock based comp is far lower than its peers and its capital allocation is more transparent. Once you adjust for these differences, Amazon’s investment profile looks meaningfully superior to many peers.
Stock-based compensation also distorts cash flow metrics. Most companies add it back as a non-cash expense, inflating cash flow from operations. As Charlie Munger famously remarked, if employee remuneration is not a cash expense, what exactly is it? Without significant adjustments, free cash flow figures become misleading, particularly when SBC is large and persistent.
In that context, the scary CAPEX forecasts for 2026 cannot be assessed through simplistic, side-by-side GAAP comparisons. The reality is far more nuanced, and the quality of capital allocation matters more than the absolute dollars being spent.
My view is that some hyperscalers will ultimately burn enormous amounts of capital and generate poor returns, with Meta being the most obvious candidate based on recent history. Others, particularly Amazon, may be entering a rare window where unprecedented scale can be achieved with genuinely accretive economics by riding the AI wave at exactly the right moment.
Jassy has repeatedly emphasized that AWS is deploying new AI and core infrastructure capacity as fast as it can be installed, with utilization effectively immediate. He has described this environment as an “extraordinary opportunity” for AWS to scale. Taken seriously, that framing suggests today’s CAPEX surge may prove far more productive than markets currently assume.
You're, of course, right that treating them as a homogeneous group misses nuance. That wasn't the point of the post. I tried to get a feel for what kind of profit growth we need to see for the basket of stocks to do well.
Funnily, I believe Meta may the one that has the most visible high ROIIC path in AI of all the stocks discussed here.
My concern relates to companies such as OpenAI, which has committed more than $500bn of CAPEX spending between now and the end of the decade, despite not being profitable and with no visible path to profitability.
My primary concern when I hear numbers from the likes of Alphabet, Amazon, Microsoft and Oracle is whether the order book on which they are basing investment decisions is robust, or whether they are relying on companies such as OpenAI which face an existential threat if they are unable to find a viable means of commercializing their technology.
In other words, is this spending based on real demand, or it is a house of cards that will implode? This is the billion dollar question.
You should be alarming instead for Amazon because their operating income is decelerating a lot from 2023 while spending much more that means they are deploying capital at much lower RONIC…..
Amazon's primary engine for profitability is still AWS. It has turnover of ~$170bn and is growing at 24% which is almost unprecedented at that scale. If it continues that rate of growth, it will double in less than 3 years and quadruple in about 5 years. Given that it is committing to huge CAPEX based on visible demand, do you think it will continue to grow at these rates (or faster) for at least a few more years?
Rather than react to a headline number, I like to deploy second level thinking. What does a $200bn CAPEX committment really mean? What does it tell me about what management can see from the inside that isn't visible to external investors? Given its very disciplined track record and stable management, is it likely to be making the right decision? With these answers I build a picture in my mind of what the future looks like for Amazon. I like what I imagine.
ROIC is not measured with revenue growth but with operating income growth..
The hyperscalers are moving from assets light business at high capex business and their ROIC will be impacted negatively and if it will be lower than their cost of capital they will start to destroy value..the long term winner in this era will be Apple not in terms of innovation but in terms of intelligent capital allocation and end user AI utilisation.
The dangerous problem that I see instead in the others is that they are investing in something that could become obsolete in less than 3 years and for that reason I would check the years taken to amortise such capex because it could become a real disaster from a financial point of view in next years.
We will see that in 2-3 years. My portfolio is exposed only to some industrials and consumer staples from end of 2025 because I do not know how it will end up but I have a feeling that it could be very bad or just a little good for that reason tech and financial sectors are my out of favour sector for 2026 and 2027.out of
Yes, they are moving from capital light to capital intensive - at least in the short term. Capital intensity may negatively impact returns on invested capital because the denominator is expanding, but what happens if the numerator expands even more? If demand ramps up, returns may increase exponentially.
I am not downplaying your concerns. They are valid. I share them too.
AI is a general purpose technology (GPT), but unlike the railroads or cable infrastructure, Moore's law means that infrastructure has a shorter shelf life and more rapid depreciation. This is an issue which is being discussed across the investment community.
But if a data center is a toll booth business, a picks and shovels operation, then the depreciation cost will ultimately be born by the user not the provider. So while the unit economics need to be carefully managed, it may not be as much of a problem as many envisage.
Add to that the continuing momentum with AWS custom chips, which will reduce dependence on overpriced NVIDIA silicon and reduce costs further:
• Trainium and Graviton now have a combined annual revenue run rate of over $10 billion and growing at a triple digit percentage year-over-year.
• Trainium2 is fully subscribed with 1.4 million chips landed, and powers the majority of inference on Bedrock, a service used by 100,000+ companies.
• Trainium2 powers Project Rainier, the world’s largest operational AI compute cluster with 500,000+ Trainium2 chips, which Anthropic is using to train its industry-leading AI model, Claude.
• Trainium3 is now delivering production workloads and seeing strong demand, with nearly all Trainium3 supply of chips expected to be committed by mid-2026.
• Trainium4 is expected to start delivering in 2027, with 6 times the FP4 compute performance, 4 times more memory bandwidth, and 2 times more high memory bandwidth capacity than Trainium3.
• Introduced Graviton5, AWS’s most powerful and advanced CPU for a broad set of cloud workloads. Used by over 90% of the top 1,000 AWS customers, Graviton is up to 40% more price-performant than leading x86 processors, and enables applications to run faster, reduce costs, and meet sustainability goals.
Amazon has a history of turning a cost line item on its income statement into a revenue line item. It is how AWS came into being in the first place. The Amazon chips may in time follow a similar path.
Alphabet, similarly, is seeing huge success with its TPUs.
You seem bullish on Apple, but it hasn't innovated since Steve Jobs and Jony Ives left the company. It has gone from being the most innovative company on the planet, to a rent extractor. It is disliked by almost everyone in tech for that reason. It has missed the AI revolution. It relies on the eco-system built by Jobs around the iPhone and Mac computer. I am less bullish on Apple under Tim Cook. I like to see evolution. I don't see that at Apple.
Just my view. Very different from yours. That's what makes a market.
I am not bullish on Apple but I see it as the one that could only benefit from AI instead of losing from it in terms of profitability and balance sheet stretch.
Great write up to put this in perspective. A few thoughts, bear with my amateur stock analysis ability.
Assuming a some of this investment is in hardware, I imagine maintenance and replacement from depreciation will eat a decent chunk of the future cash flow, hurting the ability to turn that cash flow into further reinvestments or earnings.
I did a back of the envelope summation of the revenue in the software ETF 'IGV' and got a figure of about $780B when excluding Microsoft. Assuming a 100% cannibalization, your 15% ROIIC figure already exceeds that entire industries revenue. Since the presumed purpose would be to drastically reduce B2B software expenditures it is difficult to see where this cash flow will come from. Of course other industries could be massively disrupted and the TAM could drastically expand, but still a point of note.
Lastly, you scared the hell out of me with your retweet of Matt Shumer's now viral AI warning on X today. But I wonder if even in spite of those warnings, the market is still massively overvaluing AI affiliated stocks. The recent run in energy stocks seems to be sniffing out something that most are overlooking. I think Matt can be correct and AI can be a poor investment at these prices. I would love to back into the expected return the market is pricing AI stocks at to justify these prices, but I have little confidence in my ability to accurately build a 3 stage FCF model. I guess I could just ask AI to do it.
Lots of great points Zach. Thanks for sharing. The ETF reference in particular to check if the figures pass a common sense test was particularly interesting.
Good Analysis but there are some remarks to improve it:
1)The growth to calculate ROIIC is not the one related to revenue but to operating income or better NOPLAT= EBITA (1- tax rate)
The growth in operating income is much lower because they are overspending too
In OpeX
2) ROIIC is related to incremental NOPLAT
3)R&D cannot be added to capex, you need to capitalize it, let’s say 5 years,
R&D_capitalized - R&D_ammortization= R&D_Capex
I would be happy to see your revised results, even if I already know that this huge capex will produce less profitability in the sector…ROIC will be lower for these hyper scaler based on
Extraordinary CAPEX spending has become a defining feature of the hyperscalers, but treating them as a homogeneous group misses the point. Not all capital is being deployed with the same discipline, focus, or likelihood of generating acceptable returns.
Meta is the clearest example of capital misallocation. Over recent years it has burned tens of billions annually on moonshot projects, most notably the Metaverse, a strategy so central it prompted a corporate rebrand. Framed generously, this spending looks speculative; more realistically, it shows little prospect of producing meaningful returns any time soon, if ever. From a shareholder perspective, it is difficult to justify this scale of investment relative to the visibility of outcomes.
Amazon’s CAPEX profile looks fundamentally different. Its spending is targeted and diversified across a collection of businesses with exceptional long-term potential. AWS remains the market leader in cloud infrastructure and is effectively a toll booth on the AI revolution. It is now generating roughly $170bn in annual revenue and growing at around 24%, implying a doubling in under three years and a potential quadrupling within five. Importantly, AWS operates at materially higher margins than Amazon’s retail business, driving steady expansion in group profitability.
As Andy Jassy noted on the most recent earnings call, demand is so strong that the primary constraint is not customers, but the speed at which Amazon can bring new capacity online. He frames this as investing ahead of what the filings show as a strong, contracted and pipeline demand environment, rather than a speculative spend. He highlights that AWS can still run at very high operating margins (around the mid‑30s) even while they are ramping this capex, implying that incremental scale on this base should be extremely profitable over time.
Beyond AWS, Amazon is building a portfolio of high-optionality businesses. Project Kuiper aims to compete with SpaceX in low-earth orbit satellites. Zoox is developing purpose-built autonomous vehicles rather than retrofitting consumer cars, as seen with Tesla and Waymo. Advertising continues to scale rapidly with very attractive margins. Amazon’s in-house semiconductor business is producing chips that management claims are around 40% more price-performant than NVIDIA’s offerings, and these chips are already being used by Anthropic to train the Claude large language model. Viewed this way, Amazon is less a single company and more a federation of businesses operating under one capital allocation framework.
Seen through that lens, Amazon’s projected $200bn CAPEX spend in 2026 appears far less alarming. Spread across multiple profitable and fast-growing subsidiaries, with clear demand signals and immediate utilization, the odds of achieving attractive incremental returns look materially higher than headline numbers alone would suggest.
A further distortion emerges when comparing CAPEX across large tech companies without adjusting for stock-based compensation. Many firms are effectively spending enormous sums each year simply to offset dilution. This is often buried in footnotes rather than highlighted in GAAP summaries. Alphabet’s most recent earnings report shows nearly $48bn spent on offsetting stock-based compensation and paying associated taxes upon vesting (don't look at the SBC numbers in the cash flow from operations section - that is a fiction - a market value at the time of issuance, not the real value at the time of vesting). The capital used for this purpose produces no return for outside shareholders; it is a direct transfer of wealth to insiders and employees.
Amazon stands out here as well. It does not repurchase shares to mask dilution, so its cost of issuing stock based comp is far lower than its peers and its capital allocation is more transparent. Once you adjust for these differences, Amazon’s investment profile looks meaningfully superior to many peers.
Stock-based compensation also distorts cash flow metrics. Most companies add it back as a non-cash expense, inflating cash flow from operations. As Charlie Munger famously remarked, if employee remuneration is not a cash expense, what exactly is it? Without significant adjustments, free cash flow figures become misleading, particularly when SBC is large and persistent.
In that context, the scary CAPEX forecasts for 2026 cannot be assessed through simplistic, side-by-side GAAP comparisons. The reality is far more nuanced, and the quality of capital allocation matters more than the absolute dollars being spent.
My view is that some hyperscalers will ultimately burn enormous amounts of capital and generate poor returns, with Meta being the most obvious candidate based on recent history. Others, particularly Amazon, may be entering a rare window where unprecedented scale can be achieved with genuinely accretive economics by riding the AI wave at exactly the right moment.
Jassy has repeatedly emphasized that AWS is deploying new AI and core infrastructure capacity as fast as it can be installed, with utilization effectively immediate. He has described this environment as an “extraordinary opportunity” for AWS to scale. Taken seriously, that framing suggests today’s CAPEX surge may prove far more productive than markets currently assume.
Food for thought.
You're, of course, right that treating them as a homogeneous group misses nuance. That wasn't the point of the post. I tried to get a feel for what kind of profit growth we need to see for the basket of stocks to do well.
Funnily, I believe Meta may the one that has the most visible high ROIIC path in AI of all the stocks discussed here.
My concern relates to companies such as OpenAI, which has committed more than $500bn of CAPEX spending between now and the end of the decade, despite not being profitable and with no visible path to profitability.
My primary concern when I hear numbers from the likes of Alphabet, Amazon, Microsoft and Oracle is whether the order book on which they are basing investment decisions is robust, or whether they are relying on companies such as OpenAI which face an existential threat if they are unable to find a viable means of commercializing their technology.
In other words, is this spending based on real demand, or it is a house of cards that will implode? This is the billion dollar question.
You should be alarming instead for Amazon because their operating income is decelerating a lot from 2023 while spending much more that means they are deploying capital at much lower RONIC…..
Amazon's primary engine for profitability is still AWS. It has turnover of ~$170bn and is growing at 24% which is almost unprecedented at that scale. If it continues that rate of growth, it will double in less than 3 years and quadruple in about 5 years. Given that it is committing to huge CAPEX based on visible demand, do you think it will continue to grow at these rates (or faster) for at least a few more years?
Rather than react to a headline number, I like to deploy second level thinking. What does a $200bn CAPEX committment really mean? What does it tell me about what management can see from the inside that isn't visible to external investors? Given its very disciplined track record and stable management, is it likely to be making the right decision? With these answers I build a picture in my mind of what the future looks like for Amazon. I like what I imagine.
ROIC is not measured with revenue growth but with operating income growth..
The hyperscalers are moving from assets light business at high capex business and their ROIC will be impacted negatively and if it will be lower than their cost of capital they will start to destroy value..the long term winner in this era will be Apple not in terms of innovation but in terms of intelligent capital allocation and end user AI utilisation.
The dangerous problem that I see instead in the others is that they are investing in something that could become obsolete in less than 3 years and for that reason I would check the years taken to amortise such capex because it could become a real disaster from a financial point of view in next years.
We will see that in 2-3 years. My portfolio is exposed only to some industrials and consumer staples from end of 2025 because I do not know how it will end up but I have a feeling that it could be very bad or just a little good for that reason tech and financial sectors are my out of favour sector for 2026 and 2027.out of
Yes, they are moving from capital light to capital intensive - at least in the short term. Capital intensity may negatively impact returns on invested capital because the denominator is expanding, but what happens if the numerator expands even more? If demand ramps up, returns may increase exponentially.
I am not downplaying your concerns. They are valid. I share them too.
AI is a general purpose technology (GPT), but unlike the railroads or cable infrastructure, Moore's law means that infrastructure has a shorter shelf life and more rapid depreciation. This is an issue which is being discussed across the investment community.
But if a data center is a toll booth business, a picks and shovels operation, then the depreciation cost will ultimately be born by the user not the provider. So while the unit economics need to be carefully managed, it may not be as much of a problem as many envisage.
Add to that the continuing momentum with AWS custom chips, which will reduce dependence on overpriced NVIDIA silicon and reduce costs further:
• Trainium and Graviton now have a combined annual revenue run rate of over $10 billion and growing at a triple digit percentage year-over-year.
• Trainium2 is fully subscribed with 1.4 million chips landed, and powers the majority of inference on Bedrock, a service used by 100,000+ companies.
• Trainium2 powers Project Rainier, the world’s largest operational AI compute cluster with 500,000+ Trainium2 chips, which Anthropic is using to train its industry-leading AI model, Claude.
• Trainium3 is now delivering production workloads and seeing strong demand, with nearly all Trainium3 supply of chips expected to be committed by mid-2026.
• Trainium4 is expected to start delivering in 2027, with 6 times the FP4 compute performance, 4 times more memory bandwidth, and 2 times more high memory bandwidth capacity than Trainium3.
• Introduced Graviton5, AWS’s most powerful and advanced CPU for a broad set of cloud workloads. Used by over 90% of the top 1,000 AWS customers, Graviton is up to 40% more price-performant than leading x86 processors, and enables applications to run faster, reduce costs, and meet sustainability goals.
Amazon has a history of turning a cost line item on its income statement into a revenue line item. It is how AWS came into being in the first place. The Amazon chips may in time follow a similar path.
Alphabet, similarly, is seeing huge success with its TPUs.
You seem bullish on Apple, but it hasn't innovated since Steve Jobs and Jony Ives left the company. It has gone from being the most innovative company on the planet, to a rent extractor. It is disliked by almost everyone in tech for that reason. It has missed the AI revolution. It relies on the eco-system built by Jobs around the iPhone and Mac computer. I am less bullish on Apple under Tim Cook. I like to see evolution. I don't see that at Apple.
Just my view. Very different from yours. That's what makes a market.
I am not bullish on Apple but I see it as the one that could only benefit from AI instead of losing from it in terms of profitability and balance sheet stretch.
Great write up to put this in perspective. A few thoughts, bear with my amateur stock analysis ability.
Assuming a some of this investment is in hardware, I imagine maintenance and replacement from depreciation will eat a decent chunk of the future cash flow, hurting the ability to turn that cash flow into further reinvestments or earnings.
I did a back of the envelope summation of the revenue in the software ETF 'IGV' and got a figure of about $780B when excluding Microsoft. Assuming a 100% cannibalization, your 15% ROIIC figure already exceeds that entire industries revenue. Since the presumed purpose would be to drastically reduce B2B software expenditures it is difficult to see where this cash flow will come from. Of course other industries could be massively disrupted and the TAM could drastically expand, but still a point of note.
Lastly, you scared the hell out of me with your retweet of Matt Shumer's now viral AI warning on X today. But I wonder if even in spite of those warnings, the market is still massively overvaluing AI affiliated stocks. The recent run in energy stocks seems to be sniffing out something that most are overlooking. I think Matt can be correct and AI can be a poor investment at these prices. I would love to back into the expected return the market is pricing AI stocks at to justify these prices, but I have little confidence in my ability to accurately build a 3 stage FCF model. I guess I could just ask AI to do it.
Lots of great points Zach. Thanks for sharing. The ETF reference in particular to check if the figures pass a common sense test was particularly interesting.
Good Analysis but there are some remarks to improve it:
1)The growth to calculate ROIIC is not the one related to revenue but to operating income or better NOPLAT= EBITA (1- tax rate)
The growth in operating income is much lower because they are overspending too
In OpeX
2) ROIIC is related to incremental NOPLAT
3)R&D cannot be added to capex, you need to capitalize it, let’s say 5 years,
R&D_capitalized - R&D_ammortization= R&D_Capex
I would be happy to see your revised results, even if I already know that this huge capex will produce less profitability in the sector…ROIC will be lower for these hyper scaler based on
EBITA_growth = ROIC x reinvestment rate
With
reinvestment rate =
NOPLAT / ( net capex - Change in non cash WC)
Thanks for your Analysis
The future is too uncertain for that level of accuracy and complexity
Do you think using CFROI invented by Madden makes the analysis easier without knowing special parameter related to each company?