From Capital-Light to Capital-Hungry: Can Hyperscalers Sustain the AI Infrastructure Arms Race?
"AI is probably the most important thing humanity has ever worked on... more profound than electricity or fire.” — Google CEO Sundar Pichai
Caveat: I am not a technologist or a Wall Street analyst covering the hyperscalers. I’m simply an observer, curious to explore one question: will the once “capital-light” hyperscalers remain great businesses if they must spend hundreds of billions on AI infrastructure?
TL;DR
This is a long post, so if you are pressed for time, here's a TL:DR version:
AI is capital-intensive: Hyperscalers face unprecedented infrastructure costs, projected at $5.2–$7.9T by 2030.
Balance sheet risk: The strong and growing operating cash flows of Amazon, Google, Microsoft, Meta, and Oracle face are unlikely to meet their respective FY26-FY30 capital expenditure requirements, potentially requiring debt financing.
Historical parallels: The AI buildout resembles past infrastructure booms—US railroads in the 1800s and telecom in the 1990s—where over-investment and speculative optimism often led to painful corrections.
Winners may differ: In the telecom era, consumers and application builders (e.g., Google, Amazon, Cisco) reaped the biggest rewards, not the infrastructure providers themselves. The same dynamic may unfold with AI.
Debt and funding structure matter: Unlike 1990s telcos, today’s hyperscalers have relatively clean balance sheets, but massive private equity and VC funding in AI startups could create hidden systemic risks.
Open questions remain: Will hyperscalers end up with infrastructure-style returns? Could technological advances reduce the cost of compute, leading to overcapacity? Or will new entrants and application-layer companies capture most of the value?
Full Post
The Gen AI revolution is truly underway, and hyperscalers like Amazon (AMZN), Google (GOOG), Microsoft (MSFT) plus the wannabes Facebook (META) and Oracle (ORCL) are committing staggering amounts of money to build AI infrastructure. These companies will collectively spend $360B1 in FY25 alone (i.e. ~1.2% of the US GDP).
McKinsey research shows that data centers will need to spend around $5.2T worldwide from 2026 to 2030 to keep pace with the demand for compute power. In fact, McKinsey’s accelerated demand scenario assumes a $7.9T capital expenditure (capex).
The stakes are high. Overinvesting in data center infrastructure means stranded assets, while underinvesting means falling behind. Also, like all forecasts, the above needs to be read with the caveat that there are simply too many unknowns, making it almost impossible to come up with reasonably accurate estimates.
For the purposes of this article, I have assumed that McKinsey’s capex range of $5.2T2-$7.9T is a reasonable ball-park of the investment required (hereafter, I will refer to continued momentum as ‘base scenario’ and accelerated momentum as ‘aggressive scenario’3).
The hypothesis that I wanted to prove/disprove was - will this kind of spending dramatically alter the balance sheet and future cash flows of hyperscalers?
I have also assumed that the hyperscalers maintain their FY25 share of capex spend (e.g. AMZN contributed to 33% of the overall FY25 hyperscaler capex, and it will continue to contribute the same share until FY30). These assumptions and back-of-the-envelope calculations are not intended to calculate the valuation of these hyperscalers, but rather to understand the possibilities, risks and challenges.
There are plenty of holes in the above - the biggest one being that despite the astonishing $360B that hyperscalers are spending in capex in FY25 (which is a 142% increase compared to 2 years ago), it is nowhere close to the expected annual run rate of $1T from FY26-30, to meet McKinsey’s base scenario capex estimate of $5.2T by FY30.
The above means either there’s a currently a gross under-investment in AI infrastructure, or that McKinsey’s estimates are too ambitious. If it’s the former, it means that hyperscalers are leaving room open for new entrants to take market share away. Plus, there are no guarantees that the big spending hyperscalers will win the AI race.
There are other open questions too, like - could technological advancements dramatically reduce the cost of AI compute, leading to a glut of AI infrastructure? Will hyperscalers be eventually forced to move to a fixed pricing model instead of pay-per-use (similar to telcos), which means that their returns on capital look similar to an infrastructure business?
Leaning into History
The last time one sector required such a huge scale of capex was back in the late 19th century when the US built its railway network. Between 1870 and 1890, the U.S. spent about $8–10B ($400B in today’s dollars), which was ~2–3% of their annual GDP at the time.
However, the US telecom infrastructure buildout in the 1990s is more comparable to the current AI infrastructure wave, rather than the railways. The telecom era saw the convergence of digital technology, fiber optics, and the commercialisation of the internet. Today, we are seeing the convergence of cloud technology, GPUs, and the commercialisation of Gen AI LLMs.
So how did the telecom infrastructure boom play out in the late 1990s?
The Setup
In 1996 the US Telecommunications Act was passed, which removed regulatory barriers and encouraged competition. This catalyst led to an aggressive "get big fast" mentality among new firms, encouraging an acquisition frenzy and a race to lay fiber.
From 1996 to 2000, telecommunications carriers (‘telcos’) invested over $500 billion ($1 trillion in today’s dollars). The majority of this spending was financed with substantial debt.,with telcos routinely spending 20-22 % of their sales on replacing, refurbishing or upgrading plant and equipment, earning progressively lower returns on equity. During this period, capex across all telcos grew at an average annual rate of 28%, in stark contrast to a far more modest 10% annual increase in revenues.
The Frenzy, The Hubris and the Bust
This immense capex was driven by a confluence of financial and technical optimism. A key driver was the exaggerated belief that internet traffic was doubling every 3 to 12 months, hence investors believed that whatever amount of network capacity was built would be eventually filled.
A popular narrative was “The Internet changes everything. All the old rules needed to be torn up.” Plus the TMT sector stocks had tripled in the late 1990s, which made access to funding easier. This cheap capital, largely sourced from a speculative "white-hot" IPO market, was then poured into network infrastructure. As investment increased, financial optimism became more disconnected from the underlying business fundamentals.
This ‘get big, fast’ mindset led to a massive overcapacity, excessive debt and the eventual collapse of many telcos4. As per one estimate, 39 million miles of fiber-optic strands were laid, enough to circle the Earth over 1,500 times. Merrill Lynch & Co. estimated that in 2001, only 2.6 percent of that capacity was in use.
But what led to a ‘fiber-glut’? It was a confluence of the following:
Internet growth was not as fast as expected.
The growth of data traffic was not nearly fast enough to use all of the millions of miles of fiber-optic lines that were buried beneath streets and oceans.
The equipment used to send data over fiber optic cable as well as related technology improved dramatically5, hence the cost of transmitting voice and data fell rapidly.
The Winners
The greatest beneficiary was the consumer. The aggressive infrastructure build-out resulted in a vast, interconnected network of fiber and wireless technology. Competition drove down prices for services and forced incumbents to keep innovating, which fundamentally changed how the world communicated.
Companies that built products, tools, services on top of the newly built telecommunication and internet infrastructure won big - companies like Nokia, Cisco, Siemens, Google Search, Amazon.com, Microsoft were some of the early winners.
And of course, some of the winners were executives and insiders who sold their shares before the crash and reaped hundreds of millions of dollars by liquidating their stock holdings before the market turned.
But what happend to the big spending telcos6?
WorldCom and Global Crossing filed for bankruptcy due to their unsustainable levels of debt.
Qwest Communications was involved in an accounting scandal for inflating its revenues through deceptive business practices, such as booking equipment ‘swaps’ with other carriers as sales revenue. Its CEO was also involved in insider trading.
Lucent Technologies7, which was a key equipment supplier and a "darling" of the investment community, was dangerously co-dependent on the upstart network builders, to whom it extended credit. When the market collapsed, Lucent's stock price plummeted and it was forced to lay off a massive number of its employees
Parallels and Distinctions
Mind the Gap: Economist Carlota Perez explains in her 2002 book "Technological Revolutions and Financial Capital," that there is often a gap between the installation phase of a new technology, when infrastructure is built out, and the deployment phase, when the technology becomes widely adopted and integrated into society. This gap is typically marked by a burst of the speculative bubble that formed around the new technology during the installation phase.
h/t: Tabrez Syed’s Box Cars Substack
Predicting Winners: While technological innovation can transform society, it is incredibly hard to predict the winners. This was true in the telecom frenzy, the UK railway boom, the automobile boom and is equally true for today’s Gen AI revolution.
Global Scale: The key point of differentiation between the above technological revolutions v. the Gen AI buildout today is that the Gen AI infrastructure will have a global reach.
The Narrative: The narrative of today’s AI will change the world is similar to the 1990s narrative ‘the Internet changes everything’. The frenzy to build ‘fast, at any cost’ also rhymes with the 1990s euphoria.
Capital and Leverage: A key difference is that today’s AI hyperscalers have limited debt on their balance sheets (so far), compared to the telco hyperscalers who were heavily levered. Another nuance is in the 1990s, most telcos went public early to be able to raise capital. Gen AI upstarts like OpenAI, Perplexity, Anthropic, CoreWeave have been funded by Private Equity(PE)/Venture Capital (VC) players like Softbank, Blackstone, Silver Lake. From Jan 2023 till date, $134B has been raised, some of which is funded via debt on the PE/VC balance sheets.
The Capex Burden
Let’s look at each of the hyperscalers and how the upcoming capex wave impacts their business.
Amazon
Amazon is positioning itself as the leading AI infrastructure provider. In FY25, its capex will jump 43% to $118B, pushing capex-to-sales to 17%—levels reminiscent of telcos in the 1990s. Free cash flow is projected to drop by one-third.
Looking ahead, if Amazon shoulders $1.7T8–$2.6T of capex by FY30, its own operating cash flows9 won’t be enough, leaving a shortfall of $371B–$1.3T. Borrowing to fill that gap could reshape Amazon’s balance sheet and returns profile, pulling it closer to an “infrastructure utility” than a high-margin tech giant.
Alphabet
Google is framing AI as central to its future, with CEO Sundar Pichai declaring that “AI is positively impacting every part of the business.” That optimism comes with a heavy price tag: FY25 capex is expected to jump 62% to $85B, cutting free cash flow by about 9%. At its current valuation, Google trades at 44x forward FCF, up from 32x a year earlier.
Looking out to FY30, Google spend $1.2T–$1.9T in AI-related capex. Its FY25 operating cash flow is projected at $151B with ~14% CAGR, but even under those assumptions, Google faces a potential cumulative FCF shortfall of $58B–$758B by FY30. While smaller than Amazon’s gap, this still represents a significant strain that could force the company to balance growth ambitions with capital discipline.
Microsoft
Microsoft has cast AI as a once-in-a-generation productivity shift, embedding it deeply across its enterprise and cloud ecosystems. Its FY25 capex is set to climb 45% to $72B, which will push free cash flow down modestly by 3%. At its current valuation, Microsoft trades at 53x forward FCF, up from 46x in FY24, reflecting investor confidence in its AI roadmap.
Under McKinsey’s scenarios, Microsoft may need to invest $1T–$1.6T in AI infrastructure by FY30. With FY25 operating cash flow forecast at $136B and ~15% CAGR, Microsoft could either maintain a positive cumulative FCF of $156B, or, in the aggressive scenario, face a shortfall of $344B.
Meta
Meta is betting its future on what Mark Zuckerberg calls “personal superintelligence.” To fund that vision, the company is ramping FY25 capex by an extraordinary 93%, from $37B to $72B. The near-term consequence is steep: free cash flow is expected to fall by 34%. META trades at 54x forward FCF, up from 27x the prior year.
By FY30, Meta may need to commit $1T–$1.6T to AI infrastructure, an unprecedented scale for a previously capital light business. With FY25 operating cash flow estimated at $107B and ~17% CAGR, Meta still ends up facing a projected cumulative FCF shortfall of $118B–$718B. For a company that has already endured investor skepticism over metaverse spending, such an aggressive capital program raises the question: will Meta’s AI moonshot have returns commensurate with the risk?
Oracle
A similar analysis on ORCL shows that if it spends between $300B-$500B over the next 5 years, it will have a FCF shortfall between -$62B and $262B. ORCL already has a net debt of $83B, hence the above shortfall would add further leverage to its balance sheet.
Key Takeaways
AI is capital-intensive: Hyperscalers face unprecedented infrastructure costs, projected at $5.2–$7.9T by 2030.
Balance sheet risk: The strong and growing operating cash flows of Amazon, Google, Microsoft, Meta, and Oracle face are unlikely to meet their respective FY26-FY30 capital expenditure requirements, potentially requiring debt financing.
Historical parallels: The AI buildout resembles past infrastructure booms—US railroads in the 1800s and telecom in the 1990s—where over-investment and speculative optimism often led to painful corrections.
Winners may differ: In the telecom era, consumers and application builders (e.g., Google, Amazon, Cisco) reaped the biggest rewards, not the infrastructure providers themselves. The same dynamic may unfold with AI.
Debt and funding structure matter: Unlike 1990s telcos, today’s hyperscalers have relatively clean balance sheets, but massive private equity and VC funding in AI startups could create hidden systemic risks.
Open questions remain: Will hyperscalers end up with infrastructure-style returns? Could technological advances reduce the cost of compute, leading to overcapacity? Or will new entrants and application-layer companies capture most of the value?
Some of this $360B capex is not AI related, but exact numbers are not readily available.
Initially, I found McKinsey’s $5.2T estimates too large to comprehend , however they calculate it based on the assumption of a 3.5x increase in AI workload over a five year period which is not unreasonable.
It sounds crazy, but the $3.7B capex under the ‘Constrained Momentum’ scenario above is chump change for the hyperscalers, since this can be funded easily from their FY26-30 Operating Cash Flows. Hence I disregarded this scenario from my analysis.
By 2001, there were 19 national networks, most of which either went bankrupt or were consolidated into the 3 national networks that exist today.
The cost of sending and storing information declined sharply throughout the 1990s, driven by rapid improvement in technology such as digital GSM, network slicing, Voice over Internet Protocol (VoIP), and software defined networking reduced costs for businesses and consumers.
As a result, the cost of phone calls fell dramatically. In the mid-1980s, a local call cost ~$2 per minute and an international call ~$8 per minute. However, by the early 2000s the above mentioned technological improvements forced telcos to change their pricing model from pay-per-use to monthly plans and bundles. This change in the pricing model had a dramatic impact on the telcos' returns on capital ‘sunk’ into their physical infrastructure.
Level 3 Communications saw the value of its 16,000 mile network capacity plummet as wholesale prices for bandwidth crashed. However Level 3 adopted an aggressive acquisition strategy, using its depressed stock and debt to buy distressed assets from former rivals.
Lucent Technologies, which was a key equipment supplier and a "darling" of the investment community, with its market capitalization soaring to $258 billion at its peak. However, its business was dangerously co-dependent on the upstart network builders, to whom it extended credit. When the market collapsed, Lucent's stock price plummeted over 90%, its market value was virtually wiped out, and it was forced to lay off a massive number of its employees
In my calculations, I have taken the base and aggresive scenario capex figures as total capex for the entire business which is conservative since these will continue to have other capex needs for its existing businesses divisions.
OCFs have been extrapolated for FY26-FY30 by assuming that the hyperscaler will continue to grow its OCF at the same CAGR as FY21-FY25. Its a big assumption!