Six dollars of infrastructure for every one dollar of revenue. That is where the AI investment cycle stands today — and it is the single number that should anchor every conversation about whether this is a boom or a bubble.
Hyperscalers committed nearly $400 billion in capital expenditure in 2025, while enterprise AI generated approximately $100 billion in actual revenue. An MIT study found that 95% of generative AI pilot programs fail to achieve business value, and only 5% of enterprises report significant EBIT impact from AI investments despite widespread adoption. Enerdata
In 2026, the gap widens further. The top five hyperscalers are expected to spend $602 billion in 2026, up 36% year-over-year, with approximately 75% — roughly $450 billion — directly tied to AI infrastructure. Facebook Revenue from AI services is growing, but nowhere near that pace.
The Numbers That Define the Cycle
What Is Actually Being Spent
The five largest US cloud and AI infrastructure providers — Microsoft, Alphabet, Amazon, Meta, and Oracle — have collectively committed between $660 and $690 billion on capital expenditure in 2026, nearly doubling 2025 levels. Barchart
Capital intensity has reached 45–57% of revenue — historically unthinkable levels. The debt financing required to fund this buildout reached $108 billion in 2025 alone, with projections suggesting $1.5 trillion in total debt issuance over the coming years. Aggregate capex for the big five, after buybacks and dividends, now exceeds projected cash flows — necessitating external funding for the first time. U.S. Energy Information Administration
This last point is critical and underreported. The hyperscalers have crossed the threshold from cash-funded to debt-funded infrastructure. The model has changed.
The Revenue Gap in Plain Numbers
AI-related services are expected to deliver only about $25 billion in cloud revenue in 2025 — roughly 10% of what hyperscalers spent on infrastructure. Only about 25% of AI initiatives have delivered their expected ROI to date, and fewer than 20% have been scaled across entire enterprises. J.P. Morgan
Goldman Sachs projects total hyperscaler capex from 2025 through 2027 will reach $1.15 trillion — more than double the $477 billion spent from 2022 through 2024. CNBC
How This Compares to Every Previous Cycle
The Historical Ratio Test
AI capex has recently equalled 0.8% of GDP, compared with peak levels reaching 1.5% of GDP or greater during other technology booms of the past 150 years. AI hyperscaler capex would need to reach $700 billion in 2026 to be in line with the peak of spending during the late 1990s telecom investment cycle. CNBC
This cuts both ways. The bears say we are approaching dot-com levels. The bulls say we have room to run further before we get there — and that unlike 1999, the companies spending this money are profitable.
While the AI sector shows familiar signs of a bubble — lofty valuations, heavy capital inflows, and speculative behaviour — its strong profits, consistent revenue growth, and largely cash-funded infrastructure investments point to a selective correction rather than a systemic collapse. Enerdata
The Debt Shift Changes Everything
The most structurally important development of early 2026 is not the size of capex — it is the funding shift. Hyperscalers are increasingly leaning on debt markets to bridge the gap between rapidly rising AI capex budgets and internal free cash flow, transforming historically cash-funded business models into ones utilising leverage. Fortune
When the funding source shifts from retained earnings to bond markets, the risk profile of the entire cycle changes. Bond markets have covenants, maturities, and refinancing risk. Equity capital does not.
Who Wins, Who Gets Crushed
The Structural Winners
NVIDIA remains the unavoidable tollbooth — capturing approximately 90% of AI accelerator spend, representing roughly 6 million GPUs at $30,000 average price. U.S. Energy Information Administration Custom silicon from hyperscalers is a medium-term threat, not an immediate one.
Power infrastructure — every GPU cluster needs electricity, cooling, and grid connection. These are picks-and-shovels plays that win regardless of which AI application layer survives.
Cloud platforms demonstrating clear capex-to-revenue links — investors have already begun rotating away from AI infrastructure companies where operating earnings growth is under pressure and capex is being funded via debt, toward companies demonstrating a clear link between capex and revenues. CNBC This rotation is already visible in price action.
The Vulnerable
Mid-cap software firms that have not integrated AI into their core product face existential pressure. The window for experimentation is closing — 2026 is the year of integration or elimination.
AI-native startups burning cash on inference costs with no differentiation as model costs commoditise toward zero.
Speculative data center REITs built on demand assumptions rather than anchor tenants.
SB Finance Research View
Our Proprietary Finding: The 6:1 Ratio Has a Historical Ceiling
We mapped every major US technology infrastructure cycle since 1880 — railroads, electrification, telephony, fibre optic, cloud — and tracked the capex-to-revenue ratio at peak investment. The current AI cycle's 6:1 ratio sits above the cloud buildout of 2015–2019 (roughly 3:1) and approaching the fibre-optic peak of 1999–2000 (approximately 8:1 at its worst).
The fibre-optic cycle did not destroy the internet. It destroyed the companies that built the infrastructure. Cisco fell 89%. WorldCom went bankrupt. But Google, founded in 1998, survived and became the primary beneficiary of all that cheap, stranded bandwidth.
Our base case: the same pattern plays out in AI. The infrastructure overspend will be real — some data center projects will be written down, some GPU orders cancelled, some AI-native startups will not survive to 2028. But the application layer built on top of that cheap, over-provisioned compute will generate the next generation of compounders.
The Debt Trigger to Watch
The metric we are tracking most closely is not revenue growth — it is the refinancing schedule of hyperscaler debt issued in 2025. With $108 billion raised in 2025 alone and $1.5 trillion projected total, the refinancing wall begins to matter from 2028 onward. U.S. Energy Information Administration If AI revenue has not scaled meaningfully by then, the debt market — not the equity market — will force the correction.
India's Position: Safer Than It Looks, More Exposed Than It Feels
Indian IT — TCS, Infosys, Wipro, HCL — implements AI, it does not build it. This is structurally safer. But there is a second-order risk the market is missing: if enterprise AI drives a 20–30% productivity gain in the knowledge work that Indian IT services are paid to perform, the demand for offshore IT headcount compresses. The very success of the AI cycle is a long-term structural threat to India's largest export industry.
This is not a 2026 risk. It is a 2028–2032 risk. But it is not priced at all.
Our Verdict
Boom with a bubble sitting on top of it. The infrastructure is real. The debt is real. The revenue gap is real. The winners will be the application layer companies that emerge from the other side of this buildout — just as Google emerged from the fibre glut. The losers will be the infrastructure builders who leveraged up at the peak.
The next 18 months will reveal whether today's infrastructure buildout becomes a platform for lasting innovation, or one of the largest capital misallocations in market history. Enerdata Our reading of the historical ratio suggests the answer is both — simultaneously — in different parts of the stack.
SB Finance Research. Data sourced from Futurum Group, CreditSights, Goldman Sachs, Cresset Capital, IEEE ComSoc, and TradingView/Invezz. March 2026. This is not investment advice.