In the span of three years, artificial intelligence has gone from a research curiosity to the single most important force in global financial markets. Not because of what AI does — most people still use it to draft emails and summarise documents — but because of the staggering, almost incomprehensible amount of money being spent to build it.
The numbers have crossed into territory that has no modern precedent. And with every earnings call, a familiar question grows louder among fund managers, economists, and retail investors alike: Is this a rational, once-in-a-generation infrastructure supercycle — or the most expensive bubble ever inflated?
This article breaks down the anatomy of the AI capital expenditure (capex) boom, the bull and bear cases in plain language, the accounting controversy that spooked markets, the June 2026 sell-off, and — crucially for our readers — what it all means for the Indian investor.
Section 1: The Scale — Numbers That Break the Imagination
Let’s start with the headline figure that frames everything else.
The four largest US “hyperscalers” — Amazon, Microsoft, Alphabet (Google), and Meta — are collectively guiding toward roughly $725 billion in capital expenditure in 2026, up approximately 77% from ~$410 billion in 2025. The overwhelming majority of this is AI infrastructure. Analysts already project the figure will top $1 trillion in 2027.
To put this in perspective, Goldman Sachs now expects a combined $5.3 trillion of capex across the four largest hyperscalers from fiscal 2025 through 2030. The company-by-company breakdown for 2026 looks like this:
| Company | 2026 Capex Guidance | Primary Use |
|---|---|---|
| Amazon | ~$200 billion | AWS data centres, Trainium chips |
| Microsoft | ~$190 billion | Azure, OpenAI compute, MAIA silicon |
| Alphabet (Google) | $175–185 billion | TPUs, Google Cloud, data centres |
| Meta | $115–145 billion | GPU clusters, “superintelligence” |
The single biggest beneficiary sits one layer down the stack. NVIDIA’s Q4 data centre revenue hit $62.31 billion, up 75% year-over-year, with networking revenue up 263%. In October 2025, NVIDIA’s value grew beyond $5 trillion, rising higher than the GDP of every country except the US and China.
This is what a “supercycle” looks like: a self-reinforcing wave of spending on the physical scaffolding of a new technology — chips, servers, data centre shells, networking, and above all, power.
Section 2: Why Are They Spending So Aggressively?
Before judging whether this is madness, it helps to understand the logic. Four forces are driving the spend:
- Fear of being short on compute. Every hyperscaler CEO has concluded that under-investing in AI capacity is far more dangerous than over-investing. If demand materialises and you lack the compute, you lose the market permanently. As Microsoft warned investors, the company expects to remain capacity-constrained through at least 2026.
- Real, growing revenue. This isn’t purely speculative. Microsoft’s AI business surpassed an annual revenue run rate of $37 billion, up 123% year-over-year, and Google’s cloud revenue jumped 63% year over year to $20 billion.
- The productivity promise. The bet is that AI will eventually augment or automate a meaningful share of white-collar labour, creating trillions in economic value.
- Winner-takes-most dynamics. In platform markets, the leader captures disproportionate rewards, making the race existential.
The bull case, as Jefferies analyst Brent Thill bluntly put it, is that “the AI economy is healthy” and recent revenue growth justifies the outlays.
Section 3: The Bear Case — Four Cracks in the Foundation
Now the uncomfortable part. Skeptics — including some of the most respected names in finance — point to structural warning signs.
3.1 The Revenue-Investment Chasm
The clearest concern is that the money going in dwarfs the money coming out. Consider OpenAI, the demand engine at the centre of the boom:
OpenAI’s 2025 revenue stood at approximately $13 billion, while its capital expenditure commitment over eight years is $1.4 trillion. Its projected operating loss in 2028 alone is $74 billion.
More broadly, an MIT Media Lab report found that despite $30–40 billion in enterprise investment in generative AI, 95% of organisations reported zero measurable return. A February 2026 study by the National Bureau of Economic Research similarly found that despite 90% of firms reporting no impact of AI on workplace productivity, executives projected AI to increase productivity by 1.4% — an echo of the historical “productivity paradox.”
3.2 Circular Financing — The “Financial Ouroboros”
This is the concern that draws the sharpest comparisons to the dot-com bust. A tight web of investments and purchase commitments now links the biggest players, where money appears to flow in a circle:
- Nvidia committed up to $100 billion to OpenAI — money that will be largely spent purchasing Nvidia’s own products.
- Microsoft owns roughly 27% of OpenAI and is its primary cloud provider, generating Azure revenue that gets reinvested in Nvidia chips.
- OpenAI took a stake in AMD while AMD received billions in OpenAI orders; Nvidia holds a stake in CoreWeave and committed $6.3 billion to purchase CoreWeave’s unsold capacity — data centres stocked with Nvidia GPUs.
GMO analysts describe this arrangement as “reminiscent of the circular financing of the internet bubble.” The fear is that these deals create the appearance of commercial demand while masking a financial circularity that inflates all parties’ valuations at once.
The distinction that matters, as INSEAD framed it, is whether real end-user demand is being generated, or whether money is simply moving in circles — the difference between a “flywheel” and a “house of cards.”
3.3 The Depreciation Controversy
In late 2025, “Big Short” investor Michael Burry reignited a technical accounting debate and turned it into a market event. His claim: hyperscalers are inflating profits by stretching the assumed “useful life” of their AI hardware.
Burry’s argument is that AI hardware really lives 2–3 years but is being treated as if it lives 5–6 years. He estimates that from 2026–2028, depreciation will be understated by about $176 billion, causing hyperscalers to overstate profits by more than 20%.
The logic is straightforward: NVIDIA ships a materially better chip roughly every year, so top-tier chips become uneconomic for frontier training quickly. Yet the response is genuinely contested. Defenders note that older chips retain real value for less demanding “inference” work — CoreWeave cites data showing older Nvidia A100 chips retained 95% of their original price in expired contracts. The debate is not settled, and that uncertainty alone is a risk: Goldman’s own analysis found that shortening GPU useful life from five years to three would push cumulative depreciation from roughly $3 trillion to approximately $4 trillion between 2026 and 2031 — a $1 trillion swing driven by a single accounting assumption.
3.4 Market Concentration and Cash Burn
The final crack is systemic. In late 2025, 30% of the US S&P 500 was held up by the five largest companies — the greatest concentration in half a century — with valuations reportedly the most stretched since the dot-com bubble. And the cash cost is now visible: reaching the 2026 spending numbers means a big drop in free cash flow, with Amazon projected to turn free-cash-flow negative.
Section 4: The June 2026 Wake-Up Call
For much of 2025, markets forgave the spending. That patience is fraying. The first serious investor rebellion came when Meta raised its capex guidance in early 2026 — its shares fell over 9% in a single session, the first real revolt against the spending curve.
Then came a sharper shock. In late June 2026, the fragility of the semiconductor-heavy trade was exposed:
On 23 June 2026, the South Korean KOSPI plunged and halted trading to prevent a crash, with Samsung and SK Hynix losing 12% in a single morning. The Nasdaq sank 2.2% that afternoon, and the slump went global over the following days as investors turned away from AI-related tech stocks.
This matters because it demonstrates the transmission risk. When a boom is this concentrated, a wobble in one node — a stalled deal, a disappointing model, a chip-price spike — ripples across the entire market. As one analysis soberly noted, AI-related stocks have driven roughly 75% of S&P 500 returns since 2022, so any AI de-rating would hit the broader market disproportionately.
Section 5: Bubble or Not? A Balanced Verdict
So which is it? The honest answer is that both things can be true at once — the technology can be real and the near-term valuations can be excessive. That was precisely the pattern of the dot-com era: the internet was transformative, yet most of the companies built to exploit it went to zero first.
The most useful lens comes from a definition worth remembering:
A bubble exists not only when prices exceed current fundamentals, but when prices exceed what future fundamentals can realistically deliver.
There is also a genuine bull rebuttal to the circular-financing panic, articulated by economist Noah Smith: in the core Nvidia–OpenAI relationship, revenue only flows one way — OpenAI pays Nvidia for chips it needs for its actual business, and both firms are simply doing what they are set up to do. These are heavily scrutinised public companies, not shell entities.
The most credible middle-ground view may be Jamie Dimon’s: that “AI is real,” but that some money invested now will be wasted, and that the chance of a meaningful drop in stocks is higher than the market reflects. In other words — a real supercycle with a very real bubble riding on top of it.
Section 6: The India Angle — Riding the Wave Without the Froth
For Indian investors, this is not a distant American drama. India has become one of the fastest-growing nodes in the global AI build-out, and the way to participate here is structurally different — and arguably safer — than chasing frothy US mega-caps.
At the February 2026 India AI Impact Summit in New Delhi, domestic giants made staggering commitments:
- Reliance Industries announced a multi-year commitment (reported around $110 billion) to build AI infrastructure, with a gigawatt-scale project underway in Jamnagar and a data centre partnership with NVIDIA.
- Adani Group outlined plans to invest $100 billion to develop renewable-powered, AI-ready data centres by 2035, aiming to build the world’s largest integrated data centre platform.
- Global hyperscalers poured roughly $400 billion into India’s AI ecosystem over the past year, and Nomura projects India’s data centre capacity will rise to 7 GW by 2030 as it remains cost-efficient relative to other markets.
Here is the crucial insight for our readers, and it connects directly to the “picks and shovels” philosophy:
Because there are very few listed pure-play AI model companies in India, investors focus on listed enablers — the EPC contractors, power producers, cooling, connectivity, and hardware suppliers that benefit whenever compute capacity expands.
Brokerages have mapped this value chain. Bernstein highlighted Larsen & Toubro as a direct beneficiary via its engineering and construction role, with NTPC and Adani Green as indirect beneficiaries through rising power demand. The buildout also flows through telecom (fibre and submarine cable owners like Bharti Airtel and Reliance Jio), renewable-energy IPPs, and IT-hardware and cooling suppliers.
But a note of caution consistent with our house view: many of these “infrastructure proxy” stocks have already seen sharp re-ratings, increasing their sensitivity to news flow. Buying a diversified conglomerate like Reliance or Adani Enterprises for “AI exposure” also means buying the entire rest of that business — its complexity and its risks included.
Section 7: Lessons for the Intelligent Investor
Whether this ends as a soft landing or a hard correction, the timeless principles our platform champions apply with full force. The AI capex debate is ultimately a test of temperament, not prediction.
- Volatility is temporary; the trend is long. As we explored in our Trump vs India analysis, markets have absorbed 9/11, the 2008 collapse, and COVID-19 — and recovered each time. A technology this significant will not move in a straight line.
- Beware concentration. If AI-linked names now dominate the indices, a broad index fund is no longer as diversified as it feels. Rebalancing matters more than ever.
- Demand a margin of safety. Benjamin Graham’s core teaching — never confuse a great company with a great investment at any price — is the single best defence against a supercycle’s excess.
- Separate the technology from the trade. AI being revolutionary does not make every AI stock a buy. Both can be true.
- Time in the market beats timing the market. For most retail investors, disciplined SIPs into diversified funds remain a far wiser vehicle than attempting to trade a boom you cannot fully see inside.
Conclusion: The Power of Patience
The AI capex supercycle is the defining financial story of the decade — a genuine technological transformation wrapped in unprecedented, and possibly unsustainable, spending. History suggests both the optimists and the pessimists will be partly right: the infrastructure being laid down today may well power the next era of productivity, even as many of the investments made to build it are written off along the way.
For the disciplined investor, the message is not “avoid AI” or “buy everything.” It is the same lesson that has survived every mania before this one: understand what you own, insist on a margin of safety, diversify against concentration, and let time — not headlines — do the heavy lifting.
When giants move, the ripples can drown the small — but they can also lift the patient. The difference lies entirely in how you position, and how long you’re willing to wait.
Disclaimer: This article is for educational and informational purposes only and does not constitute investment advice. AI-related equities and infrastructure themes are highly volatile and subject to rapid change. The companies and figures mentioned are illustrative, not recommendations. Please consult a SEBI-registered financial advisor before making investment decisions.
