AI, Capital Cycles and Why India Looks Reasonable
When a theme takes over a third of the book shelf, it has moved from the research desk to the dinner table.

Yesterday I walked into Bahrisons in Galleria, Gurgaon, one of my favourite bookstores, and went straight to the investing section. I did a rough count. Close to 30% of the titles on the shelf were about artificial intelligence.
My first reaction was not excitement. It was a memory of Howard Marks' Mastering the Market Cycle. Marks teaches that the most useful thing an investor can do is take the temperature of the market. You do that not by forecasting, but by noticing how people around you are behaving.
When a theme takes over a third of the book shelf, it has moved from the research desk to the dinner table. That is not proof of a top. But it is exactly the kind of observation Marks would ask us to write down.
The IPO parade
The bookshelf is not the only signal. The supply of AI paper is arriving in public markets in size.
SpaceX (which now includes xAI) listed in June 2026 after a roadshow targeting a raise of about $75 billion.
Anthropic, the maker of Claude, raised $65 billion at a $965 billion private valuation in May, filed its S-1 publicly on 26 October, and began marketing its Nasdaq listing this week under the ticker ANT. Reports point to trading as early as 19 or 20 November.
OpenAI last raised at an $852 billion valuation in March and is preparing its own listing.
Watch the valuation talk, because it tells you where we are in the cycle. Through August and September, backers spoke of roughly $2 trillion for Anthropic. After last week's tech sell-off, institutional buyers reportedly pushed that toward $1.5 trillion, with some talking of $1–1.25 trillion. A trillion dollars of "value" moved in about a month on mood alone.
Sellers like to sell when buyers are most eager. Founders and early backers are not wrong to list now. But the buyer of a record-sized IPO is rarely buying at the point of maximum pessimism.
What Howard Marks is worried about
Marks published "A Layman Thinks About AI" on 11 November 2026. He is careful to say AI is real and consequential, and that investors cannot simply abstain. His worries are about predictability, concentration and financing, not about the technology.
Too many variables. An AI investor must be right about the technology, demand for compute, competition, who captures the profit, financing and timing. Multiply even generous odds across all of them and the chance of being right on the whole thesis is small.
Who keeps the profit? If several models become equally capable, AI may become a commodity sold on price. The gains could flow to users of AI rather than its vendors. Today the clear winners are chipmakers like Nvidia and TSMC.
Earnings no one can forecast. His test: name an AI company's EPS five years out and bet it lands within ±50%. Few would take that bet.
A single fault line. AI's contribution to US GDP growth comes mostly through capex on data centres. J.P. Morgan estimates 42 AI-linked stocks produced about two-thirds of the S&P 500's price gains since January 2024. Alphabet and Amazon booked $151 billion of gains on stakes in Anthropic and SpaceX in one quarter.
Debt and circularity. AI-related bond issuance is near $600 billion this year, about a fifth of US investment-grade supply. Vendors finance buyers who buy from vendors. Hyperscalers that were once debt-free are spending their cash piles and borrowing.
Marks asked Claude itself to rate AI's predictability on a 1–10 scale. Frontier labs like OpenAI and Anthropic scored 2, comparable to early-stage venture capital. He notes the irony, and the fact that these least predictable assets carry premium valuations in the private market. His summary is that, for now, the AI investment case is largely a matter of faith.
Where the money went: capex up, free cash flow down
For three years, US market returns have been powered by AI spending more than by AI profits. The Magnificent 7 delivered 62% of S&P 500 returns in 2023 and 53% in 2024, per State Street. In 2025 they accounted for over 40% of the index's earnings growth. Remove them, and the US bull market looks ordinary.
The engine was capex. Goldman Sachs projects the big four hyperscalers will spend $1.15 trillion over 2025–27, more than double the $477 billion of the prior three years. Combined hyperscaler capex is expected to exceed $690 billion in 2026. Without AI spending, Pantheon Macroeconomics estimated, US corporate capex would be negative.
Now look at what that spending did to cash:
2024 | 2025 | 2026 (est.) | |
Free cash flow, big four hyperscalers | $237bn | $200bn | Near zero or negative for all but Alphabet and Microsoft (FactSet) |
Amazon's FCF was down to $1.2 billion on a trailing basis by Q1. Oracle's capex reached 174% of operating cash flow, and S&P cut it to BBB- in July. Reuters estimates these firms need about $1.57 of new investment for every $1 of extra cash flow.
In other words, much of the US earnings story has been one company's capex becoming another company's revenue. Nvidia's sales are Microsoft's capex. Anthropic's valuation gains are Amazon's and Alphabet's "other income". That is not fake money. But it is not yet free cash flow returned to owners either, and the sector has become a very large share of what holds the index up.
The capital cycle turns, and India is on the other side of it
Capital Returns, the collection of Marathon Asset Management's letters edited by Edward Chancellor, rests on one idea. High returns attract capital, the capital builds capacity, competition rises, and returns fall. Low returns repel capital, capacity shrinks, and returns recover. Marathon's advice is to follow the trail of investment: be wary where capex is elevated and competition is fierce, and look where capital has been starved.
By that test, AI is late in the cycle. Capex is at records, debt is funding it, IPOs are lining up, and token prices are falling fast. India sat on the other side. Through most of 2025 and 2026, foreign investors treated India as an "AI underperformer": a market of banks, consumer companies, industrials and IT services with little AI-infrastructure exposure. Before November, FPI outflows from Indian equities in 2026 had crossed ₹3 lakh crore. In October alone they sold ₹34,843 crore.
Then the flow reversed. On 5 November, all three major US indices had their worst day since April 2025, as investors questioned whether AI data-centre spending would earn its keep. In the first week of November, FPIs bought ₹38,971 crore of Indian equities, per NSDL data. Three things helped:
The AI trade cracked. Money rotated out of AI-heavy markets, including Korea and Japan, and India's low AI exposure became a feature.
Oil eased. Brent fell about 13% in the week, to near $97, on hopes of de-escalation with Iran. The rupee firmed to about 94.15 to the dollar from 95.40.
Valuations had reset. After months of selling, Indian large-caps looked reasonable, and foreign positioning in India was light.
One week is not a trend, and FPI money is famously fickle. But the shape fits the capital cycle. Capital that crowded into AI has to go somewhere when the trade wobbles, and it is going to markets it had starved. India is a services, consumption and domestic-credit story, not an AI-innovation platform. In this phase of the cycle, that may be exactly what the marginal foreign dollar is looking for.
Buffett, 1979: manias settle
Marks ends his memo by handing the last word to Warren Buffett, and so will I. In March 1979, in his Berkshire letter, and again in a Forbes article on 6 August 1979, Buffett described how US pension managers behaved through the Nifty Fifty era.
In 1971 and 1972, with the Dow near 1,000 and well above book value, they put 122% and 105% of their net new money into stocks. They were selling bonds to buy more.
In 1974, after the crash, with stocks far cheaper, they committed only 21%.
In 1978, with the Dow below book value most of the year, the figure fell to 9%.
The more they paid, the better they felt. Buffett's conclusion was that the future is never clear, and a cheery consensus is expensive. Uncertainty is the friend of the long-term buyer.
The Nifty Fifty were not bad businesses. Most were excellent. The mania was not about whether they were real, but about the price people would pay for certainty. It settled the way manias always do: not because the story was false, but because the price had run ahead of the cash the businesses could produce. AI may follow the same path. The technology will likely prove as consequential as its champions say, and many AI investors may still earn poor returns from today's prices.
Conclusion: reasonable times ahead for India
Put the pieces together. The AI trade is crowded, capital-hungry and priced on faith. US returns have leaned on capex rather than free cash flow. The IPO pipeline is full and the bookshelves are full. Those are late-cycle signs, even if no one can name the top.
India sits on the other side of that cycle. Foreign capital has spent two years leaving, valuations have cooled, and our market is built on services, consumption and credit, not on data centres. If even a slice of the money leaving the AI trade looks for growth outside it, India is one of the few large markets that can absorb it.
I would not call this a boom. Call it reasonable times ahead: fair valuations, a cleaner foreign-ownership base, and a market that no longer has to compete with the AI story for every marginal dollar.
For investors, the lesson is Marks' and Buffett's, not mine. Do not buy a story because everyone is telling it, and do not ignore a market because everyone has left it. Stay with your asset allocation, let price guide you more than narrative, and remember that a third of a bookshelf is a data point too.
.png)




Comments