I build systematic investment & trading systems at Batra Hedge for a living, so I meet artificial intelligence from two directions at once โ as a tool remaking my own craft, and as the force remaking the market I trade. I can claim no special knowledge of either. Here, anyway, are my deliberations, from an Indian seat.
Let me begin with a confession and a caveat, because this memo is built on both. The confession: I am not an AI researcher. I cannot tell you, at the level of the mathematics, why a large model does what it does. The caveat: that disqualifies me from very little, because almost nobody can, and the question in front of every investor is not whether we understand the technology but whether we can position capital intelligently around something we only half understand. On that second question I have spent my career, and I have some things to say.
This memo is written in the spirit of a recent Howard Marks piece that asked, plainly, whether a layman could think about AI and invest in it intelligently. I found the question honest and the humility bracing, and I wanted to answer it from where I sit โ not Los Angeles but Surat, not a credit fund but a quantitative investment & trading firm, not the US mega-caps but the Nifty and the rupee. The conclusions travel differently from here.
Start with arithmetic, because arithmetic is the one thing that doesn’t care about the hype. Howard Marks lays out the logic plainly: to profit from a theme from the top down, you usually have to get several independent judgments right at once โ the right technology, then the right company within it, then the right price, then the right holding period. Suppose you are a genuinely good forecaster and get each of those four right 60% of the time. If the calls are independent, your chance of getting all four right is 0.6 to the fourth power โ about 13%. AI asks you to get more than four things right, and asks it under more uncertainty than usual. The hit rate implied by honest probabilities is humbling, and anyone who isn’t humbled hasn’t done the multiplication.
So I hold three things in mind throughout. We never know everything. The most expensive mistakes in markets come not from ignorance but from false certainty โ from people who knew. And yet incomplete knowledge is not a permission slip to abstain, because sitting out a generational technology is itself a large, active bet. The job is to act despite not knowing, at a size and price that survive being wrong.
* * *
Question one. What is India actually exposed to?
This is the first thing an Indian investor must get straight, and most don’t. When we say “AI” we picture frontier laboratories training enormous models. India has almost none of that. There is no Indian company on the Nifty whose fortunes rise and fall with the training of foundation models. Our index is banks, energy, consumer, autos, and โ this is the crux โ information-technology services.
So our exposure to AI is overwhelmingly indirect, and it runs through two channels that point in opposite directions. As a consumer of AI, India is enormous and early: a vast developer base, a billion-plus digital users, a state that has shown it can build population-scale digital rails. As a producer, we are thin at the top of the stack and thick in exactly the layer โ IT services, business-process work โ that the technology most directly threatens. We are long the application of AI and, through our largest export, short the thing it automates. Any honest memo for an Indian reader has to sit with that tension rather than resolve it prematurely.
Question two. Is AI a tailwind or a solvent for our biggest export?
The Indian IT-services industry was built, over three decades, on a simple arbitrage: skilled English-speaking engineers doing at a fraction of the cost what Western firms did in-house. The industry’s own one-year numbers have been soft โ the Nifty IT exposure has lagged the broad market this year โ and the market is clearly asking whether that soft patch is cyclical or something deeper.
Here is the uncomfortable framing. If a model can write, test, and document code, and handle a tier of customer support, then the very activity that is billed by the seat is the activity most exposed to being done by a machine. A business whose unit of sale is a human hour faces a genuine question when the human hour stops being necessary. The bullish case โ that AI makes each engineer radically more productive, that demand for software expands faster than the labour needed to build it, that integration and trust and accountability remain human work โ is real and may well win. But notice what the bullish case requires: it requires the industry to climb the value chain fast, from selling hours to selling outcomes, while its old pricing model erodes underneath it. I don’t know which force wins. I do know the market is repricing the question in real time, and that a soft year is the sound of that argument being had.
“We are long the application of AI and, through our largest export, short the thing it automates.“
Question three. Who captures the profit?
There is an old and useful distinction between a transformative technology and a profitable investment. The two are not the same, and confusing them has cost investors dearly for a century. The automobile remade the twentieth century; the great majority of early car-makers went bankrupt. The internet remade commerce; most of the companies that embodied it in 1999 no longer exist. Warren Buffett has made this point for years, usually with the airlines as his example โ a wondrous industry that was, for most of its life, a graveyard for capital. A technology can be real, important, and world-changing, and still deliver its benefits to customers and competitors rather than shareholders.
For AI the question is sharper because the economics are strange. Each generation of model is enormously expensive to train and, once a comparable open or cheaper model appears, the pricing power at the frontier can evaporate quickly. If models become commodities โ interchangeable utilities bought on price โ then the profit does not sit with the model-makers. It sits, maybe, with whoever owns the scarce inputs, or whoever owns the customer relationship and the proprietary data. For an Indian investor the question becomes: of the value AI creates, how much accrues to assets I can actually own at a sane price, and how much leaks offshore to a handful of foreign balance sheets I can reach only through a feeder fund? Being right about the technology and wrong about who profits is the classic way to lose money in a boom.
Question four. Can India build the physical layer โ and should we want to?
Strip away the software romance and AI is, at bottom, a physical industry: silicon, data centres, land, water, and above all electricity. This is the “picks and shovels” layer, and it is where a surprising amount of the investable, predictable value lives โ not in guessing which model wins, but in the electricity and real estate the whole industry must consume regardless of which model wins.
India has ambitions here: data-centre capacity is being built, there is a serious push on domestic semiconductor fabrication, and the appetite for digital infrastructure is real. But the bottlenecks are equally real, and they are old-economy bottlenecks โ reliable power at scale, grid, water for cooling, land acquisition, and the capital intensity of fabs that take years and billions before a single competitive chip ships. The AI story, for India, may turn out to be far more a power-and-infrastructure story than a software story โ which is a less glamorous bet, and often a better one. Ask yourself honestly which you would rather own for ten years: the cleverest model, which may be obsolete in eighteen months, or the power contract that every model must pay.
Question five. What does it do to valuations and concentration?
Globally, a handful of companies now carry an outsized share of the major indices, and their valuations embed years of AI-driven growth. An Indian investor meets this concentration the moment they buy an international feeder โ our largest overseas ETF is, in effect, a concentrated bet on exactly these names. That is not a reason to avoid it; it is a reason to know what you own. When you buy “US tech” today you are not buying diversification, you are buying one trade held by almost everyone.
And concentration has a reflexive quality in an AI capital cycle. A large share of the spending that shows up as one company’s revenue is another company’s capital expenditure, and some of that spending is financed with debt and circular commitments between the same handful of players. None of that is proof of a bubble. But it means the layer is more fragile than its smoothness suggests, because a pause in capex by a few buyers withdraws revenue from many sellers at once. Smooth-looking systems built on circular financing have a way of discovering their correlations all at the same moment. I say this as someone whose own profession is, at its core, the pricing of exactly that kind of hidden correlation.
Question six. What does AI do to a quant’s edge?
This question is personal, so let me be plain about it. My firm runs systematic strategies. The same tools transforming every other knowledge business are transforming mine. The optimistic reading is that AI is a force multiplier โ faster research, cleaner pipelines, more hypotheses tested per week. I believe that, and we use it.
But there is a harder truth underneath. An edge that a machine can discover is an edge a machine can erase. When powerful models are cheaply available to everyone, patterns get found and arbitraged faster, crowded strategies get more crowded, and the half-life of any given signal shortens. The advantage migrates away from having a clever idea โ ideas are now cheap โ toward the things models don’t give you for free: proprietary data, execution quality, risk discipline, capital that can sit through a drawdown, and the judgment to size a bet correctly when every screen is screaming the same thing. In a world where everyone has the same brilliant analyst, the edge is no longer intelligence; it is temperament and plumbing. That is a genuinely different competitive landscape, and I’d rather name it than pretend my moat is untouched.
Question seven. Is it a bubble?
A bubble is not a high price; it is a price justified only by the assumption that someone will pay more tomorrow โ a psychology, not a number. By that definition the honest answer to “is AI a bubble” is: in parts, probably, and in parts, probably not, and which parts is largely a matter of faith about a future nobody can see.
I find it useful to score the industry by layer rather than argue about it as a whole. The frontier labs are the least predictable part โ highest potential, highest chance of being disrupted by the next lab or a cheaper open model. The physical layer is the most predictable โ power and real estate will be consumed by whichever model wins. The table below is my own rough calibration, not a forecast; its only purpose is to insist that “AI” is not one thing you are either in or out of.
| Layer (India lens) | Predictability of demand | What you’re really betting on |
|---|---|---|
| Power & grid | High | Electricity gets consumed regardless of which model wins |
| Data centres & real estate | High | Capacity and location; a landlord, not a guesser |
| Chips / fabrication | Medium | Capital intensity and years of execution before payoff |
| IT-services adopters | Medium-Low | Whether they climb from selling hours to selling outcomes |
| Frontier models(Open AI, claude) | Low | Staying ahead of the next lab and the next open model |
Notice what the table implies for behaviour. The most exciting part of the industry to talk about is the least predictable part to own, and the most ownable part is the one nobody writes headlines about. That mismatch โ glamour and predictability pointing in opposite directions โ is where most of the money will be made and lost.
Question eight. So how should an Indian investor act?
Not by abstaining, and not by betting the farm. Both are the same error in different clothes โ a refusal to hold uncertainty at a sensible size. Here is the discipline I’d argue for, and try to practise.
Decide the role before the ticker. Is this your long-term compounding core, your diversifier, or your risk capital? The answer sets the size, and size matters more than selection when the range of outcomes is this wide. Position sizing, not stock picking, is what keeps you solvent through a technology you cannot forecast.
Prefer the predictable layer unless you are being paid handsomely to take the unpredictable one. Owning the power and infrastructure that AI must consume is a quieter way to be long the theme than owning the model you hope wins. And make price the deciding variable, because a wonderful technology at a terrible price is a bad investment, while an ordinary one at a wonderful price can be a fine one. At the end of the analysis, intelligent investing is mostly about price.
Then apply a test I like for any confident forecast: take your five-year estimate of a company’s earnings and ask whether you’d bet the real number lands within plus-or-minus fifty percent of your figure. For a toll-road-like infrastructure asset you might say yes. For a frontier model-maker, be honest โ can you even bound it? If you cannot state a range you’d bet on, you don’t have a valuation, you have a hope. Hope is allowed. It just has to be sized like hope.
The Batra Hedge view
AI may be the most consequential and least forecastable development most of us will invest through. From an Indian seat the exposure is mostly indirect, the profit mostly accrues where we can least easily own it, and the surest part of the story is the unglamorous physical layer. The sensible response is neither faith nor refusal, but calibrated humility: act, because sitting out is also a bet; size small, because you will be wrong more than you think; and let price, not excitement, make the decision.
There is an old line that two kinds of investors lose a great deal of money โ those who know nothing, and those who know everything. AI has produced an unusual number of the second kind. I am trying hard to be neither. I don’t know how this ends. I know roughly what I’d pay to find out, and that, in the absence of certainty, is the only honest place to stand.
Disclosure:
This memo is published by Batra Hedge for information and education only. It reflects the author’s personal deliberations and is not investment, legal or tax advice, nor an offer or solicitation to buy or sell any security, fund or strategy. Companies, sectors and instruments are referenced to illustrate reasoning, not as recommendations. Views on an uncertain subject are opinions that may be wrong and may change. Investments in securities and mutual funds are subject to market risks; read all scheme-related documents carefully. Past performance may or may not be sustained and is not a guarantee of future returns. Assess your own circumstances and consult a qualified, appropriately licensed adviser before making any investment decision.

