A fortnight ago, the company that gave this whole boom its face said nothing about the one move everyone expected it to make with fanfare. OpenAI, the outfit whose product launches have reset every boardroom conversation I have joined since 2023, is now reportedly leaning towards pushing its blockbuster IPO to 2027 rather than floating in 2026. The moment that reporting landed, the AI trade hit a wall and dragged SoftBank down with it. This was meant to be the largest technology listing in history, and the firm that all but wrote the definition of the era would rather wait than hand it to the public just yet. Traders betting on the timing have been marking down the odds of a 2026 float ever since.

When the poster child hesitates, every chief information officer who has standardised on a single frontier lab should look up from the roadmap. I did. And I came away convinced that almost everyone is reading this moment wrong, the bubble crowd and the revolution crowd both, yelling past each other. The IPO delay is a sideshow. It distracts from a collision happening in plain sight, and once you see the collision you stop caring about the listing calendar.

The froth is real. The trillion is a distraction.

Let me concede the froth first, because pretending it away helps nobody. Anthropic's implied value on the secondary market touched 1.2 trillion dollars, up roughly 550 percent in a year, overtaking OpenAI in the process. Read the fine print in that same report and you find brokers admitting demand outstrips supply for a single reason: no one is selling. A price set when nobody sells is really just a queue with a number stuck on it. Menlo Ventures' Matt Murphy called it a noisy signal, and he was being generous.

So the warnings have merit, and much of the bubble chorus comes from people I respect. But the framing is binary and, for anyone actually running deployments, close to useless. Froth in the secondary market tells you something about capital and liquidity. It tells you almost nothing about whether the underlying technology throws off value you can bank. Those are two separate questions, and the is-it-a-bubble debate keeps welding them into one. I have sat in meetings where a genuinely sound deployment decision got postponed because someone waved a chart of secondary valuations at the room. That is the wrong chart for the wrong decision.

The question everyone is getting wrong

Here is what I keep saying to boards. Bubble-or-not is a market-timing bet, and I have no edge on market timing. Neither does your CIO, whatever the deck claims. Calling the top is a game for traders, and most of them lose it too. The operator's question sits at a right angle to all of that: is the value durable?

Look at the standoff and you can see why both camps are partly right. JPMorgan reckons the build-out could reach 5.5 trillion dollars of capital spending through 2030, with hyperscaler operating cash flow nearing 900 billion by 2027. It projects hyperscaler capex at around 650 billion dollars in 2026, climbing past 1.1 trillion in 2027, and calls the whole thing profitable for now, provided real demand shows up to justify the outlay. The bears counter that equity valuations have detached from any earnings you can point to. Both positions can hold at once. Infrastructure demand can stay hot while the equity wrapped around it is priced for a fantasy. Once I accept that, the binary loses its grip, and I am left with the only question I can act on. Can a given AI bet survive the thing already happening to it, which is a price war?

The price war is already eating the valuations

This is the collision. Trillion-dollar valuations only make sense if you assume permanent pricing power, the ability to charge a premium for frontier intelligence more or less forever. That assumption is dying in real time.

Meta's Muse Spark 1.1 API undercut both OpenAI and Anthropic and opened an outright price war on capable models. DeepSeek's V4 landed at roughly 35 times cheaper than GPT-5.5 on output tokens, about 97 percent below, and American companies from Airbnb to Cursor are wiring Chinese models like DeepSeek and Qwen into production as frontier costs climb. Amazon's chief technology officer Werner Vogels has said plainly that enterprises are shifting towards cheaper open-source models over the biggest, most expensive ones as costs mount. And Palo Alto Networks' chief executive put a number on it. Token costs, he argues, need to fall by around 90 percent before broad, profitable enterprise deployment stacks up.

Sit with that for a second. On one side, valuations that need the price war never to arrive. On the other, a price war that arrived last month. That gap is where the bubble actually lives, not in the listing calendar and not in the trillion-dollar headline. Everything else is theatre around it.

There is a version of this where the price war is very good news for buyers and quite bad news for anyone holding the equity. That is roughly where I have landed. The cost of running a capable model in production has fallen faster than almost any forecast I signed off on eighteen months ago, and it is still falling. For a chief financial officer that is a gift. For an investor underwriting permanent margins, it is the thing that undoes the model from the inside.

Follow the plumbing, not the press release

If durable value is the question, I would send a sceptical board to the balance sheet, not the launch keynote. Big Tech's combined debt has roughly doubled to about 350 billion dollars over five years to fund data centres. Nvidia and SoftBank are separately pouring money into so-called neoclouds, the specialist GPU landlords that keep the hyperscalers supplied, and the loop underneath a lot of this is close to self-referential. Nvidia funds OpenAI, OpenAI commits to Oracle, Oracle buys Nvidia chips, and demand keeps looking real because everyone in the circle is a customer of everyone else. The Deccan Chronicle laid the structure out bluntly, noting OpenAI trades near 40 times sales while losing somewhere in the 15 to 20 billion dollar range.

Put the unit economics beside that. OpenAI is reportedly losing 1.22 dollars for every dollar it earns. Debt-funded capital spending, revenue that partly circles back on itself, and negative margins on the transaction itself. That configuration does not survive a 90 percent price cut. This is not a moral judgement about anyone's ambition. It is arithmetic.

Why India's boring AI story ages better

Which brings me home, to an argument I have been making in some fairly sceptical conference rooms. India is oddly insulated here, for an unglamorous reason. Our AI story was never built on valuation. It was built on revenue and delivery, on money that has already changed hands for work that has already shipped.

Take TCS. Last quarter it reported an AI annualised run-rate of about 2.6 billion dollars, anchored by a landmark 800 million dollar multi-year AI-led transformation deal with the Swedish manufacturer SKF, with AI moving off the slide and into the profit and loss statement where it can be audited. Sarvam, meanwhile, became India's newest AI unicorn at 1.5 billion dollars on a 234 million dollar round led by HCLTech, building sovereign, Indian-language models aimed squarely at regulated banking, insurance, government and defence workloads. When frontier access gets restricted or repriced, a sovereign fallback stops being a nice-to-have and becomes a procurement requirement.

I will not sell you a fairy tale, though, because contrarian discipline cuts both ways. If the bubble bursts, India feels it. Foreign money would leave, the rupee would soften, and the hiring freezes would land hardest on the global capability centres in Bengaluru and Hyderabad, exactly as the Deccan Chronicle set out. And Fractal Analytics, India's first AI company to IPO, debuted at a discount of about 2.67 percent, a small sign that public markets are already pricing in doubts about durability rather than hype. The lesson for Indian CIOs and CFOs is not immunity. The lesson is that vendors who put revenue first, and can fall back on something sovereign if they have to, will outlast the price war far more comfortably than a trillion-dollar frontier bet ever could.

My take

I am not going to tell you when the bubble pops. I have no idea, and anyone who claims to is selling something. What I can tell you is how I now size up any AI vendor bet, and it comes down to five questions rather than one wager.

First, does it survive a 90 percent price cut? Is there any margin under today's price, or does the whole model rest on premiums that the Palo Alto number says are already draining away?

Second, is the revenue something I can see and verify, like a signed SKF contract booked into a P&L, or is it a run-rate narrative floating on a pitch deck?

Third, is the demand real or circular? Trace who ultimately pays. If the customer is funded by the supplier, you are staring into a hall of mirrors.

Fourth, what does re-platforming cost me if this vendor folds or triples its prices next year? Did I plan the exit and keep a hybrid, open-weight or sovereign option in reserve, or did I hand someone a switch they can flip on my budget?

Fifth, and this is the one people dodge. Am I buying a capability or buying a story? An AI system that flags a security hole nobody spotted for ten years in widely used code earns its keep, and it goes on earning it whatever the market does. A valuation propped up by the fact that no one is selling does not clear that bar.

My bottom line is unfashionable, and I will own it. I am not betting on the timing of the pop. I am assuming the price war wins, because it already is, and I am buying accordingly. The durable winners will be whoever ships the concrete capability at a price that outlasts the war. From where I sit in India, that looks less and less like the trillion-dollar names we are all told to envy, and more like the boring, revenue-first operators who were never invited to the party. Boring is starting to look like the smartest seat in the room.