For two years, enterprise AI ran on a single, intoxicating question: how fast can we spend? Bigger context windows, more tokens, more agents, more pilots. The number on the invoice was treated as proof of ambition. In June 2026, that question quietly changed, and the new one is far less comfortable: what did any of it actually return?

I have spent this year deploying AI inside a large enterprise, sitting in the rooms where these budgets are defended. So let me say plainly what the month's news is circling: the era of AI as a science experiment is ending, and the era of AI as a line on the P&L has begun. That shift will separate the organisations that got value from the ones that merely got invoices.

The tell: from "tokenmaxxing" to efficiency

The clearest signal came from the demand side. Reporting late in the month described enterprise users shifting from "tokenmaxxing" — maximising usage at any cost — toward cost efficiency, just as OpenAI and Anthropic, the chief beneficiaries of the spend-at-all-costs mindset, gear up for trillion-dollar IPOs. The most vivid example in that coverage: the chief executive of one AI startup switched his company entirely off premium frontier models, moved 100% of its traffic to a cheaper open-weight alternative, and watched his cost curve, in his words, crash to the ground.

One anecdote is not a trend. But it rhymes with everything else that happened this month, and the pattern is unmistakable. The buyer has stopped asking "how powerful is it?" and started asking "what does it cost me per outcome, and can I get the same outcome cheaper?" That is not a small change in tone. It is the moment a technology stops being a frontier and starts being a budget.

The CFO has entered the chat

The people now shaping AI strategy are not the ones you would expect. Across industries, chief financial officers are moving aggressively to impose budget controls on AI projects, replacing the open-ended experimentation that defined the last two years. And the numbers explain why. As enterprise AI moves from pilot to production, only about one in four AI initiatives is delivering the ROI that was promised, even as vendors cite headline returns of several rupees for every one invested. Gartner now expects more than 40% of agentic AI projects to be cancelled by 2027, and reports that only a fifth of companies running autonomous agents have a mature governance model for them.

Read those two facts together and the cancellations stop being a mystery. We did not have an intelligence problem. We had a discipline problem. Thousands of pilots were funded on the promise of what AI might do, with no gate for what it actually delivered, and no control plane to govern it once it did. The bill for that indiscipline is exactly what is now coming due.

Why the cost curve is finally bending

The supply side is responding to the same pressure, and this is genuinely good news for buyers. The defining infrastructure story of the month was vertical integration aimed squarely at the cost of inference. OpenAI unveiled its first custom inference chip, Jalapeño, built with Broadcom, to reduce its structural dependence on Nvidia GPUs. Microsoft pushed its own in-house MAI models, explicitly designed to lessen reliance on OpenAI and lower costs for developers. And open-weight pressure from cheaper, often Chinese, MIT-licensed models kept dragging the floor price of "good enough" intelligence lower every month.

The strategic lesson here is the one I have argued all year, now proven in hardware: the model is a commoditising input. When the most powerful labs in the world are racing to make inference cheaper and the open-weight tier is closing the quality gap, the rational enterprise posture is not loyalty to a single supplier. It is portability, so you can ride the cost curve down without re-platforming every time a cheaper option appears.

But cheaper compute is not the same as ROI

Here is the trap I would warn every leader about, because the efficiency narrative has a seductive false floor. Cutting your cost per token is not the same as creating value. You can make a worthless pilot 90% cheaper and it is still worthless. The danger of an efficiency drive is that it optimises the wrong number, celebrating savings on initiatives that should never have been scaled in the first place.

The organisations actually getting returns are doing something quieter and harder. They are governing. When Microsoft and KPMG scaled agents across a 276,000-person workforce in June, the headline word was not "powerful" — it was "governed": centralised control of agents operating across systems, data and processes. The most valuable agents this year are not the ones dominating demo reels; they are the unglamorous ones running autonomously inside companies, where a retailer can quietly let an agent negotiate supplier contracts. Value is coming from disciplined, governed, narrowly-scoped deployment, not from spend and not even from savings alone.

What this means for enterprise leaders

I am telling my own organisation to treat this month as the end of the experimentation budget and the start of the accountability budget. Concretely, four commitments. First, treat the model as a swappable, commoditising input and build for portability, so falling prices flow straight to your margin instead of locking you in. Second, put a CFO-grade ROI gate in front of every agent before it scales — a defined outcome, a baseline, and a number, because an agent with no measured return is a liability with an API key. Third, fund the governance and control plane first, since the reason 40% of these projects will be cancelled is ungoverned sprawl, not weak models. Fourth, measure the quiet wins, not the demos, and reward the teams shipping boring, dependable, governed automation over the ones with the most impressive prototype.

There is an India-shaped advantage threaded through all of this. The hype era rewarded whoever could spend the most, and that was never going to be us. The efficiency era rewards disciplined, cost-aware, domain-wrapped engineering, which is precisely the muscle Indian technology has spent decades building. As the model layer commoditises and the global question becomes "what is the governed, ROI-positive outcome?", the advantage shifts from the firms with the biggest compute budgets to the firms with the deepest domain discipline.

My take

The hype era funded the experiments. The efficiency era will decide who actually got value from them, and the verdict will be harsher than the spending implied. The winners of 2027 will not be the organisations that spent the most on AI, nor even the ones that saved the most. They will be the ones that finally started governing AI like a P&L instead of a science project: portable at the model layer, disciplined at the ROI gate, serious about governance, and honest about which pilots deserved to die. The bill has come due. The good news is that paying it, properly, is also how you finally start earning a return.