30 July 2026
For the last three years, the most expensive experiment in technology history has run on someone else’s money. Investors — venture funds, hyper-scalers and sovereign wealth — have poured record sums into artificial intelligence. Global corporate AI investment more than doubled in 2025 (Stanford’s 2026 AI Index), and the four biggest US tech firms are on track to spend close to US$700 billion on AI infrastructure this year alone. OpenAI reportedly spends around US$60 billion a year on compute against roughly US$13 billion of revenue. That gap has been a gift to the rest of us: powerful tools, cheap or free.
The gift is being withdrawn.
In the past few months, the bill has started moving downstream — to the companies and people actually using AI. In July, Tesla capped staff at US$200 a week on external AI tools after engineers ran up thousands in token charges. It’s not alone. TechCrunch reported that Uber burned through its entire 2026 AI coding budget by April and Microsoft revoked developers’ Claude Code licences months after issuing them. CNBC describes a market-wide pivot from “tokenmaxxing” to efficiency as businesses scramble to rein in “out-of-control token spend.” Even the giants feel it: Meta is trimming thousands of roles partly to help fund an AI capital programme north of US$115 billion.
Read those stories together and a clear pattern is emerging. The subsidised era of AI is ending. Usage-based pricing means every prompt, every reasoning step and every autonomous agent now carries a real, metered cost — and someone must pay it.
That word — agent — is where this gets serious for New Zealand businesses. We’ve moved well past chatbots. Autonomous agents now commit spend, reconcile transactions and make decisions that bind the organisation. They’re also expensive in a way traditional software never was. Because agents reason in loops — planning, acting, re-planning — one agent can consume an order of magnitude more compute than a simple query. As our own research puts it, “an agent stuck in a reasoning loop… can generate substantial costs before anyone notices.” For a business without Silicon Valley’s balance sheet, that math matters even more.
Some of that cost is not inherent in the intelligence itself. It comes from asking a highly capable model to operate inside businesses whose meaning was never made machine-readable. What counts as a customer, which margin definition applies, when a discount needs approval — experienced staff know; agents do not. They compensate by searching, querying and reasoning in loops, often still reaching the wrong answer. Without a shared meaning layer, businesses pay twice: once for the extra compute, and again for the mistake.
The numbers back it up. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, blaming escalating costs and inadequate controls. These aren’t build-phase failures — they’re run-phase failures. Yet most organisations running agents today can’t see what they cost in real time, set no per-agent budget, and get no alert when one goes rogue. They find out when the invoice lands.
This is exactly why Fusion5 is investing in building an Agent Operations Centre.
The industry has spent enormous effort making agents easy to build — and almost none making them safe and affordable to run. Every mature technology eventually earns an operational discipline: networks got the NOC, security got the SOC. Agents need the same. Our Agent Operations Centre is that layer — a dedicated function that watches every agent in real time, tracks cost per agent with thresholds and alerts, it monitors each agent’s guardrails and escalates to a human the moment an agent steps out of line. We set out the full thinking in our Agent Operations Centre whitepaper (Whitepaper: The Agent Operations Centre | Fusion5 New Zealand) and across our Microsoft AI Tour 2026 sessions, “Prepare Your Organisation to Operate AI at Scale.”
My message to any leader eyeing the potential of agents is simple. The technology is real and the upside is genuine — we’re all-in on it. But harnessing it without monitoring and controls is no longer a technical nicety. On the evidence now landing from around the world, it’s the fastest route to a very large bill and a nasty governance surprise.
The businesses that build the discipline to operate AI — not just deploy it — will be the ones still standing when the music, and the free money, stops.