AI Is Getting Expensive, and the Executives Paying for It Are Starting to Notice

Key points

• A KPMG survey of more than 2,000 senior executives across 20 countries found 29 percent struggled to understand the running costs of their AI deployments.
• Nearly half said they were looking to slow or reshape AI projects when costs outweigh the benefits.
• Gartner analyst Nitish Tyagi forecasts that by 2028, the per-developer cost of AI coding assistants will exceed the global average developer salary.
• In countries with lower wages, such as India, that crossover is already happening.
• Anthropic, OpenAI, and GitHub have shifted from flat monthly fees to usage-based billing, where companies pay per "token," meaning each small chunk of text the AI reads or writes.

Businesses rushed to adopt AI tools over the past two years. Now the invoices are arriving, and the numbers are alarming boardrooms.

KPMG, the global consultancy, polled more than 2,000 senior executives across 20 countries and found 29 percent couldn't get a clear picture of what their AI deployments cost to run as they scaled. Reported first by The Register AI, the findings capture companies that jumped in fast and are now scrambling to understand what they signed up for. We first covered KPMG's AI cost reporting on 13 July 2026, when open-source alternatives were already being floated as a pressure valve.

The confusion has a direct cause. Flat monthly subscriptions made budgeting easy: fixed fee, unlimited use, predictable bill. That model's gone. Anthropic, OpenAI, and GitHub have all moved to token-based billing. A token is a small unit of text, roughly three quarters of a word, and every prompt sent and every reply received burns through them. Usage spikes; the bill follows without warning.

Nearly half the executives surveyed said they were considering "re-phasing" their deployments: slowing down, narrowing scope, and mixing in cheaper or smaller models rather than defaulting to the most capable option.

What does this mean for workers whose companies use AI tools?

It may actually protect some jobs. Gartner's Nitish Tyagi found that in lower-wage markets, the cost of running an AI coding assistant already exceeds what a human developer earns. Globally, he projects that crossover arrives in 2028. When a tool costs more than the person it was meant to replace, the financial case for replacement collapses.

Gartner's research also found something counterintuitive: using more tokens doesn't automatically produce better code. Developers who wrote precise, disciplined prompts got higher-quality results and spent less than those who used the AI freely and often. That's a skill gap most companies haven't budgeted to close.

The wider financial picture is hard to ignore. One major investment analysis found the AI industry is on track to spend 1.5 trillion dollars on data centre infrastructure through 2030. That spending has to be recouped, which is exactly why pricing is shifting toward usage-based models that grow with each customer's footprint.

For companies already dependent on these tools, switching isn't simple. Developers who've used AI assistants daily for a year or two can lose fluency with the underlying skills, and that dependency gives providers real pricing use. Charge too little, and labs never turn a profit. Push rates too high, and enterprise customers will route around it: open-source models, smaller specialist tools, or simply rehiring the staff they let go.

The honest read on all this is that usage-based billing isn't a temporary growing pain. It's the business model, and for any company that didn't stress-test their AI costs before getting hooked on the productivity gains, the reckoning is already underway.