top of page
Gammatek ISPL LOGO

Gammatek ISPL

Gammatek ISPL

Gammatek_green_LOGO_FINAL.png

AI Costs Are Rising for Australian Businesses Even as Tokens Get Cheaper

  • Writer: Gammatek ISPL
    Gammatek ISPL
  • 1 day ago
  • 6 min read

By Gammatek ISPL, Industrial Systems & Compliance Analyst at Gammatek ISPL Published: August 2026 | 10 min read

Author block:Gammatek ISPL advises manufacturing, chemical, and pharmaceutical plants on industrial safety, compliance, and technology adoption at Gammatek ISPL. This analysis draws on Gammatek's own experience deploying AI-driven monitoring tools (including FixitX) across industrial clients, combined with reporting from ABC News, Epoch AI research, and McKinsey's 2026 enterprise AI cost analysis.

Factory dashboard graphic showing AI spending rising while token prices fall, 2026
Australian manufacturers are discovering the same paradox hitting every industry: cheaper AI, bigger bills.

Why Should You Care?

If your plant has started using AI — for predictive maintenance alerts, compliance report drafting, quality inspection, or scheduling — you've probably budgeted for it based on "per-token" pricing that looked cheap. Here's the problem: token prices have fallen by more than 90% since 2023, yet Australian businesses are spending more on AI than ever before, with national AI spending already estimated at $5-8 billion a year and projected to climb toward $20-40 billion within a decade. If you run a plant, a maintenance team, or a compliance function, this isn't an abstract tech-industry story — it's about to show up as a bigger, harder-to-predict line item in your own budget, and most manufacturers aren't set up to catch it before it happens.

The Paradox: Cheaper Tokens, Bigger Bills

A token is the small unit of text an AI model reads and processes — a short sentence might be a handful of tokens, a detailed maintenance report could be thousands. According to research institute Epoch AI, the cost of running top-tier AI models has fallen dramatically since late 2023, driven by more efficient chips, more competition between AI providers, and a wave of new data centre capacity. By most industry measures, the price of processing the same amount of AI work has dropped by a factor of ten or more per year.


Logically, that should mean AI gets cheaper to run over time. Instead, the opposite has happened at the enterprise level. Reporting on the Australian market has found firms across sectors are spending significantly more on AI year over year, even as the per-token cost keeps falling. The explanation, according to technology leaders quoted in that reporting, comes down to volume: businesses aren't doing the same amount of AI work for less money — they're finding far more tasks to hand to AI, and each of those new tasks adds tokens on top of what came before.

Why This Happens: The "Cheap Steam Engine" Problem

There's a useful historical parallel here. When steam engines became more fuel-efficient, total coal consumption didn't fall — it rose, because cheaper-to-run engines got used for far more jobs than before. AI is following the same pattern. Once a task becomes cheap enough to automate, teams don't do the same volume of work for less — they find dozens of new places to apply it.

For a plant, this looks like: last year, AI might have drafted the odd compliance summary. Today, the same AI tool might be running continuous equipment monitoring, generating maintenance alerts, drafting incident reports, checking regulatory language, and answering staff questions — all day, every day. Each of those is individually cheap. Multiplied across a full operation running continuously, the total adds up fast, and most finance teams don't have a clean way to see it coming.


There's a second factor specific to how modern AI tools actually work: agentic AI. Rather than a single question-and-answer exchange, many AI systems now plan, check their own work, call other tools, and loop through several steps before finishing a task — and at every step, they resend the full context of what they're working on. By the twentieth step of a complex task, a system may be re-processing the same background information twenty times over. That means a single "task" can consume far more tokens than the deceptively cheap per-token price suggests.

The Industrial-Sector Angle Most Coverage Misses

Most coverage of this trend focuses on white-collar use cases — chatbots, coding assistants, office productivity tools. Manufacturing, chemical, and pharma operations face a version of this problem that's arguably harder to manage, for three reasons specific to industrial environments:


1. Continuous monitoring means continuous token consumption. Unlike an office worker who queries AI a few times a day, an AI-driven monitoring system watching equipment sensors, safety thresholds, or compliance conditions runs continuously — 24/7, across every line, every shift. That's a fundamentally different cost profile than "seat-based" software pricing most plants are used to budgeting for.


2. Compliance and safety use cases can't be the place you cut corners to save tokens.

One of the most consistent warnings from AI cost experts is that businesses should match the AI model to the task — don't use an expensive, powerful model for something simple. That's sound advice for general business tasks. But for safety-critical or regulatory functions, using a cheaper, less capable model purely to save on tokens introduces a different kind of risk entirely: a compliance summary or safety alert that's subtly wrong because it was generated by a model chosen for cost rather than accuracy. In our own work advising plants at Gammatek, this is the tension we see most often — teams under budget pressure default to the cheapest available model across the board, without separating "this can be a lightweight model" tasks from "this needs to be right every time" tasks.


3. Token estimation is unreliable — and that's a bigger problem on a factory floor. Research on AI cost estimation has found that predicted token usage for a given task can differ from actual usage by a factor of up to 30. In an office setting, that's an annoying budget surprise. In an industrial setting where AI is tied to safety alerting or compliance deadlines, unpredictable cost behaviour makes it much harder to build reliable, auditable systems — which matters enormously if a regulator later asks how a given alert or report was generated and what it cost to verify.


Typical Office AI Use

Industrial Plant AI Use

Usage pattern

Occasional, user-initiated

Continuous, sensor/system-initiated

Cost predictability

Moderate — tied to how often staff use it

Low — tied to plant uptime and monitoring scope

Risk of under-spec'd model

Lower-quality email or summary

Missed or inaccurate safety/compliance signal

Budget owner

IT/software budget

Increasingly split across IT, safety, and compliance budgets

Audit requirement

Usually none

Often required to justify AI-generated compliance decisions

What Australian Manufacturers Should Actually Do About It

Based on Gammatek's own client conversations and the broader guidance emerging from AI cost analysts, a few practical steps stand out for industrial operators specifically:

  • Separate "always-on monitoring" AI from "occasional task" AI in your budgeting. These have completely different cost trajectories, and lumping them into one software line item makes the bill impossible to forecast.

  • Match model tier to task risk, not just task complexity. A lightweight model may be fine for drafting an internal memo. It's a much riskier choice for anything feeding into a safety alert or compliance record — the cost saved is rarely worth the accuracy risk.

  • Ask any AI vendor how token usage is estimated for your specific use case, not just their published per-token price. As Australian AI experts have noted, even the AI systems themselves are often poor at predicting their own token consumption in advance — get this in writing where you can.

  • Build in an audit trail for AI-generated compliance or safety content, independent of the AI vendor. If regulators or auditors later ask how a decision or alert was produced, "the AI said so" isn't an adequate answer — you need documentation that sits outside the AI tool itself.

  • Treat AI cost review as an ongoing discipline, not a one-time budget line. Because usage tends to expand into new tasks continuously, a cost review that made sense six months ago may already be outdated.

This is precisely the gap between a plant's technology stack (the AI tools and monitoring systems themselves) and its compliance stack (the audit trail proving those systems are being used safely and correctly) — and it's a gap that's widening as more industrial AI tools come to market faster than the compliance frameworks around them.

Where This Is Heading

Australia's federal government has already flagged this as a national economic issue, not just a business cost problem — officials have pointed out that the overwhelming majority of AI spending flows offshore to overseas providers, effectively functioning as a new import bill for the country. For individual manufacturers, that macro trend translates into a simple operational reality: AI costs are not going to become predictable on their own, and the businesses that build proper cost governance and compliance documentation around their AI use now will be in a far stronger position than those scrambling to explain a surprise line item at the next board review.

 
 
 

Comments

bottom of page