EBIT-T -DA: AI's hidden line item, solving AI's growing cost problem with observability
Tue, 4th Aug 2026 (Today)
There is a new line-item reshaping the P&L of Australian enterprise, and most boards have not named it yet.
"EBITDA is now EBIT-T -DA" - earnings before interest, tax, tokens, depreciation and amortisation. The line surfaced first at the ADAPT Cloud and Infrastructure Edge and resurfaced through the day, including at a roundtable Avocado Consulting co-hosted with Dynatrace, attended by technology and operations leaders from banking, insurance, energy, healthcare, government, mutuals and retail.
With 84% of companies now reporting AI costs eroding gross margins by more than 6 percentage points, "EBIT-T -DA" is not just a clever line, it reflects the reality many organisations are living right now. The extra 'T' names a cost line that almost none of those organisations can yet quantify with confidence: tokens.
Part of the problem is a behaviour driving this: "token maxxing." As AI usage becomes a visible productivity signal - tracked on internal leaderboards, flexed in stand-ups, treated as proof of "leaning in" to AI - employees and teams are incentivised to maximise token consumption rather than optimise it.
More tokens starts to look like more output, more effort, more value. But token volume and business value are not the same curve. Without observability into what that consumption is actually producing, organisations are left funding a behaviour, not a result.
Tokens aren't even the biggest, unexpected cost today - data platform usage and networks currently top that list. That's not a reason to worry less about tokens; it makes it more concerning.
Everyone is cost-constrained. Few have an ROI story
Every organisation at our roundtable operated under cost constraint, whether driven by a regulator, a board, or a customer base that ultimately funds every dollar of spend. What differs is how each is answering for that constraint as AI enters the budget.
Several delegates pointed to a recurring gap: productivity gains are arriving quickly, but hard-dollar ROI is not. One delegate, from banking, was reluctant to attach a ROI figure to current AI initiatives, noting that time saved is typically reabsorbed by the same team, rather than removed from the cost base - evidence of productivity, but not of a saving that can be booked.
Other delegates offered counter-examples. One, working on a large-scale cloud transformation, described legacy troubleshooting that previously took three to five weeks, reduced sharply through AI-assisted analysis, which in turn compressed an end-to-end delivery cycle from 16–20 weeks to 8–10 weeks - a figure straightforward enough to present to a board.
The limitation is that these results are typically localised. A regulated, cost-conscious organisation running AI across dozens of teams has no consistent way to aggregate that value story; each team is operating its own baseline, on its own assumptions, with no shared measure of the total return. That inconsistency is now measurable at an industry level: In the 2025 State of AI Governance report, only 34% of companies say they have an "advanced" AI cost management program in place, defined as having tracking, cost attribution and governance policy all functioning together.
The question that exposes the gap: cost versus speed
Luke Napoli, Head of Enterprise Sales (ANZ) at Dynatrace put a direct question to the delegate who had described the compressed delivery cycle: would that outcome still hold if a cheaper, slower model were used instead of a faster, more expensive one? Specifically, would a third of the token cost be worth accepting a slower result?
The delegate's answer was that "it depends" entirely on the value on the other side of the trade-off - if a two-week time saving is worth more than the additional token cost, the faster model is justified; if not, it is not. The principle is straightforward. What was missing, for nearly every organisation represented, was a repeatable way to make that calculation, because few had real-time visibility into what any given model call was costing them, by use case and by team.
That gap is more structural than it first appears. According to a State of AI costs report, Enterprise spend on LLM APIs rose 36% in a single year, from an average of $62,964 to $85,521 a month, and it rose during a period when the underlying unit price of intelligence was falling faster than at almost any point in computing history. Independent research from Epoch AI, which tracks the price of matching a given level of model performance over time, found that the median rate of price decline accelerated to roughly 200x per year after January 2024, up from an already fast 50x per year before that. Falling unit prices are not translating into falling bills, because consumption is growing faster than price is falling: new features, longer context windows and multi-step agentic workflows all add token volume that a per-unit price cut cannot offset. That marks a real difference from the cloud cost optimisation exercises most IT organisations have already run: the unit economics move independently of anything the organisation does, and consumption compounds in ways a single pilot project will not reveal.
A defensible cost-versus-speed decision requires instrumentation that shows, in real time, what is being spent and what is being returned for it. Without it, "it depends" is where the analysis stops - and that unresolved question is a significant reason boards remain cautious about AI budgets.
It is also a circular problem. Observability data is needed to build the ROI case that secures funding, but funding is needed to build the observability layer in the first place. Because that instrumentation work is unglamorous, it is routinely deprioritised, and organisations proceed with AI deployment ahead of the cost visibility that should have preceded it.
The foundation AI needs, and what happens without it
Two examples illustrate the point from opposite directions.
The first - an Avocado client - involved a Western Australian energy provider, where multiple operational teams were each running their own dashboard for critical infrastructure, with no shared view across the organisation. Consolidating
The second example at the roundtable shows what happens without that foundation. An organisation with several overlapping monitoring and data tools opted to put a general AI layer across all of them rather than consolidate the underlying data first. Without a single, trusted source of truth for what was actually happening across the environment, the AI system could not reliably distinguish correlation from causation. Acting on its own reasoning, it began making changes that took applications down rather than resolving the issues it was meant to address. The failure was not the model itself; it was the absence of AI-specific observability capable of validating whether the AI's conclusions were grounded in accurate, unified data before it was allowed to act on them.
Read together, the two cases make the same point: AI observability is not something layered on after the fact. It depends on the same consolidated, trustworthy data foundation that good observability practice already requires. Without it, an AI system is reasoning from incomplete information, with consequences that can be operational as well as financial.
The cost of not knowing
The financial risk compounds when trust in AI's output is also in question. Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027, attributing this to escalating costs, unclear business value and inadequate risk controls - notably, not to any shortfall in model capability. A separate analysis of that same failure pattern names the absence of observability as one of the recurring technical causes behind those cancellations.
Dynatrace's State of Observability research points to the same gap from the buyer's side. Observability budgets are rising sharply - 70% of organisations increased spend this year and 75% plan to increase it again next year, with AI capability now the leading criterion in platform selection, ahead of cloud compatibility. Yet only 28% of organisations currently use AI to connect observability data back to business outcomes. Spend on visibility is rising faster than the ability to translate that visibility into a business case a board will accept.
As Dynatrace has put it in its own analysis, AI changes what "healthy" means: a system can be running perfectly and still produce a wrong answer, breach a policy, or quietly burn money at scale. This is why token spend, hallucination rates, governance and compliance need to sit within the same observability discipline rather than separate ones. A platform that reports spend without reporting the reliability of the output is only answering half the cost question.
Where this leaves technology leaders
The practical implication is straightforward: AI observability should be treated as the infrastructure that makes an AI cost case provable, not as overhead alongside it. That infrastructure is under growing strain - AI workloads have driven a 93% increase in log volume over the past twelve months alone, which makes consolidating fragmented monitoring and data platforms progressively more urgent, not less. For AI specifically, that consolidation is the difference between knowing where token spend is going and discovering it after the fact.
The organisations making genuine progress are not the ones with the most ambitious AI programs. They were the ones prepared to fund instrumentation and baselining ahead of deployment, rather than after the invoice arrived.
That's the gap AI observability closes. It's the difference between "it depends" and knowing.
Want to see where your AI, data and network spend is actually going - and how to optimise it? Book a free observability discovery