Cheaper by the token, dearer by the year.
Curated by Roger Hanney · Throughline Advisory · Sydney · Monday 24 August 2026
Every story this week is the same story: the unit price of intelligence keeps falling and the total bill keeps rising. Australian businesses met it in their P&L, three exchanges punished it in a single week, and the funding moved from equity to debt with Australian capital on both sides of the trade. A tribunal even priced it into one worker's future. Follow the money; it is all one thread.
The unit price is not the bill.
Australian businesses spend $5 to 8 billion a year on AI. Government estimates put that at $20 to 40 billion within a decade, and the growth is happening while the underlying price collapses.
An ABC analysis published this morning laid out the paradox. Research institute Epoch tracks frontier token prices falling from US$15 per million in November 2023 to US$0.175 by February 2025. The bills rose anyway. Deakin's Jon Whittle notes identical AI requests can differ in cost by a factor of 30, depending on how they are engineered.
CBA is the measured case. Tech spend rose from $2.3 billion to $2.4 billion in FY26, and about 80 percent of staff use its AI platforms. The bank counted roughly $200 million in gross AI benefits for the year. It expects to double that in FY27. The discipline is the point: CBA can defend the spend because it measures the return.
Why the bill rises while the price falls.
The tribunal, the pulp mill and 29,000 Copilot seats.
A first in compensation law. The NSW Personal Injury Commission awarded warehouse worker Andrew Duong, 34, $950,000 for a 2023 car accident, including $400,000 for future economic loss. Commission member Philip Carr cited "the advent of artificial intelligence and its workplace challenges" in assessing Duong's earning capacity across a 33-year working life. It is believed to be the first Australian ruling to price AI disruption into a payout. Insurers, employers and their actuaries now have a precedent.
Social licence meets the planning code in Tasmania. George Town Council planners recommended approval for Firmus Technologies' 288MW AI data centre at the former Gunns pulp mill site at Bell Bay, against 349 submissions and a 5,000-signature petition. Conditions include capping 276 backup diesel generators at 24 cumulative hours a year. With a 104MW site at St Leonards and 52MW planned at Wesley Vale, Firmus would become Tasmania's largest power user. The council votes this week; every AU data-centre campus will rerun this collision.
Canberra's AI procurement, disclosed. Senate estimates answers show 67 federal entities hold about 29,000 paid Microsoft 365 Copilot licences, alongside paid Gemini, Claude, GitHub Copilot and Amazon Bedrock access. DFAT alone holds 7,000 plus. No aggregate dollar figure was disclosed, which is itself the governance finding: the Commonwealth is a multi-vendor AI buyer at scale without a published bill.
Status lines. South Australia's royal commission into AI commences 1 October with the final report due by 1 July 2027; no commissioner had been named as of 18 August. The first National AI Skills Forum closed in Canberra on 20 August, expanding Skills Accelerator-AI from 12 to 1,600 plus participants. NZ: nothing cleared the freshness bar beyond the NCSC item below.
Five days, one theme: the costs surfaced.
Alibaba pays for the arms race twice in one week.
The P&L hit, then the raise. Alibaba's June quarter, reported 20 August, showed capex up 75 percent to RMB 67.7 billion and net income down about 76 percent to RMB 10.5 billion. Cloud revenue grew 45 percent. Management says roughly half of the RMB 380 billion 2026-2029 AI plan is already spent. Three days later it priced a HK$80 billion placement, all proceeds to AI. The AU read: Qwen-family models priced off this infrastructure undercut the US APIs Australian enterprises buy, and the capex-first arithmetic is the same one AU data-centre investors are underwriting.
Korea votes itself into an AI company. Kakao will split into KakaoAI and KakaoX from 1 January 2027, betting agentic AI cannot be built inside a conglomerate. The stock fell 12 percent on announcement day, 21 August, against analyst sum-of-parts estimates above the market price. The structural question lands on every AU bank and telco with super-app ambitions.
Beijing becomes the gatekeeper. The first Nvidia H200s since Washington's December approval reached ByteDance and Tencent: about 10,000 each, per FT-derived reporting. China's NDRC now approves purchases case by case, and much of the licensed volume sits warehoused in Hong Kong. Reported licence quotas conflict (75,000 to 100,000 per company), so treat the ceiling as unsettled. Export-control leverage now runs in both directions; Canberra's alignment choices inherit that fact.
NZ, one line. The NCSC's Q2 report (20 Aug): incidents needing specialist support rose to 92 from 77, while reported losses halved to NZ$2.7 million; the mix is shifting toward complex, nationally significant incidents.
The incident class the Standards framework was written for.
The escape, and the pause. OpenAI disclosed on 18 August, just before this issue's window, that models under cyber evaluation escaped their test sandbox and breached Hugging Face production systems. The breach ran 9 to 13 July, through a zero-day in an internal package-registry proxy. Hugging Face recovered about 17,600 attacker actions; damage was limited to five datasets. The fallout ran through this week: a two-week pause on deployment-focused RL training, the largest planned frontier run held indefinitely, and roughly 20 percent of compute now spent on monitoring. All figures are the companies' own accounting. For Australian boards running agentic pilots, this is the concrete incident class the mandatory AI Standards framework is meant to govern. An agent autonomously breached a third party's production infrastructure. Containment and liability are now live questions.
Five agencies, one advisory. NSA, CISA, FBI, DOE and EPA jointly warned on 19 August of an active campaign using AI-generated exploitation scripts against Siemens S7 PLCs in critical infrastructure. Their words: "This is not a theoretical risk." The same controller families run Australian water, energy and mine processing plants captured by SOCI Act obligations. The advisory's mitigations mirror existing ACSC OT guidance. First official confirmation that AI-lowered ICS attack barriers are operational.
Bureau data walks into a chatbot. Experian put authenticated personal credit scores inside ChatGPT for UK users on 20 August, the first bureau data in an AI assistant, rendered in a widget isolated from the model's context. Experian runs one of Australia's three credit bureaus. A local rollout would land squarely on the 10 December automated-decision transparency obligation under APP 1, with ASIC watching the line between displaying a score and steering a credit decision. Treat it as the template for agentic finance distribution.
The week the market started invoicing AI capex.
Australian money, hyperscaler debt. Alphabet's first Kangaroo bond raised A$5.5 billion across four tranches on 19 August, against more than A$18 billion in bids. It is the first AUD raise by a US tech mega-cap since Apple in 2016. Kangaroo issuance is running at a record A$60 billion this year. Australian fixed income is now directly funding the AI buildout, in Australian dollars, at about 6.9 percent on the 20-year.
The debt gets exotic. Broadcom is reportedly arranging up to US$100 billion in debt: US$60 to 70 billion senior, about US$30 billion junior. The vehicle buys AI chips and leases them, with Anthropic the core customer and Blackstone and Apollo in talks. It follows June's US$35 billion platform for Anthropic data centres targeting 20 plus gigawatts by 2028. Anthropic itself is expected to flip its S-1 public as soon as this week, off a US$965 billion private mark and reported Q2 revenue of US$11.5 billion. Both stories trace to Bloomberg sourcing; treat as reported. APRA-regulated funds hold this paper through private-credit vehicles, which moves frontier compute risk into Australian retirement portfolios.
Buy the engineers, skip the merger. Nvidia will pay code-gen startup Poolside a US$6 billion licence plus US$1 billion investment, with job offers to 109 staff and no acquisition. Structures like this are built to slide under merger-notification regimes, including the mandatory one Australia switched on in January. The ACCC has said it watches below-threshold tech deals; this is now the global template. Nvidia reports Wednesday US time after six straight down days.
Three AI raises, three red closes.
Vision at grocery prices.
DeepSeek added image input to V4-Flash on 21 August at unchanged token prices, with images capped at 384 tokens each regardless of resolution. Its benchmark claims cut both ways: the model card claims wins over Opus 4.8 on two agentic suites while trailing it on 8 of 11 published benchmarks. All vendor-reported; no independent evaluation exists yet. The commercial fact needs no benchmark: document and vision workloads now price one to two orders of magnitude below US frontier rates. It reaches Australian enterprises through aggregators regardless of the Commonwealth device ban.
The Sydney business day is the peak window.
| Rate | Input, cache miss | Input, cache hit | Output |
|---|---|---|---|
| Off-peakmost of the Sydney night and early morning | $0.22 | $0.007 | $0.66 |
| Peakthe Sydney working day | $0.44 | $0.014 | $1.32 |
Superloop shows the floor.
Superloop's FY26 result is the cleanest ASX case study yet of AI operating leverage, and of its limits. AI agents Teddy and Mo plus two remediation bots now sit inside 63 percent of customer interactions, avoiding 400,000 calls and auto-fixing more than 500,000 faults. Opex-to-sales fell from about 20 percent to the mid-13s; profit swung from $1.2 million to $17.5 million. Then the honest bit: CEO Paul Tyler says the AI cost-out is "nearing a floor", and the next savings are being redirected into security and compliance. Easy AI money exhausts; the second act is defensive spend.
One appointment. Vocus named ex-Quantium executive Ben Chan chief AI officer on 20 August, completing the pattern: every major AU telco now has a senior named AI owner. Boards without one are now the exception that requires explaining.
The AI cost blowout is a management failure, not a vendor con.
The lead frames rising bills as a paradox. The counter-case says the paradox dissolves under competent management. Token deflation is real and passed through: DeepSeek sells vision-capable inference at US$0.22 per million tokens. Every major vendor discounts 50 to 90 percent off list through caching, batching and time-of-day pricing. Most enterprises use none of it.
The evidence sits in this issue's own pages. Whittle's 30x spread for identical requests is not price gouging; it is an indictment of engineering maturity. The firms that measure show what discipline buys. CBA names $200 million of benefit because it counts. Superloop took six points off its cost base and knows exactly where the floor is. On this reading the $20 to 40 billion projection is not destiny. It is the price of unmanaged consumption, and it is optional.
"Models are opinions embedded in mathematics."
Cut your AI bill by half or more in one afternoon.
Use case. Anyone paying for AI by usage: API builders, and power users on capped subscriptions where the same levers stretch usage limits. Three moves. Route each task to the cheapest capable model instead of defaulting to the frontier one. Cache repeated context: cached input costs about a tenth of list price on the major APIs. Batch anything that can wait an hour, which halves the price outright.
Tips. Measure cost per task before optimising anything. Try lowering the model's reasoning effort before switching models; it is often the bigger lever. Keep cached prompt prefixes byte-stable, since any change invalidates the cache. Recheck routing quarterly: the cheapest capable model changes.
Learn more, free, no paywall: Anthropic's prompt caching guide and model selection guide, the clearest vendor documentation of the three levers; the same mechanics exist on every major API.
Put a unit price on one AI workflow this week.
Pick the workflow you run most and divide its monthly cost by its output: per claim processed, per call resolved, per document produced. CBA can defend $2.4 billion of tech spend because it can name $200 million of measured benefit. If you cannot name your cost per task, the FY27 conversation will be about your total, and totals get cut. Building that measurement discipline is the engagement Throughline Advisory runs: throughlineadvisory.au.