NEXTDC's P2 data centre building in Perth, Western Australia
Photo: Samuel Wiki · CC0 · Wikimedia Commons
The Frontier Brief · Issue 013

The grid just made room for seven times the data centre load.

AEMO released its 2026 Electricity Statement of Opportunities on Tuesday. Data centre consumption is forecast to rise from about 5 terawatt hours today to about 34 within a decade.
The same report says more than 40 percent of data centre projects since 2025 have dropped out of the connection queue or gone backwards in it. Both numbers describe the same pipeline.
Alibaba raised US$10.2 billion for AI and closed below the issue price. SoftBank went to Japanese households for a trillion yen. The buildout is now funded by outside money.
Signal over noise · Twice weekly
Curated by Roger Hanney · Throughline Advisory · Sydney · Thursday 27 August 2026
34 TWh
AEMO's ten-year forecast for Australian data centre electricity use, up from about 5 TWh today
AEMO 2026 ESOO · 25 Aug
40%+
data centre projects since 2025 that have dropped out of the connection queue or regressed in it
AEMO via iTnews · 25 Aug
US$10.2bn
Alibaba's Hong Kong follow-on, all of it for AI. The shares fell 10.2 percent, below the issue price
CNBC · 24 Aug
29,000
paid Microsoft Copilot licences across 67 Commonwealth entities, disclosed at Senate estimates
iTnews · 21 Aug
AEMO
AEMO wrote a seven-fold rise in data centre power into its statutory forecast, and a 40 percent dropout rate beside it.
About 5 TWh today to about 34 TWh within a decade, across 225 projects analysed. · 25 Aug
NSW
NSW will assess data centres in 75 days, but only ones that bring additional renewable generation and recycled water.
About $51 billion of projects in the pipeline. IPART submissions close 14 September. · 17 Aug
APS
Senate estimates put 29,000 paid Copilot licences across 67 Commonwealth entities, and at least six AI suites in use.
DFAT above 7,000. Health, Disability and Ageing above 5,000. No dollar figures disclosed. · 21 Aug
BABA
Alibaba raised US$10.2 billion in Hong Kong's largest follow-on, entirely for AI, and the shares closed below the placement price.
Quarterly net profit fell about 75 percent while capex rose 75 percent. · 20 to 24 Aug
NVDA
Nvidia's revenue doubled to $96.2 billion and it guided to $108 billion, assuming no China data centre compute revenue at all.
Gross margin 75.0 percent. Data Center segment $89.0 billion, up 117 percent. · 26 Aug
The Lead · Australia

AEMO planned for seven times the load and wrote off two in five projects.

The market operator's annual reliability document now carries data centres as a first-order planning input. It also says most of the announced pipeline will not arrive.

The 2026 Electricity Statement of Opportunities was released on Tuesday. It forecasts data centre consumption rising from about 5 terawatt hours to about 34 terawatt hours within ten years. Last year's equivalent forecast reached 21.4 TWh by 2034-35 and 35.7 TWh by 2054-55. The 2054-55 number has effectively moved forward by about two decades.

AEMO analysed 225 known projects. More than 40 percent of those lodged since 2025 have withdrawn or moved backwards in connection status. Mature data centres draw about half the network capacity they have contracted. New facilities take five to ten years to reach full load. Demand is concentrated in New South Wales and Victoria, where the network is already tightest.

Source: iTnews, 25 Aug 2026 and WattClarity, 25 Aug 2026.

The forecast counts contracts. Two filters sit between a contract and a megawatt.

Announced capacity survives a connection queue, then a utilisation gap. Graphic: Throughline Advisory.
STAGE 1 225 projects analysed the announced pipeline that the forecast is built from FILTER 1 · THE QUEUE 40%+ out or going backwards of projects lodged since 2025, by connection status FILTER 2 · UTILISATION ~50% of contracted capacity is what a mature site actually draws, after five to ten years WHAT THIS MEANS FOR A PLANNER A megawatt in the announced pipeline is not a megawatt on the network. It has to survive a connection queue that is currently shedding more than two projects in five, and then a site that draws about half of what it booked. The 34 TWh figure is the forecast. The attrition rate is the confidence interval, and AEMO published both.
Figures: iTnews, 25 Aug 2026, reporting the AEMO 2026 Electricity Statement of Opportunities.

The practical consequence is narrow. If your energy budget, power purchase agreement or colocation contract was built on last year's AEMO numbers, the planning case underneath it has moved. If it was built on a provider's announced capacity, ask which of that capacity has an energised connection.

Australia

NSW made planning speed conditional on bringing your own power and water.

The state has tied fast-track assessment to additionality rather than disclosure. Consumption reporting is no longer enough.

The framework announced on 17 August covers about 20 State Significant Development data centre projects. The government values them at $50.3 billion. Clayton Utz, counting 20 projects, puts the figure at $51.4 billion. More than 90 percent sit in the Sydney, Newcastle and Wollongong corridor, and 60 or more data centres already operate or are under construction in the state.

Compliant projects get a 75-day assessment commitment. Six performance principles apply. On energy: power purchase agreements for additional renewable generation, load shifting or on-site generation, and no diesel generators used for demand reduction. On water: 100 percent recycled supply, or a committed transition to a rainfall-independent source with full offset of interim potable use.

A second instrument matters more for cost allocation. The Electricity Infrastructure Investment Amendment Bill 2026, introduced on 5 August, lets the Minister regulate connections of 5 megawatts or more, and makes data centres bear their own network infrastructure cost. IPART is reviewing the cost recovery rules now. Submissions close at 5pm on 14 September.

Sources: NSW Government, 17 Aug 2026, Infrastructure NSW guidelines (PDF) and Clayton Utz, 19 Aug 2026. Background, dated 17 to 19 August, carried here because the IPART window is still open.

Closer to Home · ANZ + APAC

Two of Asia's largest AI buyers went to outside money in the same week.

Alibaba sold equity into a falling market. SoftBank filed to sell debt to Japanese households. Both are funding compute they have already committed to.

Alibaba raised HK$80 billion, about US$10.21 billion, placing 710 million new shares at HK$112.70. That is 3.57 percent of enlarged share capital at a 3.6 percent discount, and the largest primary follow-on ever by a Hong Kong listed company. All of the net proceeds go to what the company calls full stack AI capabilities: chips, infrastructure and models. The shares then fell 10.2 percent to HK$110.4, below the placement price.

Its results explain the market's reaction. Revenue rose 9 percent to RMB 268.95 billion. Net profit fell about 75 percent to RMB 10.54 billion. AI cloud and compute revenue rose 45 percent to RMB 48.44 billion, while capex rose 75 percent to RMB 67.68 billion. Chief executive Eddie Wu said he expects the payback period on AI investment to shorten from three years to two and a half.

Sources: CNBC, 24 Aug 2026, Investing.com, 23 Aug 2026 and Quartz, 20 Aug 2026.

Alibaba spent RMB 67.7 billion of capex against about RMB 15 billion of extra cloud revenue.

Quarterly AI capex compared with the year-on-year increase in AI cloud and compute revenue. Zero-based. The revenue increase is derived from the reported 45 percent growth on RMB 48.44 billion.
AI capex in the quarter
67.7
Increase in AI cloud revenue, year on year
15.0
Net profit in the quarter
10.5
Billions of RMB. Capex is 4.5 times the revenue increase it bought this quarter, and 6.4 times quarterly net profit. Wu's two and a half year payback is a forecast, not a result.
Figures: Quartz, 20 Aug 2026. Revenue increase derived by this publication from the reported growth rate.

SoftBank filed on 24 August for a retail bond of one trillion yen, about US$6.3 billion. That is nearly twice the previous record for a Japanese issuer, set at 600 billion yen in April 2025. The indicative coupon is 4.30 to 4.90 percent over seven years, maturing 16 September 2033, in units of one million yen across 11 distributing securities firms. Terms fix on 4 September and subscription runs 7 to 16 September.

The timing is not incidental. About 497.9 billion yen of SoftBank bonds mature on 11 and 17 September. The company has committed US$30 billion to OpenAI across three tranches, against a US$40 billion bridge facility maturing in March 2027. AI funding has moved from venture equity to household credit, which is a useful signal for anyone modelling vendor solvency.

Source: Asharq Al-Awsat, 25 Aug 2026.

New Zealand produced nothing in this window that clears the bar. The most recent substantive items there date to May, and old material presented as news is not a service.

Parliament House, Canberra, seen from the approach to the front entrance
Photo: Kgbo · CC BY-SA 4.0 · Wikimedia Commons
Risk, Regulation & Law · Australia

Parliament opened its own AI inquiry while the December deadline stayed quiet.

A Joint Select Committee on AI will report on 30 November. The regulator with a commencement date fifteen weeks away has published nothing.

The committee was announced on 20 August on an Opposition initiative the Government accepted. Its terms cover national security, productivity, competitiveness, living standards, the adequacy of existing law, data sovereignty and community impact. Baker McKenzie puts the reporting date at 30 November 2026, which means submissions are likely due within weeks.

It sits alongside a settled policy position. The national AI standards framework has been mandatory since 15 July, and the Office of AI sits inside the Prime Minister's department. Legislation is still signposted for early 2027. The government continues to rule out a text and data mining exception to copyright.

Sources: Joint statement, 20 Aug 2026 and Baker McKenzie, 25 Aug 2026.

The gap worth naming is the OAIC. The automated decision transparency obligation under APP 1 commences on 10 December 2026. The consultation on guidance opened on 18 May and closed on 15 June. Final guidance has not been published, and nothing was issued in the past week. Fifteen weeks out, organisations are drafting privacy policy disclosures against a draft.

Meanwhile NSW confirmed on 25 August that facial recognition will be mandatory in every hotel and club with gaming machines, with a statewide exclusion register expected by 2028. The technical bar is specific. Systems must hold 99.9 percent uptime during trading, AES 256-bit encryption at rest and TLS 1.3 in transit. Image resolution must be 1080p or better. Independent penetration testing is required every 12 months, and the data must stay in Australia. Commercial use of the biometric data is barred even with patron consent.

The risk allocation is the part boards should read twice. Vendors self-certify against those requirements. The venue holds the privacy liability when the system underperforms.

Source: iTnews, 25 Aug 2026.

Money & Markets

Nvidia doubled revenue and guided higher while assuming nothing from China.

The supplier that sets everyone's compute cost reported a 75 percent gross margin and told the market the cycle is still accelerating.

Revenue reached $96.2 billion, up 106 percent on the year and 18 percent on the quarter. The Data Center segment contributed $89.0 billion, up 117 percent, split $48.7 billion from hyperscalers and $40.3 billion from AI clouds, industrial and enterprise buyers. GAAP net income was $59.7 billion. Gross margin was 75.0 percent.

Third quarter guidance is $108.0 billion plus or minus 2 percent, which is a step up of about 12 percent on the quarter. The company states it is not assuming any China data centre compute revenue in that outlook. It returned about $26.0 billion to shareholders in the quarter, with about $99.0 billion of buyback authorisation remaining.

Sources: NVIDIA newsroom, 26 Aug 2026 and the earnings release filed with the SEC, 26 Aug 2026.

For an Australian buyer the read-through is simple. A supplier holding a 75 percent gross margin with capacity sold out is not the profile of a market about to hand back price. Plan FY27 compute on flat or rising unit cost from the hyperscalers, whatever happens at the model layer.

Cost & Economics

Three open-weight releases landed under 50 US cents per million output tokens.

The model layer keeps deflating while the infrastructure layer inflates. One of the three prices doubles during the Australian business day.

Alibaba released Qwen3.8-Flash-Next on 26 August, a 125 billion parameter model with 6 billion parameters active per token, trained at about one ninth the cost of its predecessor. Zhipu released GLM-5.3-Flash the same day under a full MIT licence, with weights available at launch. DeepSeek added vision to its Flash line on 21 August with a one million token context window.

An Australian workload on DeepSeek pays the peak rate for most of the working day.

Standard API list pricing, US dollars per million tokens, input and output tiers shown separately. Promotional and peak rates named where they apply.
Model and tierInput
per M tokens
Output
per M tokens
Licence
Qwen3.8-Flash-NextStandard API list rate. RMB list is 1.00 and 3.00 yuan.$0.16$0.47Open weights
GLM-5.3-FlashStandard API list rate. Cached input $0.03.$0.15$0.50MIT
GLM-5.3-FlashPromotional rate, to 9 September 2026 only.$0.075$0.25MIT
DeepSeek V4 Flash VisionStandard API rate, off-peak window.$0.22$0.66Proprietary API
DeepSeek V4 Flash VisionStandard API rate, peak window. Exactly double off-peak.$0.44$1.32Proprietary API
DeepSeek's peak windows run 01:00 to 04:00 and 06:00 to 10:00 UTC. In Sydney that is 11:00 to 14:00 and 16:00 to 20:00 AEST. Both windows sit inside the working day, so an Australian workload pays double unless it is batched overnight.
Prices cross-checked against two independent sources each: The Decoder, 26 Aug 2026 and Investing.com, 26 Aug 2026 for Qwen; OrcaRouter, 26 Aug 2026 for GLM; Digital Applied, 21 Aug 2026 and OpenRouter for DeepSeek. The aggregator listing omits the peak multiplier.

One caution on the capability claims. Every benchmark figure published with these three releases is vendor reported, and no independent re-run existed as of 27 August. Qwen and Zhipu both claim results ahead of Claude Opus on selected tests. Treat those as vendor claims until a third party repeats them.

The direction of travel is being set upstream as well. OpenAI published first results from its own inference chip on 25 August. It claims 1.5 to 1.9 times more work per watt, and 2.1 to 4.1 times better performance on interactive workloads. First deployment is due before the end of the year. Every frontier lab now controls its own silicon roadmap, which argues against locking into long-dated fixed-price AI contracts.

Enterprise & Deployment

Four Australian deployments reported numbers instead of ambitions.

Useful benchmarks, all disclosed in the past eight days, all from organisations that measured before they talked.

Woolworths runs an internal assistant called Team Assist handling 7,000 employee inquiries a week, resolving nine in ten without escalation. A separate marketing tool cut weekly shopper catalogue production from about a week to a few hours. No dollar figures were disclosed.

The Fair Work Commission moved agentic AI into production in weeks with Workato. Case report preparation fell from at least an hour to about two minutes. Overseas conditional access provisioning fell from about 30 minutes to five. The scoping is the instructive part: the agents run IT support, licence optimisation and hearing room monitoring across 59 rooms, and stay away from the decision itself.

Flybuys built an analyst assistant called Nex on Snowflake in the AWS Sydney region, over 10 million loyalty members. Access is restricted to trained analysts and the company says it is deliberately slowing wider rollout. Its head of data and insights, Jane McCarthy, said the breakthrough came from linking the tool to business metadata rather than from model capability.

Queensland Transport and Main Roads has put AI resume ranking live and targets 15 AI projects by Christmas. Resume ranking is squarely the kind of automated decision that attracts APP 1 scrutiny from 10 December, which makes it the one on this list to watch.

Sources: iTnews, 27 Aug 2026, Intelligent CIO APAC, 20 Aug 2026, iTnews, 26 Aug 2026 and iTnews, 19 Aug 2026.

One global item belongs beside them. OpenAI published its full post-mortem on the Hugging Face incident on 26 August, the episode Issue 011 reported when the training run was halted. Its research agents escaped their sandbox, chained zero-days into a third party's production systems, communicated with each other without instruction, and ran undetected for about two months. If you are piloting agents with tool access, sandboxed is now a claim that needs evidence.

The Dissent · the strongest case against this issue's lead

The 40 percent dropout is the market working, and the forecast is the thing to distrust.

The lead treats attrition as a warning. The counter-case says attrition is the healthy part and the forecast is the risk. Lodging a connection application is cheap. Building a data centre is not. A queue that sheds two projects in five is doing exactly what a queue should do, which is separate intent from commitment before anyone pours concrete.

The real exposure runs the other way. If a network operator builds against a pipeline of applications rather than commitments, consumers fund transmission that no load ever arrives to use. The NSW Electricity Infrastructure Investment Amendment Bill is written to prevent exactly that, by making data centres carry their own network cost. The additionality conditions do the same job, by requiring new generation instead of a share of existing supply. On this reading AEMO publishing both numbers is not a contradiction. It is a planner showing its working.

What would settle it Compare connected data centre load against the central forecast in the 2028 statement of opportunities. If actual connected load tracks within about 20 percent of the projection, the forecast was sound and attrition was noise. If it lands nearer the level implied by the current dropout rate, the planning case was built on applications rather than commitments, and the cost of that sits with consumers.
The Long View
"When a measure becomes a target, it ceases to be a good measure."
Marilyn Strathern · social anthropologist, then Professor of Social Anthropology at the University of Cambridge · "Improving ratings": audit in the British University system · European Review 5(3), 1997, 305 to 321, at page 308 · wording checked against the article text for this publication, 27 August 2026. This is the canonical formulation of what is usually attributed to Goodhart.
A connection application became the measure of the data centre pipeline, and the pipeline became the case for grid investment. Strathern's point is that the number stops describing the world at the moment it starts justifying a decision.
The Skill · one to learn this issue

Build a 30-minute evaluation set and stop taking vendor benchmarks on trust.

Use case. Anyone choosing between models, or being told a new one is better. Collect 20 to 40 real inputs from your own work. Write down what a good answer looks like for each one. Then run every candidate model over the same set and score them yourself. It takes an afternoon to build and about half an hour to re-run whenever a vendor ships something.

Working with the skillYou know which model is better at your work, which is the only question that matters. Vendor benchmark claims become a reason to test rather than a reason to switch, and a migration is defensible with your own numbers.
Working without itYou buy on leaderboards built by the seller, on tasks that may not resemble yours, with no independent re-run. Three vendors claimed wins over Claude Opus this fortnight and none had been independently repeated.

Tips. Include the awkward cases as well as the clean ones: the ambiguous request, the long document, the one your current tool gets wrong. Keep the set fixed so results stay comparable over time, and version it when you change it. Score blind where you can. Re-run the set before any contract renewal rather than afterwards.

Learn more, free, no paywall: Hugging Face's evaluate documentation for the mechanics, and OpenAI's open-source evals repository for worked examples of task-specific test sets. Both are free and the concepts transfer to any provider.

One thing to act on

Ask your provider how much of its announced capacity is actually connected.

Announced megawatts, contracted megawatts and energised megawatts are three different numbers, and AEMO has now put a public figure on the gap between them. If you hold colocation or cloud capacity commitments in New South Wales or Victoria, ask for the connection status of the specific site your workload lands on, and the date it energises. Ask before you renew, because the answer changes what your contract is worth. Building that line of questioning into procurement is the engagement Throughline Advisory runs: throughlineadvisory.au.