Energy Consumption & the Data-Centre Boom + Australian AI Startups

Eva Trade Desk × Tam Research — joint study brief (10 Sep 2026)

Authors: Eva (desk data + listed-market lens) · Tam (health/tech/energy research) — findings registered to vault Research/trade-desk/ and Research/tam/tech

PART ONE — Energy: the numbers are staggering (AEMO 2026 ESOO, 25 Aug)

Today: ~165 operating data centres on the National Electricity Market consume ~5 TWh/yr = 3% of total NEM consumption. Average monthly demand grew 200MW (2018) → 700+MW (2026).

FY26 demand split by state:

• NSW 66% (Sydney = Australia's data-centre capital)
• Victoria 30% · SA 2% · Qld 1% · Tas <1%

The forecast (Step Change scenario):

• Effectively as much power as every home in NSW + Victoria combined (38 TWh) by 2036
• Total NEM consumption +40% in a decade (176 TWh → 250 TWh), while ~15 GW of coal/gas exits (Yallourn 2028, Eraring + Gladstone 2029, Loy Yang A 2035) — the demand curve crosses the retirement curve
• A year ago AEMO didn't see this level until after 2050 — one year later, pulled ~15 years forward

Reality checks AEMO flags (important for skeptics):

• 36% of 2025-listed projects cancelled (even with grid agreements!)
• Operating centres used only 26% of their contracted capacity in Q1 2026
• Either number says: model the pipeline with a high haircut — but even haircuts double the load

Policy: national cabinet settled rules requiring new centres to guarantee new renewable generation + firming (Albanese, Aug 25) — grid integration, water, and "locations appropriate" constraints now attach.

Market response — the buildout is real money

NEXTDC (ASX: NXT) FY26 (27 Aug):

• Contracted utilisation 524 → 740MW (+202%) — largest contracting year in history
• Forward Order Book 565MW (>3.2× billing); every MW is a binding contract
• Billing utilisation 175MW (+58%); 197MW converts to billing in FY27, +221MW in FY28 (74% of FOB in 2 years)
• Contracted EBITDA >$1.0bn already in contracts vs FY26 EBITDA $235-248m — the hockey stick is contracted, not hoped
• FY27 guide: revenue $615-640m, EBITDA $385-410m (both +50%+); capex $5.25-5.75bn
• Geography confirms AEMO: S4 Sydney 350MW, M3 Melbourne 225MW, M5 Melbourne 1.2GW planned — NSW/Vic again
• Liquidity $8.7bn; capex $3.4bn FY26 overshot guidance by $400m (accelerated delivery)
• Desk note: NXT was added to the S&P/ASX 50 in the Sep-2026 rebalance (with MIN, replacing COH/WTC) — index money now chases this theme
• Stock: ~$11.35 (−3.1% today, from highs ~$14+). Fwd EV/EBITDA rich; the bet is FOB conversion

Firmus (private, pre-IPO): AI-factory builder, $500m raise at ~A$10.5bn valuation (August, led by Nvidia/Coatue, fourth raise inside a year: $1.9bn → $5.5bn → $10.5bn). 150MW under construction in Tasmania (Project Southgate, Launceston — world-first AI Factory Zone, cheap renewable power is the location logic: hydro). Q4-2026 IPO window flagged, ~A$420m raise. The Tasmanian angle matters: renewable-rich grid beats metro congestion.

Gridsight (Sydney): US$26m Series B (Insight Partners) — AI digital-twin of distribution networks; customer Endeavour Energy used it to double solar export capacity 5kW→10kW → 600MW unlocked without new wires. The AI-on-the-grid play.

Metacognition AI (Adelaide/ANU): A$10m pre-seed (Main Sequence/CSIRO) — 7-10× the Australian median pre-seed — for persistent memory + reasoning layers under existing frontier models. Van den Hengel: "We're not building frontier models, we're building better ways to use them."

PART TWO — The Australian AI startup market (the pattern)

Where the money concentrates (Sept snapshot):

• AI infrastructure >> AI models. Firmus (factories) + Gridsight (grid twins) + Metacognition (AI memory/OS layer) — the local VC thesis is "picks and shovels + application layers", explicitly NOT competing with OpenAI/Anthropic on frontier models
• Main Sequence (CSIRO-backed) is the axis: Firmus, Metacognition, Number 8 Bio all on its book
• Deal count collapsed, cheque size exploded (Airwallex-type megadeals dominate): 53 disclosed FY-rounds (down from 156 at peak) but $958m to one company's stack — barbell: huge infrastructure cheques / broken seed pipeline
• Seed→Series A gap widened (2.75× → 6.3×) — seed is hard, scale-ups get 9-10-figure rounds
• Exit path finally real: Firmus IPO Q4-2026 = the first big local AI-infrastructure listing; watch it (EV/EBITDA will be extreme; it's a theme-print not a value-print)

ASX-listed exposure map (what trade-desk users can act on):

🎯 Eva's Read

1. The uranium-data-centre knot is now explicit. AEMO: 13% of the grid by 2036 vs 15GW of coal exiting — Australia needs firm 24/7 baseload, exactly where our uranium thesis sits. Nuclear re-entry (Coalition policy path) + data-centre load = the macro tailwind behind PDN/DYL/BOE beyond spot prices. This study connects our theme to the country's biggest infrastructure story.
2. NXT is the listed expression — contract book is binding, FY27 +50% growth is real; valuation is the risk (rich on EV/EBITDA vs the conversion risk of a 3.2× order book). Post S&P-50 inclusion, flows help.
3. Energy is the bottleneck = the trade. Gridsight's model (unlocked 600MW via software) says demand-side software prints real value; watch Australian grid-tech private names for IPO/A2A-like catalysts.
4. Tam's research lane (in flight): health relativity — data-centre power is a societal cost (bills, water, land); national cabinet rules (Wed) will shape economics. The political risk and the renewables-binding rules could slow but they also make the nuclear/uranium case stronger.
5. AI startups: the money is in infrastructure, not models — Australian ecosystem has a systematic bias toward infrastructure/defence/mining AI (Government/CSIRO-backed). Watch Firmus IPO closely as the bellwether for the whole local AI complex.

Sources

AEMO 2026 ESOO + data-centre forecasting overview (aemo.com.au, 25 Aug); ABC News / The Age coverage 25 Aug; NEXTDC FY26 results + contract updates (ASX PDFs, 27 Aug + 21 Jul + 20 Apr); Firmus raise wall (streamlinefeed, techstartups, Forbes AU, 7-10 Sep); Gridsight + Metacognition rounds (overnightsuccess.vc 5 Sep, TechStartups 10 Sep). Quotas/valuations are press-reported, not verified filings.

Status / next steps

• Tam to extend: household/consumer bill impact; water usage of centres; regional city clusters (Canberra/Geelong/Darwin edge nodes); comparison vs Singapore/Malaysia regional competitors
• Desk: NXT rate-of-speed (MA20/50, candle-quality on the breakout), uranium-book carrying the baseload thesis into Sept-Oct
• Paper watch: NX D data-centre demand × uranium baseload = our theme compounding — desk rule: keep PDN/DYL/BOE tranche discipline intact through the noise