Amazon's AU$2.8B renewable bet isn't a sustainability story. It's an infrastructure finance story.

The deal

On 16 April, Amazon announced nine new PPA (Power Purchase Agreement) structures across New South Wales and Victoria. The portfolio adds 430MW of wind, solar, and battery storage — worth AU$2.8 billion — and brings Amazon's total Australian renewable capacity to 990MW, nearly 1GW. That makes it the largest corporate power buyer in Australia this financial year. (Sources: Amazon announcement; RenewEconomy; Data Center Dynamics)

In Edition 1, we asked whether geopolitical chokepoints were being underpriced in data infrastructure. This edition picks up a connected thread: not who controls the cables, but who controls the power.

The engineering read

Eight of the nine deals include BESS (Battery Energy Storage Systems) — a first for Amazon outside the US. That detail is easy to skim past. It shouldn't be.

AI workloads are pushing power density to new extremes. A traditional server rack draws around 3–8kW. Modern AI racks can exceed 30–80kW — a 5–15× increase in power density. AI jobs can technically be checkpointed, but doing so is costly and disruptive to training cycles. Operators optimise for continuous, stable power delivery.

Storage-backed PPAs don't replace on-site UPS (Uninterruptible Power Supply) or generators — but they turn intermittent renewable energy into more predictable, dispatchable supply, reducing reliance on an already capacity-constrained grid. (IEA Energy and AI, 2025)

At utility scale, BESS is almost always built on LFP (Lithium Iron Phosphate) batteries — LFP now makes up ~85% of the stationary storage market, up from 48% in 2021. Third-generation cells (e.g. CATL 587Ah) are rated at 12,000+ cycles under managed conditions; CATL's TENER system is the first mass-producible system rated for zero capacity degradation in the first five years. Round-trip efficiency: 85–92%. Equipment costs ~US$125/kWh outside China, down 90%+ since 2010. (IRENA; Energy Storage News; CATL TENER, 2024)

The capital allocation read

Amazon spent ~US$125B on capex in 2025 — mostly AWS — and has guided to US$200B in 2026, with a US$200B infrastructure backlog. AU$2.8B is a line item. The structure is what matters. (Amazon Q4 FY2025 earnings; Platformonomics, Feb 2026)

1. Storage-backed PPAs convert variable energy risk into contracted infrastructure.

Plain renewable contracts leave the buyer absorbing intermittency risk. Storage-backed structures shift that risk upstream to the developer. For a hyperscaler, this turns an unpredictable operating cost into a fixed, contracted obligation — the same capital logic as any long-term offtake agreement in project finance.

2. The timing premium is real — and late movers will pay it.

As AI demand materially increases Australian grid prices, Amazon has already locked in its rates. The best co-located storage projects are finite; early movers secure both price and counterparty quality.

3. The PUE (Power Usage Effectiveness) gap reflects more than cooling efficiency.

Hyperscalers operate at fleet-wide PUE of ~1.10 — meaning ~91% of energy goes to compute. Industry average sits at 1.55–1.58; most colocation at 1.2–1.5, legacy 1.6+. That gap reflects custom hardware design, workload optimisation, scale economics, and vertical integration. Colocation operators can improve cooling; matching full-stack hyperscaler optimisation is materially harder.

The Effective Capacity formula introduced in this edition:

In a capacity-constrained grid, power availability is the binding variable. Storage-backed PPAs raise the floor on that variable — directly increasing effective utilisation and revenue stability.

What this means for the broader market

When your largest potential tenant is building near-1GW renewable portfolios with battery backing, "we use green energy" stops being a differentiator. It starts becoming a baseline qualification criterion.

Australia tends to move first in APAC. What plays out here usually arrives in Singapore, Malaysia, and Japan within a few years. Those markets have deep datacenter pipelines but slower, more complex paths to grid decarbonisation. The gap between hyperscaler expectations and what local grids can deliver is a tension worth watching closely.