AI & THE WRONG LAYER.

AEMO's official forecast: data-centre electricity demand in the National Electricity Market triples to about 12 TWh by 2030 — roughly 6% of the grid, or enough to power every home in Victoria — then climbs to about 34 TWh (12% of the grid) by 2050. Under an accelerated AI-uptake sensitivity, the post-2030 figure could run 40% higher.

The connection-request pile is far larger than the real demand. AEMO received 44 GW of data-centre connection requests in its 2025 planning round. Independent modelling (Oxford Economics) found roughly 6 in every 7 megawatts requested is "phantom demand" — projects that won't materialise. If all 44 GW ran flat out, they'd approximate current global data-centre consumption. So the headline pipeline numbers are wildly inflated — but even the realistic slice is a step-change load on a grid simultaneously trying to retire coal.

NSW pipeline alone: 44 data centres (as at 31 March 2026) totalling 11.4 GW — the output of nearly four Eraring power stations, the country's largest coal plant. The proposed Mamre Road facility in Sydney would, at 1.2 GW, be Australia's single largest energy user — larger than the Tomago aluminium smelter — with 852 diesel back-up generators and 14.4 million litres of diesel storage.

The catch: data-centre load is flat and constant, which favours baseload — coal and gas — at exactly the moment the grid is going renewable-and-intermittent. Bloomberg NEF estimates ~83% of incremental global data-centre demand will be met by fossil fuels to 2030. CommBank's read: stronger-than-forecast data-centre growth is the key upside risk to household electricity prices. You may end up paying more for power so a machine can do a job it's worse at than you are.

Cooling those chips takes water. Industry estimates put total Australian data-centre water demand more than tripling — from 5.5 GL to about 17 GL — within five years. Sydney Water reports receiving applications for single data centres wanting up to 40 million litres a day — the equivalent of 16 Olympic swimming pools, every day, for one building.

By 2030, data-centre water use is projected at 1.9% of Sydney's supply and 0.9% of Melbourne's — and connection requests suggest the real appetite could be far higher. This lands as climate change intensifies drought and water stress. The government's 15 July framing named water explicitly as a risk; the framework to price or ration it does not yet exist.

Australia already has at least 162 operational data centres and around 90 more in the pipeline (not all of which will be built). On investment: Microsoft A$25B to 2029 (a 140% footprint expansion on 29 sites — implying on the order of 70 new sites or capacity-equivalents), AWS A$20B, plus Google's flagged (then reportedly paused) A$20B. NSW's Investment Delivery Authority endorsed a $51.9B, 15-project tranche for priority support in March 2026.

Add the announced numbers and you're past $65 billion in five-year commitments. The honest caveat, from a hyperscaler developer quoted at the Clean Energy Summit: "every property developer ever born is now a data centre developer" — so the gap between applications and final builds is huge. The real count will be a fraction of the requests. But it is still the fastest, largest infrastructure build in the country, and — per the national accounts — the biggest thing propping up the growth figure right now.

Technical analyses converge on a useful lifespan for AI chips of one to three years — from thermal/electrical wear and, more decisively, obsolescence when a new generation ships. Nvidia's Blackwell generation offered 4–5× the inference of the H100; the moment a 2× performance-per-watt jump lands, the old fleet's cost-per-token effectively doubles while the silicon still works.

The hyperscalers depreciate this hardware over five to six years — longer than the facts arguably warrant. Microsoft's own Satya Nadella: "I didn't want to go get stuck with four or five years of depreciation on one generation." Short-seller Michael Burry has publicly questioned whether the industry is overstating earnings by stretching those lives. The Economist called it a "$4 trillion accounting puzzle." There's a legitimate counter — silicon can cascade from training to inference to batch work and earn its keep longer (theCUBE/SiliconANGLE make this case) — but even the optimistic framing is a 6-year clock on a building meant to serve for decades.

The second clock is the one that could strand the whole model. Through 2026, capable AI stopped being something you rent from a distant data centre and started being something that runs on hardware you already own.

Small language models (roughly 1–14B parameters) now run locally on phones and laptops via dedicated neural processing units. Apple's M-series Neural Engine, Qualcomm's Snapdragon X2, Intel and AMD's 40–50 TOPS NPUs — every Copilot+ PC clears the bar. Offline translation in 150+ languages, on-device photo editing, local coding assistants, private health analysis: all running with no data centre in the loop, no per-query cost, no network. As Meta's on-device lead put it, mobile NPUs now approach the capability of a 2017 data-centre GPU. The industry consensus is a "local-first" split — the device handles the common, latency-sensitive majority; the cloud is a capability ceiling reached only for the hardest queries.

The strategic risk is obvious. If the bulk of everyday inference migrates to the appliance in your pocket — for privacy, cost, latency and offline resilience — then a large slice of the centralised capacity being built today is sized for a workload that partly walks out the door. AEMO itself lists "efficiency gains" and "project deferrals" as the downside scenario that flattens demand beyond 2030. You don't need on-device to win everything. You only need it to win enough to leave a continent's worth of half-empty, coal-fed, water-cooled sheds depreciating on somebody's balance sheet.

Gallup's most-replicated finding: managers account for roughly 70% of the variance in team engagement. The 28%-of-time-spent-managing figure cuts both ways — the human 28% may be the part that actually holds a team together, not the part to delete. Newsweek's reporting on the 2026 cuts quotes psychologists noting managers do invisible work: interpreting organisational change, absorbing emotional shock. Cut them and, per Forbes, you may be "dismantling the apparatus that produces senior leaders in 2028." Bayer flattened hard — and one analysis calls it a failed experiment with flat growth; Fortune documents the "megamanager" with 90 reports as a burnout machine, not a triumph. Automating the coordination layer could hollow out the exact function that makes the frontline liveable.

Princeton's Arvind Narayanan argues Moravec's paradox describes "what the AI community finds it worthwhile to work on" more than a hard limit on what's possible. Computer vision was the canonical "impossible" case — until deep learning cracked it around 2012–13, almost overnight. Humanoid robotics and physical AI are advancing fast; the frontline's safety may be more temporary than this dossier implies. Betting that the machine will stay bad at the physical world is, in his words, "false comfort."

Every past efficiency gain in computing produced more demand, not less. On-device AI could expand the total market rather than cannibalise the cloud — more capable local models create appetite for even more capable cloud models behind them. AEMO's demand curve keeps rising to 2050 for exactly this reason. And the "GPUs are stranded" case has a serious rebuttal (theCUBE, SiliconANGLE): silicon cascades from training to inference to batch analytics, earning revenue across a genuine 6-year life, just as CPUs did before. The build-out may be prudent capacity, not a bubble.

Sydney University's Prof. Uri Gal makes the case that "AI" is a convenient label for post-pandemic over-hiring corrections, investor signalling, and capex funding — decisions that were coming regardless. Australia's own DEWR report (8 July 2026) found no evidence of broad AI-driven labour-market upheaval: unemployment 4.4%, youth outcomes holding, no accelerated occupational reshuffling. If AI isn't actually driving the cuts, then "AI is aimed at the wrong layer" is answering a question that reality hasn't yet asked. We take this seriously. Our reply: the layer-choice argument is about where the technology and the deployments point, which is observable now (Cowork, Amazon's memos, the reversals) — regardless of whether the current headcount numbers are AI-caused. But we record the counter-evidence in full, because it's real.

The counter-case is strong enough that we hold the thesis as a direction, not a certainty: the evidence that AI is better at coordination than at the frontline is solid; the evidence that it's currently pointed the wrong way is solid; the claim that this will play out to mass frontline displacement and stranded sheds is a projection that the counter-arguments could genuinely overturn. We think the layer-choice is the right thing to watch. We do not think it's decided. That's the honest position, and it's the one we'll hold.