The AI Stack

• Alphabet (Q3 results, 28 Oct 2026): 2026 capex of $195-205B, unchanged; CFO said capex will increase significantly in 2027.
• Amazon (Q3 results, 29 Oct 2026): about $220B of 2026 cash capex reaffirmed; CEO said it expects to spend meaningfully more in 2027.
• Meta (Q3 results, 21 Oct 2026): $135-145B including finance-lease principal; 2027 plans to be shared on the Q4 call.
• Definitions differ by company (calendar year, cash, finance leases), so the sum is indicative only. Source: company earnings releases as summarised by Luminix, Oct 2026.

Microsoft (fiscal Q1 2027 results, 28 Oct 2026) guided calendar 2026 capex to about $170B, down from about $175B, citing delays in third-party data-centre deliveries. It expects capex above $50B next quarter, and its CFO said demand still exceeds supply. Source: Microsoft earnings materials as summarised by Luminix, Oct 2026.

The IEA's April 2026 update reported data-centre electricity use rose 17% in 2025, more than five times faster than the roughly 3% growth in global demand, and that AI-related demand is expected to triple by 2030 while total data-centre demand doubles. Source: IEA via W.Media, 20 Apr 2026.

H100 prices were near $8/hr in early 2024. One-year contracts were about $2.35/hr by March 2026 (SemiAnalysis), and the trade-based Ornn index hit $2.94 on 25 May 2026 before easing. Nebius raised its on-demand H100 rate from $3.85 to $4.50 per GPU-hour from 1 Oct 2026, so the decline has paused or partly reversed at some providers. Sources: KuCoin News on Ornn, May 2026; IntuitionLabs, Oct 2026.

Not refreshed: the inference-cost series (Stanford HAI 2025), the capex share breakdown, GPU depreciation lives, nuclear and SMR timeline status, sovereign AI deal counts, and the 2024 baseline of 415 TWh. The 2030 figure of 945 TWh was not re-verified at source in this update.

Every AI answer is produced by a physical machine, acres of silicon, steel and power lines that someone had to build and pay for. The cloud is not a cloud. It is the largest industrial build-out since the railroads.

1. Prompt: your words travel to a data centre, often in another state.
2. Route: a scheduler assigns your request to a rack of GPUs.
3. Compute: the model's weights run in a few hundred milliseconds.
4. Return: the answer streams back, costing the operator a fraction of a cent.

The IEA expects this to more than double by 2030, and its April 2026 update shows data-centre use grew 17% in 2025 alone. The internet's physical footprint is becoming visible in national energy statistics.

An AI data centre is a factory: its raw material is electricity, its machines are GPUs, and its product is computation. Thinking of it as a factory, not an office, explains every design decision.

• Density: AI racks draw 5-10x the power of normal racks, so cooling becomes the design problem.
• Liquid cooling: air can no longer remove the heat; water and direct-to-chip loops take over.
• Networking: 100k GPUs must talk constantly; the network is as critical as the chips.
• Location: sites are chosen for power and water first, latency second.
• Speed: the race rewards whoever energises capacity first.
• Depreciation: the $60B of GPUs inside will be obsolete within five years.

1. Classic racks: fans and cold aisles (air-cooled).
2. Rear-door: chilled doors on racks.
3. Direct-to-chip: liquid loops on the silicon.
4. Immersion: servers submerged in fluid (liquid-cooled).

• Nuclear revival: hyperscalers signing deals to restart and build reactors.
• Behind-the-meter: data centres built at the power plant, skipping the grid.
• Renewables + storage: solar and batteries where land and sun allow.
• Gas turbines: fast to deploy, climate targets quietly strained.
• Grid investment: transformers and lines becoming strategic goods.
• Efficiency: each GPU generation does more per watt, but demand grows faster.

The IEA projects US data centres will use more electricity than aluminium, steel, cement and chemicals production combined. Computation is becoming the defining industrial load of the American grid.

1. 2023First PPAs: hyperscalers sign nuclear power deals.
2. 2024Three Mile Island: Microsoft funds a reactor restart.
3. 2025SMR orders: Google and Amazon back small reactors.
4. 2028eFirst SMRs: new small reactors targeted online.
5. 2030sGigawatt fleets: dedicated nuclear for AI campuses.

You no longer buy AI compute, you rent it by the hour, and the rental price just collapsed. What happened to GPU prices in two years is the most under-reported story in AI.

• Hyperscalers: AWS, Azure, GCP. Full platforms, premium prices, deep pockets.
• Neoclouds: CoreWeave, Lambda, Nebius. GPU specialists, aggressive pricing.
• NVIDIA itself: sells the chips and increasingly rents them too.
• Model labs: OpenAI and Anthropic buying capacity years ahead.
• Spot markets: idle GPUs auctioned by the hour, the price floor.
• Sovereign clouds: nations buying compute as strategic reserve.

• HBM memory: SK Hynix and Micron, the scarce memory bolted to every AI chip.
• Networking: NVIDIA InfiniBand, Broadcom, Arista, the fabric between 100k GPUs.
• Transformers & switchgear: multi-year waiting lists; the grid's quiet chokepoint.
• Liquid cooling: Vertiv and peers, where plumbing becomes strategic.
• Construction & EPC: data-centre builders booked out for years.
• Land & fibre: sites with power, water and connectivity command premiums.

Source: Government announcements; industry reporting (2024-26).

1. Cheaper compute
2. More AI usage
3. More revenue
4. More capex
5. Better chips (which leads back to cheaper compute: the build-out flywheel)

A two-by-two map with scarcity of the input on the horizontal axis and stickiness of the customer on the vertical axis. Quadrants: top-left loyal, abundant; top-right the fortress; bottom-left commodity; bottom-right scarce, fickle. Apps sit in the loyal, abundant quadrant (high stickiness, low scarcity). Cloud sits near the centre line on stickiness and just left of centre on scarcity. Models sit near the centre. Silicon sits in the fortress (high scarcity, high stickiness). Power & grid sits at the right edge on scarcity, near the middle on stickiness. Bare GPU rental sits in the commodity corner (low scarcity, low stickiness).

• The railroad lesson. The track builders went bust; the economy that ran on the track thrived.
• The fibre lesson. Dot-com fibre was overbuilt, bankrupted its owners, then powered two decades of the internet.
• The likely rhyme. Some of today's AI capex will never earn a return, but the capacity will not be wasted.
• The investor's task. Distinguish owning the boom's spending from owning what the boom leaves behind.

The stack's iron rule: value settles where the constraint is, and the constraint is moving from chips to power to distribution.

2023's bottleneck was GPUs. Today's is grid connections. Tomorrow's is likely customer relationships.

• Return on capex: roughly $720-740B of 2026 hyperscaler capex guidance (Q3 2026 earnings; definitions differ) must eventually meet matching revenue.
• GPU depreciation: 3-5-year asset lives against 30-year financing instincts.
• Power denial: grid queues and local opposition stalling projects.
• Price collapse: compute deflating faster than contracts assume.
• Concentration: a handful of firms are the whole demand side.
• Policy shock: export controls, energy rules, data-centre moratoria.

The AI build-out is on track to absorb a trillion dollars a year, more than global oil-and-gas upstream investment. Whether that is visionary or reckless will define this market decade.

1. Does compute keep deflating? The pace of GPU price decline sets everyone's margins.
2. Do the gigawatts arrive? Energised capacity, not announced capacity.
3. Does a neocloud stumble? The first casualty will mark the cycle's turn.
4. Do nuclear deals materialise? Power partnerships moving from press release to plant.
5. Does capex guidance crack? One hyperscaler blinking changes the whole narrative.

The cloud turned out to be made of concrete, copper and gigawatts. Whoever controls the scarce layer of that machine, this year chips, next year power, quietly sets the price of intelligence.

1. AI is an industrial system: six layers of physical infrastructure, not an app.
2. Power is the new chip: electrons, not GPUs, now gate the build-out.
3. Compute is commoditising: the middle of the stack is being repriced.
4. Value sits at the edges: scarce inputs and sticky customers win.
5. Watch the constraint: it moves, and profit moves with it.

• IEA: 'Energy and AI' report; data-centre demand projections (2025).
• Stanford HAI: AI Index Report 2025 (inference cost trends).
• GPU cloud price trackers: H100/B200 rental rates (2023-26).
• Company disclosures: hyperscaler capex and data-centre plans.
• Epoch AI: compute and training-cost trend data.
• McKinsey: data-centre and power-demand analyses (2024-25).
• S&P Global / DCD: grid interconnection and energisation reporting.
• Figures marked 'estimate' or 'illustrative' are compilations.