FREE WITH EVERY PLAN · OCTOBER 2026 EDITION
The AI Infrastructure Primer 2026
How neoclouds, AI factories and AI datacentres actually work — from the GPU-hour and the token to grid-to-chip power, liquid cooling, networking, financing and the global map. Written for investors, lenders and operators.
- pages
- 70
- chapters
- 16
- figures
- 26
- tables
- 13
Data as of 14 October 2026. Every material figure is sourced and dated; estimates are labelled. Reissued every April and October — each new edition goes to every subscriber.
Get the Primer
Free with Compute Open: confirm your email and we send the PDF link, plus the daily 08:00 ET compute read. Unsubscribe any time.
Already subscribed? Sign in — the Primer is in your library.
AI Factories &
AI Data Centres
16 chapters · 26 figures · 13 tables
What's inside
A plain-language technical tour of the machinery beneath the contracts: the silicon, how clusters are installed, powered, cooled, kept productive and networked, how sites are chosen and financed, and where the build-out is going. Six structural observations anchor the document:
- The unit of compute is now a rack, not a server.
- Electricity efficiency is a stack of small losses.
- Installed ≠ productive.
- Networking is the speed limit.
- Economics reward the integrated few — and 2026 pricing moved their way.
- The capital stack has institutionalised — and is now stratifying.
26 figures, 13 tables, a glossary and full source notes. Every exhibit is drawn for this edition.
Sixteen chapters in five parts
PART I · THE INDUSTRY
- The neocloud sector
- The AI factory and token economics
- The demand side
- Market structure and the competitive landscape
- The value chain
PART II · THE TECHNOLOGY
- The silicon arc: Hopper to Blackwell to Rubin
- Deployment: how GPUs actually get installed
- Power: from the grid to the chip
- Cooling: getting the heat back out
- GPU productivity: installed is not productive
- Networking: from chip to fibre
PART III · ECONOMICS AND CAPITAL
- Site selection and energy economics
- Financing the buildout
PART IV · GEOGRAPHY AND RISK
- Geography: the global AI datacentre map
- Risks and regulation
PART V · SYNTHESIS
- The operator scorecard
Plus: Glossary · Sources and methodology.
A few of the numbers inside
From Table 1, “The sector on one page”. Each figure carries its source and date in the Primer.
| Metric | Value | Date / source | Ch. |
|---|---|---|---|
| Neocloud / GPU-as-a-service revenue | 2025: >$25bn (Q4-25 $9bn, +223% y/y); 2031E ~$400bn (~58% CAGR) | Synergy Research, 2 Apr 2026 | 1 |
| One-year H100 contract | $1.70 → $2.35 per GPU-hour | SemiAnalysis, Oct 2025 → Mar 2026 | 1 |
| Contract value per MW of IT load | ~$20–25m per year (multi-year); ~$40–50m (short-dated) | Nebius Q2 (12 Aug), IREN, CoreWeave, Sep–Oct 2026 | 1, 12 |
| 2026 capex guidance | Amazon ~$220bn; Alphabet $195–205bn; Microsoft ~$175bn (basis ambiguous); Meta $130–145bn | Q2 2026 calls, late Jul 2026 | 3 |
| Big four + Oracle capex | ~$800bn (2026) → ~$1.2tn (2027) | Goldman Sachs, 25 Sep 2026 | 3 |
| NVIDIA Q2 FY2027 | Revenue $96.2bn; data centre $89.0bn; Q3 guide $108bn ±2% | NVIDIA, Aug 2026 | 5, 6 |
| All-in capex per critical-IT MW | GB300-class ~$45–55m (IT ~62%); Rubin-era ~$60m (~$60bn per GW) | Author estimate (Jul 2026); J. Huang, 26 Aug 2026 | 5 |
| Anchored vs merchant spread | IG SOFR+213–300 vs BB SOFR+450–550 (~250–325bp), pre-hike | 2026 deal prints | 13 |
Source: AI Infrastructure Primer 2026, October edition, data as of 14 October 2026 (Table 1), each row with its own source and date as printed. Selected rows; figures from licensed index providers appear in the PDF only. Not investment advice.