Artificial Intelligence

• Fiscal 2026 (year to January 2026): data-centre revenue $193.7B, up 68%; total revenue $215.9B (NVIDIA Q4 FY26 release, 25 Feb 2026).
• Q2 fiscal 2027 (quarter ended 26 July 2026): total revenue $96.2B, data-centre $89.0B, up 117% year on year; Q3 revenue guided to about $108B, assuming no China data-centre compute revenue (NVIDIA, 26 Aug 2026, as reported by trade press).
• China demand is disputed: Culper Research, a short seller that posted on 13 May 2026 that it is short NVIDIA, says over 20% of FY2026 compute revenue stayed driven by China through Southeast Asian intermediaries (Culper Research, 13 May 2026, report not opened; Benzinga, May 2026). The outlook above excludes China data-centre compute revenue; NVIDIA has said diversion claims about one customer were investigated and debunked, and I found no reply to Culper by name (The Hill, 29 Oct 2026). The NVIDIA page (/investing/nvidia) sets out both sides.

• Five-company 2026 capex forecast: $803-828B; Q3 spend about $227B, up roughly 20% from $188.5B in Q2 (Business Insider via Hargreaves Lansdown, 30 Oct 2026).
• Amazon held about $220B of cash capex for 2026 and signalled meaningfully more in 2027; Microsoft trimmed its calendar-2026 outlook to about $170B on third-party data-centre delays; Alphabet's CFO said 2027 capex should increase significantly.
• This tally includes SpaceX and is wider than the four-company series in the capex chart.

• Anthropic: annualised run-rate revenue reported above $65B by end of July 2026, from about $47B in May (Bloomberg, as cited by Analytics India Magazine, 21 Aug 2026); Q2 revenue $11.6B (WSJ, as cited).
• OpenAI: Q2 2026 revenue $6.7B, up 18% on the quarter; FT reported annual revenue of about $40B (as cited, 21 Aug 2026).
• Run rates are company or press figures, not audited accounts, and the deck's frontier-AI revenue estimates in chapter 06 pre-date them.

• Anthropic raised $65B at a $965B valuation in May 2026 and confidentially filed for a US IPO in June (Bloomberg, FT, as cited, Aug 2026).
• OpenAI's latest private valuation is $852B after a $122B round in March 2026; it also filed confidentially in June and has indicated a 2027 listing (CNBC, as cited).

The ninth AI Index (13 April 2026) was reported, via secondary coverage, to again show GPT-3.5-level inference falling from $20 to about $0.07 per million tokens (over 280x) between late 2022 and late 2024, and global AI compute capacity growing about 3.3x a year since 2022. The report itself was not read directly.

Not refreshed: training-cost bar chart (no 2025-26 frontier cost figure verified), Epoch compute curve, McKinsey and Goldman Sachs estimates, the 2025 market-size figure, occupation exposure data, the hyperscaler four-company capex series by year, and the frontier-AI revenue estimates in chapter 06.

AI is software that learns patterns from data, instead of following rules a human wrote by hand.

That one shift, from programming to learning, is why AI can now do things no one could explicitly code.

1. Collect data: trillions of words and images are gathered and cleaned.
2. Train: the model tunes billions of internal 'dials' to predict that data.
3. The model: a frozen file of learned parameters, the finished 'brain'.
4. Inference: the model answers your prompt, the part you actually touch.

1. 1950Turing: the test is posed
2. 2012AlexNet: deep learning works
3. 2017Transformer: the modern architecture
4. 2022ChatGPT: AI goes mainstream
5. 2025Agents: AI begins to act

Almost everything you use today descends from one 2017 idea: the transformer. It let models learn from enormous datasets in parallel, the unlock that made scale pay off.

The chart's axis runs from 10^0 in 2012 to 10^10 by 2025, labelled approximately 10 billion times since 2012.

Labelled values: $670 (Transformer, 2017), $4.6M (GPT-3, 2020), $78M (GPT-4, 2023), $170M (Llama 3.1, 2024), $192M (Gemini Ultra, 2024).

The central paradox of AI: building the frontier has never been more expensive, while using it has never been cheaper. Hold both facts in mind: nearly every debate about AI as an investment is really an argument about which curve wins.

A matrix plots durability of moat (vertical) against value captured today (horizontal). Chips and Cloud sit top-right, in the 'Durable winners' quadrant (Chips furthest right and highest). Apps and Tooling sit left of centre, near the middle of the durability axis, with Apps slightly higher. Models sit lower right, near the 'Emerging' quadrant. The quadrants are labelled Fragile winners (top-left), Durable winners (top-right), Contested (bottom-left) and Emerging (bottom-right). This is an analytical placement, not measured data.

The picks-and-shovels layer is capturing the profits. Whether the application layer ever does is the trillion-dollar question. In past platform shifts the apps eventually won the biggest prizes, but only after the infrastructure was built and cheap.

Sources for these figures: Goldman Sachs (2023); McKinsey (2023); Grand View Research.

Goldman Sachs estimates 300 million full-time jobs are exposed to AI automation worldwide. 'Exposed' is not 'eliminated', but it does mean the task content of most knowledge work is about to change.

• Automation has always displaced tasks. Yet total employment has risen for 200 years.
• The transition is the hard part. New jobs arrive, but rarely for the same people, in the same place, at the same time.
• White-collar goes first this time. Which is economically and politically novel.
• Productivity is the swing factor. The optimistic GDP case only lands if AI genuinely makes workers faster.

• Railroads and fibre looked like bubbles too. The capital looked 'wasted', but the infrastructure powered the next century.
• The spenders are profitable. Hyperscalers fund capex from cash flow, not speculative debt, unlike 1999.
• Deflation drives demand. Every fall in inference cost unlocks a new tier of viable applications.
• But timing is everything. Even a real revolution can bankrupt those who arrive too early, or lever too hard.

1. Does inference keep deflating? Continued price collapse is the strongest bull signal.
2. Do the hyperscalers blink on capex? A guidance cut would reset the whole sector's narrative.
3. Does an application layer break out? Durable, profitable apps would move value up the stack.
4. Do open models stay a step behind? If they close the gap, proprietary pricing power fades.
5. Does productivity show up in the data? The macro case needs measurable output-per-worker gains.

Building the frontier has never been more expensive. Using it has never been cheaper. Everything about AI as an investment follows from that single sentence.

1. A general-purpose technology: priced and adopted like no other before it.
2. An inverted cost structure: fortunes to build, pennies to use.
3. Value pooling at the base: chips and cloud win first; the top of the stack is still open.
4. A real prize and a real bubble risk: both measured in trillions, at the same time.
5. Augmentation over replacement: for most work, for now, but the transition will be uneven.

• Stanford HAI: Artificial Intelligence Index Report 2025 (Chapter 1: R&D).
• Epoch AI: 'Trends in Machine Learning' and compute/cost trend insights, 2024.
• McKinsey Global Institute: 'The economic potential of generative AI', June 2023.
• Goldman Sachs Research: 'The Potentially Large Effects of AI on Economic Growth', 2023.
• Grand View Research: Artificial Intelligence Market Size & Outlook to 2030.
• NVIDIA Corporation: Quarterly SEC filings and data-centre segment reporting.
• Company guidance: Alphabet, Amazon, Meta, Microsoft capital-expenditure disclosures.
• Vaswani et al.: 'Attention Is All You Need', 2017 (the transformer paper).
• Figures marked 'estimate' or 'illustrative' are compilations, not primary data.