Rhys Lindmark: AI 2026, Part 1: Chips

A chart-heavy piece arguing that Nvidia sits at the centre of the AI data centre buildout, but that the chip market is far from settled. Roughly 100 GW of capacity and about $2T of chips are still to be built, which leaves room for challengers.

Links

• Essay: https://www.rhyslindmark.com/posts/ai2026-pt1/ (undated; refers to Aug 2026)

Key points

• Scale: the AI buildout is about 2% of GDP, comparable to railroads, cars and telecoms. AI chip capacity grew about 10x, from 2M to over 20M H100-equivalents.
• Nvidia's position: revenue up from about $100B to $300B a year, with GenAI over half of it. Operating margin about 66% against about 33% for Big Tech. It filled about 70% of the roughly 12 GW built in 2020–25.
• What is next: about 100 GW of data centre capacity at roughly $20B of chips per GW, a $2T opportunity. Timing estimates run from 2030 (Epoch) to 2040 (the author).
• Power: US electricity capacity is about 1 TW (goal 2 TW); China is about 4 TW (goal 8 TW).
• Four groups of challengers: hyperscaler chips (Trainium, TPUs; Anthropic has committed to 5 GW through Amazon), inference startups (Cerebras, Etched, SambaNova), model labs (OpenAI's chip "Jalapeño") and Chinese makers (Huawei about 3 years behind; GLM 5.3 served entirely on Chinese hardware).
• Open questions: supply limits (power, EUV, CoWoS, HBM), the bubble question, GPU depreciation (about 6 years) against infrastructure payback, and where defensible moats come from.

Charts

About a dozen charts, mostly sourced from Epoch AI, EIA and CompaniesMarketCap. Among them: buildout against past technology investment, data centre against office construction, Nvidia revenue, margins and profit against Big Tech and earlier tech eras, Nvidia's market cap, and the US–China electricity and chip gaps.

Caveat

Opinionated and at times hyperbolic ("Today is Day 1 of the Singularity"). Many figures come from third-party sources, some linked without the underlying data, so check the numbers before reusing them.

Related

• claude-haiku-5-5 (the efficiency side of the same compute story)