Benedict Evans — "AI Eats the World" (Spring 2026)
Annual macro deck (79 slides, Keynote-exported PDF) from Benedict Evans, dated May 2026, hosted on his ben-evans.com Squarespace site. Structured in three acts: Capital, Deployment, Change. This is his recurring State-of-AI presentation — the spiritual successor to his "Mobile is eating the world" decks from the 2010s.
Why this is in the vault
Evans is one of the few independent macro analysts who consistently zooms out and asks the structural-economics questions that hyperscaler earnings calls obscure. The deck lands directly on top of two RDCO active theses (memory cycle / hyperscaler capex AND innermost-loop AI infrastructure thesis), and the "mobile networks 2010 = LLMs 2026" analogy is a clean Sanity Check candidate. He is already a tracked author in the vault.
Key claims:
- Capex — $700bn planned 2026 capex across Meta + AWS + Alphabet + Microsoft. For comparison: global Telecoms ~$300bn, Oil & Gas ~$1tr. Capex/sales ratios push 40-60% at some hyperscalers — these are no longer asset-light businesses.
- OpenAI's $1.4tr / 30GW deals — "Other People's Balance Sheets, circular revenue and a lot of plate-spinning." Evans is sharp-tongued here, suggests aspiration of 1GW/week construction at $20bn/GW ≈ $1tr annually.
- Nvidia can't keep up — and can't get TSMC to expand fast enough. Quarterly revenue chart shows the run-rate divergence vs Intel.
- Bottlenecks everywhere — GPUs/memory, multi-year power backlog (except China), community/political backlash. He calls this "Pig in a python."
- Mobile-networks analogy (the deck's core thesis) — Bits=tokens, marginal cost, opaque pricing in 2010 ≈ LLMs today. Equilibrium converged on flat-rate capped bundles aligning price/usage/capacity, with WiFi/edge offloading. Long-term pricing power for the labs is "unlikely."
- Models are commodities — Selected frontier LLMs by aggregate benchmark score show convergence. No network effects, no value capture up the stack (telco mirror: trillion-dollar industry but all the value built by other people).
- OpenAI net revenue ~$4bn/month by mid-2026, Anthropic gross ~$4bn/month. 900m+ weekly ChatGPT users but only 5% paying.
- Consumer use is "a mile wide and an inch deep" — top 20% of users sent <1,000 prompts in 2025. Daily-habit penetration much lower than weekly-touch.
- Workplace adoption — Tech ~75% any-use, but Healthcare/Retail/Manufacturing/Government still in early innings.
- "AI gives you infinite interns" — best one-liner of the deck. Reframes the question from "what's expensive that becomes cheap" to "what was impossible that now becomes cheap" (e.g., listen to a million calls/day).
- Provisional thesis — Chat is a terrible UX. Labs can't build (or generate) all the apps. Models are commodities with no network effects. Innovation will move up the stack.
Mapping against Ray Data Co
Memory cycle / hyperscaler capex thesis (active anchor at investing/anchors/hyperscaler-capex/) — Strongly corroborates. Evans' $700bn-for-2026 number and the capex/sales curve are exactly the macro frame the thesis rests on. His semiconductor billings chart and TSMC-bottleneck framing reinforce the memory/HBM upstream pull. No challenge — this is supporting evidence. Worth pulling the capex/sales chart values into the next quarterly capex pulse brief.
Innermost-loop AI infrastructure thesis (memory + power + DC infra) — Strong corroboration on all three layers. Power backlog ("multi-year, except China") + DC construction overtaking office construction in US Census data is direct evidence. The power layer thesis benefits most from the explicit "build can't keep up" framing.
LLM-as-commodity thesis — Evans' frontier-benchmark-convergence chart and the telco-analogy ("commodity infra rarely captures value up the stack") aligns with the [[2026-02-05-stratechery-interview-benedict-evans-ai-software]] interview already in the vault. The mobile-networks-2010 frame is a sharper articulation than anything else we have on file.
Harness-engineering thesis — Neutral. Evans doesn't engage with agent harnessing / Claude-Code-style productivity frames. His "infinite interns" line gestures at the same territory but he treats it as a deployment question, not a tooling question. Not a contradiction, but not reinforcement either — he's working a level up the stack.
Sanity Check newsletter hook (strong candidate) — The "mobile networks 2010 = LLMs 2026" analogy is the most sharable single idea in the deck and not yet broadly absorbed. A Sanity Check piece could re-frame this as: "The capex you're staring at is the 3G buildout. The value capture isn't where the capex is." This is exactly the kind of original re-frame the founder wants from SC (per feedback_no_derivative_sanity_check_pieces) — not a restatement of Evans, but using his analogy as scaffolding for "so where DOES the value go in an LLM-commodity world?" Queue for Notion content calendar.
Tracked-author flag — Evans is already tracked via Stratechery interview. No new author candidates from this deck (it's a single-author presentation, not a multi-source survey). His Spring + Fall annual decks should probably be added to a recurring watch in the vault — flag for follow-up.
Related
- [[2026-02-05-stratechery-interview-benedict-evans-ai-software]] — Stratechery interview, same author, complementary frame
- [[2026-02-06-stratechery-weekly-saasmageddon-super-bowl]] — Weekly digest that cites Evans interview
- Mainframes, ML and Digital Transformation — Benedict Evans (archive/readwise-thin)
investing/anchors/hyperscaler-capex/— Anchor that this corroborates