Epoch AI

Where the book uses it

• Chapter 1 — The Stakes

  What the race is for: the belief that a lead in AI can compound into a lasting advantage, how long such leads have actually lasted, and what a trillion dollars a year is being bet against.

Where the record uses it

• Dated test Chapter 2, the ladder of cost, due 31 December 2027

Sources that use it

• Epoch AI, “Largest AI data centers ranked by power capacity, 2026,” the page as updated 9 October 2026 and its dataset as downloaded 11 October 2026 (Epoch revised two of the thirteen campus estimates the book’s ledger takes from it, against its dataset of 10 September, which matched the ledger on all thirteen: Meta Prometheus at 31 August 2026 now 471 MW, was 562, its end state 1,022 MW in the third quarter of 2028, was 854, the row reading “Buildings 7-10 (tents) are operational, based on infrared drone video from RAAD (raad.com) captured on this date”; Google New Albany now 333 MW, was 453; the other eleven unchanged; on the revised figures the ledger’s evidenced total is 4.56 GW against 38.4 GW announced or estimated on the twenty-six campuses with a figure, 11.9 percent; held by Sourcing as two dated snapshots). In the Sources
• Daniel Davies, “SpaceX considered as a leasing company,” FT Alphaville, 28 August 2026 (the column’s premise, that Anthropic “has agreed to take on more or less the entire compute of the Colossus I data centre near Memphis”, which “cost between $7bn and $13bn to build and kit out with GPUs” by Epoch AI’s or Global Data Center Hub’s estimate, uses “around 350MW” and “should generate $15bn of rental payments in the first year”; on a three-year depreciation life the return on investment “is likely to be more than 100 per cent”, and the column’s own reckoner table has it above 100 percent on every life from two to five years only at build costs of $7 billion to $9 billion, 114 percent at $10 billion on three years and 80 percent at $13 billion; “lease revenue of this kind is something you should expect to have for a good time rather than a long time”; the $1.25 billion a month is read in SpaceX’s final prospectus of 12 June 2026, as the rent Anthropic agreed to pay through May 2029; the $7 billion is read in the two estimators’ own free pages, both held in the Library, Epoch AI’s “Trends in AI supercomputers,” by Konstantin F. Pilz, Robi Rahman, James Sanders and Lennart Heim, 23 April 2025, as Colossus’s estimated hardware cost in March 2025, and Global Data Center Hub’s “How Did Elon Musk Build a $7B AI Factory in Memphis Faster Than Microsoft or Meta Ever Could?”, 21 July 2025, neither of which carries the $13 billion; the column read whole, its table included, on a subscription on 11 October 2026, after captures that carried the headline and the standfirst alone). In the Sources
• Sri Muppidi, “Anthropic Lowers Gross Margin Projection as Revenue Skyrockets,” The Information, 21 January 2026 (paywalled below its first two paragraphs, which are read in the Wayback Machine’s capture of 22 January 2026: “Anthropic last month projected it would generate a 40% gross profit margin from selling AI to businesses and application developers in 2025, according to two people with knowledge of its financials. That margin was 10 percentage points lower than its earlier optimistic expectations”; “The lower-than-expected gross profit margin resulted from the costs of running Anthropic models for paying customers, in a process known as inference, on servers from Google and Amazon. Those inference costs were 23% higher than the company had anticipated, the projections showed. Anthropic calculates gross margins by subtracting inference costs and other costs of selling its products”; Epoch AI’s footnote linking this report dates the projection: “Anthropic projected in mid-December a gross margin around 40%”). In the Sources
• Amelia Michael and Ben Cottier, “Total cost of ownership of a one-gigawatt AI data center,” Epoch AI data insight, 14 May 2026 (Epoch’s own model on its own assumptions, $38 billion up front and $8.5 billion a year for one gigawatt; “A typical one-gigawatt AI data center requires $38 billion in up-front capital expenditure (CapEx) and $0.9 billion in annual operating expenses (OpEx)”; annualized, $8.5 billion a year, on IT equipment assumed to last five years and a facility fourteen; servers “60% of the total”; “Shortening the IT lifespan to 3 years raises the total annual cost to $12B; extending it to 7 years lowers the cost to $7B”). In the Sources
• Ben Cottier, Ben Snodin, David Owen and Tom Adamczewski, “LLM inference prices have fallen rapidly but unequally across tasks,” Epoch AI Data Insight, 12 March 2025 (from the page’s method block, read 19 September 2026: “Across all of these benchmarks and performance thresholds, we found prices declining between 9x per year and 900x per year, with a median of 50x per year”, on six benchmarks with a rate fitted at each score threshold that a new model had set; “The fastest price drops in that range have occurred in the past year, so it’s less clear that those will persist”; “We set a threshold of at least 4 data points to include a regression result”; “All prices were a 3:1 weighted average of input and output token prices”; “We excluded reasoning models from our analysis of per-token prices. Reasoning models tend to generate a much larger number of tokens than other models, making these models cost more in total to evaluate on a benchmark”; with all model data before January 2024 removed, “the median rate increased from 50x per year to 200x per year”; the release dates in the page’s table of priced models run from 20 November 2021 to 5 February 2025, and the data files behind the page are stamped as updated on 20 November 2025; the method block re-read live on 21 September 2026, every passage above unchanged, archived as VERIFY 260921/R68 B). In the Sources

And 39 more: search the Sources.

Related terms

• Algorithmic efficiency
• Evidenced
• Application layer
• Arsenal behind every arsenal
• Capital expenditure
• Clocks of diffusion
• Cluster
• Exclusive generation
• Exclusivity of the model
• Falsifier
• Frontier
• General-purpose technology