Libertarians at the gate amv

E Y E O N T H E M A R K E T • M I C H A E L C E M B A L E S T • J . P . M O R G A N • Se p t e m b e r 1 0 , 2 0 2 6 US Economy and Iran War Tracker Eye on the Market Recommendation Report Card

Libertarians at the Gate: the remarkable and disturbing last 72 hours in AI A brief note on the extraordinary events of the last 72 hours: Anthropic power agreements, OpenAI Millennium and ARC prizes on remarkable jumps in model capabilities, disclosures by OpenAI and Anthropic regarding sandbox jailbreaks and coordinated AI agent swarm attacks, and the relentless campaign by frontier labs to argue for regulatory moats to dampen competitive threats from US/Chinese open models. [1] Power, power and more power Anthropic began the year with 1-2 GW of compute1 but according to a Sept 6 report has now committed to pay for access to another 15 GW of power that it will access over the next few years. The estimated cost: ~$500 bn to rent cloud capacity, chips and leases on data centers. This compares to $180 bn of projected data center expense that Anthropic mentioned last year through 2029. Anthropic is catching up to OpenAI which increased its estimate of data center expense from $660 bn to $750 bn through 2030, which is shooting for 30 GW by 2030 (equivalent to 30 modern 1 GW nuclear reactors, and the US only has ~95 of them). Most of this power continues to be either grid-based or behind the meter dedicated gas fired capacity (see box). Notable recent deals secured by Anthropic for AI servers Date announced/ Counterparty reported Amount Capacity First capacity Google Cloud Oct 23, 2025 Tens of billions 1,000 MW 2026 Fluidstack Nov 12, 2025 $50bn Undisclosed 2026 Microsoft Azure/Nvidia Nov 18, 2025 $30bn 1,000 MW Undisclosed CoreWeave Apr 10, 2026 Multi-billion Undisclosed 2026 Amazon AWS Apr 20, 2026 $100bn 5,000 MW 2026 According to a SemiAnalysis report on Google Apr 24, 2026 $200bn 5,000 MW Undisclosed Sept 10, the pipeline for data center SpaceX May 6, 2026 $45bn 300 MW May 2026 behind the meter onsite generation Akamai May 7, 2026 $1.8bn Undisclosed Undisclosed has risen to 75 GW, almost entirely AMD Jul 22, 2026 Undisclosed 2,000 MW H1 2027 comprised of natural gas turbines, Volta Infra Aug 4, 2026 $10bn 121 MW Dec 2026 reciprocating engines, steam turbines Nscale Aug 26, 2026 $45bn 460 MW Q4 2027 and solid oxide fuel cells. In other Lambda Aug 31, 2026 $35bn Undisclosed Undisclosed words, all based on natural gas rather Source: The Information, "How Anthropic clinched $517 bn in compute deals in 11 months", Sept 6, 2026 than renewables + battery or nuclear.

[2] Evidence of increasing AI capabilities: GPT-6 Astra and the ARC Prize Foundation The Millennium Math Prize allegedly solved by OpenAI got a lot of press this week2, but capabilities of OpenAI Astra 6 are just as interesting. Francois Chollet, a French computer scientist who worked at Google, has been a notorious AI skeptic. The Atlantic in April 2025 cited him in an article titled “The Man Out To Prove How Dumb AI Still Is”. In 2019, Chollet created the Abstraction and Reasoning Corpus for Artificial General Intelligence (ARC-AGI), an exam designed to show the gulf between AI model memorized answers and the fluid intelligence that people have. The exam assesses the ability to quickly acquire skills and solve unfamiliar problems from first principles, rather than just memorizing enormous amounts of training data and regurgitating information. GPT- 3 scored a zero on ARC-AGI-1 (humans score 60%-70%), and OpenAI o1 scored just 3% on ARC-AGI-2.

1 Epoch AI estimated Anthropic's December 2025 compute at 1.2 million H100 equivalents. H100s require 700W of thermal design power; after factoring in all the other components in the server drawing power, that leads to roughly double the power per H100, or 1,400 W, which when multiplied by 1.2 million H100s = 1.7 GW 2 OpenAI announced that an unreleased model produced a proof for the unsolved Navier-Stokes Millennium Prize problem related to fluid dynamics and turbulence; the mathematical community is still reviewing and verifying the work. Just this morning, OpenAI stated to the NY Times that they have made substantial progress on…