Anthropic’s $30 Trillion Moonshot

Before we get to the What I Read This Week below… join me, the besties, an incredible group of speakers, and thousands of founders, investors, and executives at this year’s All-In Summit!

You’ll get access to the most important ideas, conversations, and people you wouldn’t encounter anywhere else, including your future potential investors, customers, and partners. See you there!

Get tickets here ➡️ All-In Summit 2026 September 13 – 15, 2026 • Los Angeles, CA

1) Anthropic’s $30 Trillion Moonshot

The Wall Street Journal reported on August 25 that Anthropic is preparing to tell IPO investors that its total addressable market exceeds $30 trillion. American GDP is about $30 trillion, so the pitch is that the annual value of work AI can do equals the entire US economy. The company is weighing a raise of up to $100 billion at a valuation near $2 trillion; this would make it the largest raise in history, surpassing SpaceX’s IPO this year.

Anthropic’s second-quarter revenue passed $11.5 billion, posting its first profitable quarter, a 14-fold growth since the same time last year. Anthropic is separately projecting 2028 revenue of $190 billion to $200 billion.

Anthropic plans to unveil its IPO prospectus, the main legal document a company uses to sell shares to the public, after Labor Day, with a potential listing in late September ​or early October. Anthropic is considering allowing existing shareholders to sell ⁠shares as part of its IPO while also weighing lockup ​periods longer than the customary 180 days.

2) Salesforce’s Best Day Since 2020

Salesforce sells the software companies use to manage their customer relationships. Every call, deal, and complaint gets logged in one place for easy organization. On August 26, it reported second-quarter revenue of ~$11.3 billion, up about 11%, and net income of $3.53 billion, up from $1.89 billion a year earlier. The gain includes $2.6 billion from strategic investments, most of it in its Anthropic stake. Salesforce Ventures put roughly $50 million into Anthropic’s Series C in early 2023, and Bloomberg valued the position near $5 billion in June. The stock rose 22% the next day, its second-best session on record.

Marc Benioff and Dario Amodei announced Claudeforce the same afternoon. A plugin called Salesforce in Claude is now in pilot, with open beta in September and 37 prebuilt sales skills. It runs so a seller works inside Claude and never has to open a Salesforce screen. Claude also becomes the default model behind Agentforce and Slack AI.

Benioff has spent the year calling the SaaSpocalypse “nonsense,” arguing that frontier models depend on Customer Relationship Management (CRM) software instead of replacing it. His answer is to hand Claude the interface while keeping the data and the permissions underneath. Their core product continues with per-seat pricing while Agentforce bills by usage, and is now running at $1.5 billion annualized, up over 240% Y/Y.

3) One-Shot Learning Hits Robotics

Teaching a robot a new task usually starts with teleoperation, where a person controls the robot through the motions and records the demonstration. Those recordings then train a policy, the software that turns what the robot sees into motor commands. The data closely matches the robot, but collecting enough of it is slow.

In August, several robotics companies showed different ways to reduce that bottleneck.

Dyna Robotics released Dyna-2, trained on more than one million hours of human video captured from head-mounted cameras while people cooked, tidied, and folded laundry.

Skild AI released S1 and showed how performance changes as the amount of training video increases. At 1,000 hours, written instructions worked better than a video demonstration on tasks the robot had already practiced. But on unfamiliar tasks, the advantage of video grew sharply as training scaled. At 100,000 hours, the robot shown a demonstration succeeded 66% of the time, compared with 9% for the robot given only written instructions.

Generalist released GEN-1.5, which can watch a single demonstration of a new task and then attempt it without retraining. Across 10 tasks, Generalist reported a 59% average success rate, while noting that these prompted skills remained more brittle than versions retrained specifically for each task.

All-In: Nvidia’s Historic Quarter, SaaS Comeback, Bessent vs Druck, America’s Debt Crisis, Cancer Vaccine

Aug 28

California is the 5th largest economy in the world, but 43% of its residents say the budget is a "big problem." So let’s take a closer look at the budget...

Read full story