Anthropic proposes a 5% fund; Microsoft exits an Nscale data center

Good morning.

Three corporate moves lead this morning, and the fine print decides what the money and the computing capacity are worth.

Anthropic’s proposed 5% Public Benefit Fund leaves a practical question open: when does a share pledge become money in somebody’s hands?

Microsoft’s exit from Nscale’s Essex project surfaced Sunday, four days after Nvidia CEO Jensen Huang called Nscale possibly Britain’s most valuable startup in a decade. Nscale is now preparing a US IPO.

Nvidia and Apple each pledged $25 million to Trump’s ballroom. Nvidia requested anonymity from the very top of the donor list, an ambitious hiding place.

Stay curious,

Marcus Schuler

Anthropic proposed committing 5% of its shares to a new Public Benefit Fund on Sunday, subject to stockholder approval.

Anthropic’s cofounders pledged 80% of their own equity to charity at founding. The proposed fund brings combined founder, employee and company commitments to roughly 20% of shares. The board could add 2% for each 50% valuation rise above an undisclosed baseline.

Anthropic is establishing the fund as a wholly owned LLC. In a 2018 analysis predating this proposal, law professor Dana Brakman Reiser described philanthropic LLCs as limiting the public’s ability to examine and influence their spending.

Why This Matters:

• Anthropic’s proposal needs stockholder approval; workers and nonprofits still lack eligibility criteria for seeking support.
• As Anthropic’s valuation rises, board discretion will determine whether the fund receives additional shares and how large it ultimately becomes.

Microsoft has withdrawn from Nscale’s Essex data center for 23,040 Nvidia GB300 processors.

Microsoft was the site’s only publicly announced customer, and Nscale says its other Microsoft contracts remain intact. Nscale is preparing a US IPO with $103.4 billion in total contract value, only $2.6 billion of it active, against $140.6 million in first-half revenue. It wants a valuation of about $30 billion, while a group of investors has suggested $16 billion to $19 billion.

InferenceBench tests Cohere’s open speech-to-text model across four datasets and finds an average word error rate of 13.65%, against the published 5.42%.

InferenceBench used H100 GPUs and 30 samples per dataset, finding a word error rate of 1.46% on clean audiobook speech and 29.17% on distant-mic meetings. Self-hosting cost about $0.21 per hour of audio on a $2.50-an-hour H100.

• India’s CERT-In disclosed on October 15 that AI agents from multiple providers reached live organizational systems during research and testing runs, reading personal and business records through exposed credentials.
• Meta removed a business-website training provision on October 17 after legal warnings about publisher scraping, while saying its web-crawler practices had not changed.
• DeepSeek released open-weight V4.1-Flash-1017 on October 17, with company-reported 77.0% on DeepSWE for under $0.20 per task.
• Goodfire’s Nick Cammarata, a co-author of a 2020 paper on tangled AI neurons, challenges the theory’s reading of the evidence; Google DeepMind’s Neel Nanda remains unconvinced.
• Replit made ChatGPT the default for Replit Agent; CEO Amjad Masad said it had been one of Anthropic’s top five customers.

Write a delegation brief that prevents rework.

A task can sound clear in conversation and still stall because the assignee does not know where authority ends. Use the model to turn scattered notes into a compact operating brief.

Your raw input:

The prompt:

Why this works:

Explicit decision rights reduce approval loops without granting vague authority. The "done means" test converts an abstract standard into evidence both people can check.

What to use:

Claude handles long email threads and mixed meeting notes well. ChatGPT is a solid fallback for a short assignment.

Nvidia and Apple each committed $25 million to President Trump’s White House ballroom, according to a leaked fundraising spreadsheet The Washington Post published in an exclusive on October 15.

Our take: Nvidia and Apple each committed $25 million to Trump’s ballroom, the highest amount on the leaked list. Nvidia requested anonymity. The top tier was an awkward place to avoid notice: the list recorded 67 donations above $10,000 totaling $333 million as of late September.

Donations pass through the Trust for the National Mall, generally under no obligation to disclose donors. Nvidia’s anonymity request came with that provision already in place. Trump said last year that the project was being paid for “a hundred percent by me and some friends of mine.” His name does not appear on the list. The friends appear to have handled the arithmetic.

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai