AI Daily Briefing: March 27, 2026
It's been a packed Friday in AI. From courtroom drama between Anthropic and the Pentagon, to SoftBank writing a $40 billion cheque for OpenAI, to a new benchmark that makes frontier models look embarrassingly human — here's everything that happened today.
Anthropic Wins Injunction Against Pentagon Blacklisting
The week's biggest story got its next chapter today. A federal judge in San Francisco issued a preliminary injunction blocking the US Department of Defense from labelling Anthropic as a "supply chain risk" — a designation that would have effectively banned the company's Claude models from US government use.
Judge Rita Lin ruled that the Pentagon's move appeared to be retaliation for Anthropic's publicly stated positions on AI safety, potentially crossing the line into viewpoint discrimination. The injunction is temporary while the full lawsuit proceeds.
Anthropic filed the suit after the DoD added it to a restricted list alongside Chinese and Russian tech firms — a classification Anthropic argued was politically motivated, given that its biggest competitor, OpenAI, wasn't similarly flagged. The case is being closely watched across the industry as the first major confrontation between an AI lab and the US military over AI ethics and national security framing.
SoftBank Secures $40 Billion Bridge Loan to Double Down on OpenAI
SoftBank confirmed today it has secured a $40 billion unsecured bridge loan — arranged by JPMorgan Chase, Goldman Sachs, Mizuho Bank, SMBC, and MUFG — to fund further investment in OpenAI and for general corporate purposes. The loan matures in March 2027.
This comes on top of the $30 billion SoftBank had already committed to OpenAI through Vision Fund 2. CEO Masayoshi Son has bet his legacy on AI, and at this scale, it's becoming one of the largest concentrated private investments in a single company in financial history.
For OpenAI, the cash couldn't come at a better time — the company is burning heavily on compute, its restructuring to a for-profit entity is ongoing, and competition from Anthropic, Google, and Meta is intensifying.
ChatGPT Ads Hit $100M Run Rate in Six Weeks
OpenAI reported that its ChatGPT advertising pilot in the United States has already reached a $100 million annualised revenue run rate — just six weeks after launch. The programme is currently US-only, but the speed of monetisation signals that OpenAI's pivot to ads is landing hard.
This is notable because OpenAI has historically been subscription and API driven. Ads represent a new and potentially massive revenue layer — one that, if it scales globally, could fundamentally change how the company finances its operations.
Trump's White House Science and Technology Council: The AI Dream Team
The White House confirmed this week that President Trump has appointed a who's who of the tech industry to a new Presidential Council of Advisors on Science and Technology (PCAST). The list includes:
- Jensen Huang — CEO, Nvidia
- Mark Zuckerberg — CEO, Meta
- Larry Ellison — Executive Chairman, Oracle
- Sergey Brin — Co-founder, Google
- Marc Andreessen — General Partner, a16z
- Lisa Su — CEO, AMD
- Michael Dell — CEO, Dell Technologies
The council is co-chaired by David Sacks, who has dropped his informal "AI Czar" label but reportedly retains significant policy influence. Notably absent from the list: Elon Musk and Sam Altman — an omission that has generated considerable commentary given both men's proximity to the administration in its early months.
ARC-AGI-3 Drops — and Frontier AI Scores a Humbling 12%
The ARC Prize Foundation — the organisation run by François Chollet and Zapier co-founder Mike Knoop — released ARC-AGI-3 this week, and it's already reshaping how we think about benchmarks.
Unlike previous ARC tasks which were static puzzles, ARC-AGI-3 is interactive and agentic: AI systems must learn rules through multi-turn interaction with novel environments, rather than pattern-match against training data. It's designed to be fundamentally resistant to memorisation.
The results from the preview phase are stark. The best AI system scored 12.58% against a human baseline of 100%. Google's Gemini Pro led the pack — at 0.37% on the most challenging sub-tasks. Even on the broader evaluation, no model came close to human-level reasoning.
Chollet has long argued that ARC is the real test of general intelligence, not MMLU or HumanEval. ARC-AGI-3 takes that argument further: if you can't learn from interaction in a novel environment, you're not intelligent — you're interpolating.
OpenAI Leases 200,000 sq ft of Hardware Space in Richmond
OpenAI has leased over 200,000 square feet at the Portside Commerce Center in Richmond, California. The scale suggests dedicated hardware and infrastructure — possibly for custom silicon development, large-scale training runs, or the physical backbone of its agentic product ambitions. Details remain sparse, but the facility reportedly joins a growing network of compute infrastructure OpenAI is building independently of its cloud partnerships.
EU Antitrust Chief Meets Google, Meta, OpenAI, Amazon CEOs
EU antitrust chief Teresa Ribera met with the CEOs of Google, Meta, OpenAI, and Amazon this week amid growing scrutiny of AI practices under the EU AI Act and Digital Markets Act. No formal actions were announced, but the meetings signal that European regulators are moving from observation to engagement — with formal enforcement inquiries expected in the coming months.
Research Highlight: ARC-AGI-3 Paper (arXiv:2603.24621)
The formal ARC-AGI-3 paper dropped on arXiv today, authored by the ARC Prize Foundation. The abstract frames it as "an interactive benchmark for studying agentic intelligence through novel, abstract, turn-based environments in which agents must reason through exploration rather than recall." It's already one of the most discussed AI papers of the week.
Worth a read if you're following the frontier intelligence debate seriously.
The Big Picture
Today's news tells a coherent story: AI is no longer a technology story — it's a geopolitical and financial story. A federal court is adjudicating what AI companies are allowed to say. A $40 billion loan is being written to fund a single AI startup. The President of the United States is appointing the CEOs of Nvidia, Meta, and Oracle to advise on science policy.
And at the same time, a new benchmark quietly showed that all of this investment and political power hasn't yet produced a system that can learn to play a simple new game the way a human child can.
That gap — between the money and the capability — is the real story of 2026.
Published by BuildrLab — building AI-first software. Follow us at buildrlab.com
Top comments (0)