AI Weekly: Breakthroughs, Funding Rounds, and Policy Shifts (May 19‑25 2026)

OpenAI proves the Erdős unit‑distance conjecture – on May 20 the company announced that its internal general‑reasoning model independently solved the 1946 discrete‑geometry problem, producing a 125‑page proof using algebraic‑number‑theory and the Golod‑Shafarevich criterion. The proof was verified by Fields Medalist Tim Gowers and Princeton mathematician Will Sawin, who called it a milestone for AI mathematics.

Anthropic’s Q2 results – the firm reported roughly $109 billion in revenue, a $5.59 billion operating profit, a $440 billion annualized revenue run‑rate and an ~80× quarter‑over‑quarter growth. Growth drivers include Claude Code and an enterprise‑grade financial AI agent built with JPMorgan, while high‑spending customers (>$100 k annual spend) grew seven‑fold and $1 M‑scale clients rose from 12 to over 500.

Anthropic‑xAI compute deal – Anthropic signed a contract to purchase 300 MW of compute from xAI’s Colossus 1 data centre in Memphis, at a monthly cost of $12.5 billion, with a term through May 2029 and a discount period during the first two months.

Anthropic financing – the company is close to closing a financing round exceeding $300 billion, valuing it at over $9 trillion, led by Sequoia Capital, Dragoneer, Altimeter Capital and Greenoaks, with existing investors such as Founders Fund and General Catalyst contributing about $2 billion. Completion would make Anthropic the first private AI firm to surpass OpenAI’s valuation.

OpenAI IPO filing – on May 22 OpenAI submitted a confidential S‑1 to the SEC, targeting a September 2026 IPO with a valuation goal of $8.52 trillion‑$10 trillion and a typical 4‑6 month window from confidential filing to public offering. The filing aligns with a broader $3.7 trillion AI‑IPO wave that includes SpaceX and Anthropic.

Policy developments – President Trump cancelled a planned AI‑safety executive order after direct calls from Elon Musk, Mark Zuckerberg and David Sacks. The cancellation contrasts with the EU’s AI Act and signals a retreat in U.S. AI safety regulation.

Meta surveillance revelation – leaked audio showed Meta tracking employee Gmail, coding sessions and internal tools to train AI, coinciding with a layoff of 8 000 staff.

Papal AI encyclical – Pope Leo XIV released the first AI‑focused papal encyclical, “Magnifica Humanitas,” addressing AI’s impact on human dignity, creative work, labor rights, concentration of power, deception and autonomous weapons, and will be read by 1.4 billion Catholics.

Karpathy joins Anthropic – Andrej Karpathy, co‑founder of OpenAI and former head of Tesla Autopilot, moved to Anthropic to rebuild its pre‑training research team, signaling a major talent shift in the LLM field.

Google Gemini integrations – at I/O 2026 Google announced that Gemini will embed Canva’s “Magic Layers,” Adobe’s Firefly AI Assistant (for Photoshop, Premiere, Illustrator), and CapCut’s mobile video editing tools, expanding Gemini into a full‑featured creative studio.

Google Search redesign – Google is rebuilding the search bar to generate custom visuals, interactive graphics and mini‑apps directly on the results page, shifting focus from click‑through rates to AI‑generated overviews.

Intuit restructuring – Intuit plans a 17 % workforce reduction (~3 000 jobs) and expects over $5 billion in annual cost savings in H2 2026, redirecting resources toward AI agents, operational efficiency and mid‑market services.

Anthropic acquires Stainless – the acquisition adds a startup that builds high‑quality SDKs for APIs (Python, TypeScript, Java, Go, Ruby), promising to close the gap between Claude and OpenAI SDK quality within 6‑12 months.

OpenAI macOS certificate revocation – on June 12 the company will revoke macOS app signing certificates, causing the existing ChatGPT app to be flagged as untrusted; users must reinstall from a re‑signed build after a supply‑chain attack on TeamPCP.

TencentDB Agent Memory open‑source – Tencent released a MIT‑licensed, fully local AI‑agent memory system with a four‑layer architecture that provides persistent, structured memory without cloud dependence.

Perplexity’s Bumblebee – Perplexity open‑sourced a read‑only, agent‑less supply‑chain scanner covering npm, PyPI, Go modules, MCP configs and browser extensions, addressing a gap in AI‑development environment security.

ByteDance AI budget increase – the company raised its 2026 AI infrastructure capex by ~25 % to ¥200 billion, allocating ¥85 billion to AI chips and over $5 billion to domestic chip purchases (Cambricon, Huawei Ascend, etc.).

Baidu’s Wenxin 5.1 model – Baidu unveiled the “Wenxin 5.1” large model using a “multidimensional elastic pre‑training” technique that reduces pre‑training cost to about 6 % of industry benchmarks while achieving top rankings on the LMArena search leaderboard.

White House AI model review draft – the National Economic Council’s Kevin Hassett disclosed that the administration is drafting an executive order to review new AI models using a process modeled on FDA drug approvals.

Five‑Eyes AI safety guide – the intelligence alliance released joint guidance warning of unique risks from agentic AI systems (productivity loss, service disruption, privacy breaches, cyber incidents) and recommends limiting use to low‑risk tasks, enforcing identity management, continuous monitoring, human‑in‑the‑loop for high‑cost operations and regular red‑team testing.