The Weekly AI Digest
Week of 19 June – 2 July 2026
Top 5 AI Stories This Week
Anthropic's Fable 5 and Mythos 5 Restored as Export Controls Lift
On 30 June the US Department of Commerce lifted the export controls it imposed on 12 June, and Anthropic began restoring Fable 5 globally from 1 July after adding a new safety classifier it says blocks the reported jailbreak in over 99% of cases. Mythos 5 was restored to a set of approved US organisations from 26 June through the Glasswing program. Commerce Secretary Howard Lutnick said the department worked with Anthropic for two weeks before approving the return, and blocked requests now route to Anthropic's Opus 4.8 model.
Why it matters: Frontier model access can be switched off and back on by policy, not just capability. The episode shows deploying across multiple clouds does not insulate a business from an action against the underlying model provider.
Google Loses Two Senior AI Researchers to OpenAI and Anthropic
Noam Shazeer, co-lead of Gemini and a co-author of the 2017 paper that introduced the Transformer architecture, announced on 18 June he is leaving Google for OpenAI as Lead for Architecture Research. Within the same week, AlphaFold lead John Jumper left Google DeepMind for Anthropic. Google had paid a reported $2.7 billion in 2024 to bring Shazeer back. Alphabet shares fell around 5% to 6% on 22 June, with reports linking the decline in part to investor concerns over AI talent retention and spending.
Why it matters: The scarcest resource in frontier AI is the people who can build the systems, and the largest labs are competing hard to hold them. Where senior architects land signals where the next model advances may come from.
European Parliament and Council Approve Delay to EU AI Act Deadlines
On 16 June the European Parliament approved the Digital Omnibus by 423 votes to 57, with 174 abstentions, pushing the compliance deadline for standalone high-risk AI systems from 2 August 2026 to 2 December 2027. The Council gave final approval on 29 June. The package also adds an explicit ban on AI tools that generate non-consensual intimate imagery and child sexual abuse material, effective 2 December 2026. Article 50 transparency obligations still apply from 2 August 2026.
Why it matters: The EU has bought high-risk AI deployers more time, but the framework and its penalties, up to 35 million euros or 7% of global turnover, remain. Which deadline applies now depends on how each system is classified, not a single headline date.
OpenAI and Broadcom Unveil a Custom Inference Chip
OpenAI and Broadcom unveiled Jalapeño on 24 June, OpenAI's first custom-designed AI inference processor, built specifically for large language model inference and developed with Broadcom and manufacturing partner Celestica. The companies said engineering samples are running in the lab, with initial deployment planned by the end of 2026 through data centre partners including Microsoft, and further generations to follow. Detailed performance figures have not yet been released.
Why it matters: The largest AI companies are moving to design their own silicon rather than rely solely on merchant chips. Control over inference hardware affects both cost and capacity for everyone building on their platforms.
MGX Closes One of the Largest AI Funds Ever at $49 Billion
MGX, an Abu Dhabi investment firm and backer of both OpenAI and Anthropic, closed its Fund I on 1 July at $49 billion, above its $45 billion target, making it one of the largest AI-focused vehicles ever raised. The firm has backed 14 companies and co-led or joined recent mega-rounds at Anthropic, OpenAI and xAI.
Why it matters: Capital pooled specifically for AI continues to scale, keeping the funding environment for model developers and infrastructure well supplied. It signals sustained investor appetite even as the leading labs move toward public markets.