AI Briefing

Today's biggest plot twist: Elon Musk, who spent months calling Anthropic 'misanthropic,' just rented his entire Memphis Colossus 1 data center to them, complete with a side-quest about building gigawatts of AI compute in orbit. Meanwhile OpenAI quietly swapped GPT-5.5 Instant in as ChatGPT's new default for 900M weekly users, and Apple agreed to write checks to roughly 37 million iPhone owners over Siri AI features that never shipped. The thread tying it all together is infrastructure: who controls the GPUs, the protocols, and the trust to make this stuff actually work at scale.

01IndustryCNBC ↗

Anthropic Rents All of Musk's Colossus 1 Data Center, Doubles Claude Code Limits, and Hints at Compute in Space

In a stunning detente, Anthropic signed a deal to use the entire 300+ megawatts of compute (roughly 220,000 Nvidia GPUs) at SpaceX's Colossus 1 data center in Memphis, formerly Musk's xAI flagship facility. The agreement immediately doubles Claude Code's five-hour rate limits and removes peak-hour throttling for paid tiers, and the two companies even floated jointly developing 'multiple gigawatts' of orbital compute capacity. Musk, who in February said Anthropic 'hates Western Civilization,' now says he is 'impressed' after spending time with the team — a remarkable pivot given he merged xAI into SpaceX (now SpaceXAI) earlier this year.

02ProductTechCrunch ↗

OpenAI Quietly Replaces ChatGPT's Default Brain With GPT-5.5 Instant, Cutting Hallucinations 52% in High-Stakes Domains

GPT-5.5 Instant is now the default model for all 900 million weekly ChatGPT users, replacing GPT-5.3 Instant. OpenAI says it produces 52.5% fewer hallucinated claims on high-stakes prompts in medicine, law and finance, jumps from 65.4 to 81.2 on the AIME 2025 math benchmark, and uses about 30% fewer words to make the same point. The bigger shift is memory: GPT-5.5 Instant can now reach back into your past chats, uploaded files, and connected Gmail to personalize answers, with visible source links so users can audit and delete what it remembers.

03PolicyTechCrunch ↗

Apple Pays $250M to Settle Class Action Over Apple Intelligence Marketing — Roughly 37M iPhone Owners Eligible for Up to $95

Apple agreed to a $250 million class action settlement covering about 37 million U.S. devices sold between June 10, 2024 and March 29, 2025, after plaintiffs alleged the company oversold the AI features (especially the upgraded Siri) it pitched alongside the iPhone 16 launch. Eligible owners can claim a base $25 per device, scaling up to $95 depending on how many people file. A court approval hearing is set for June 17 before Judge Noel Wise in the Northern District of California, in what is shaping up as the first major U.S. consumer payout for AI marketing that didn't ship on time.

04Open SourceOpenAI ↗

OpenAI, Nvidia, AMD, Intel, Microsoft and Broadcom Open-Source MRC, a New Networking Protocol Built for Gigascale AI Training

OpenAI released Multipath Reliable Connection (MRC), a new GPU-to-GPU networking protocol co-developed with Nvidia, AMD, Intel, Microsoft and Broadcom, and donated it to the Open Compute Project so any cloud or chipmaker can adopt it. MRC targets a stubborn bottleneck in modern training runs: at gigascale, network packet loss and tail latency can leave expensive GPUs idle. Nvidia's Spectrum-X Ethernet fabric is shipping with MRC support today, marking a rare moment where six fierce rivals agreed to standardize on the same low-level transport before forking incompatible alternatives.

05FundingTechCrunch ↗

Sierra Hits $15.8B Valuation on $950M Raise as 40% of the Fortune 50 Now Run Customer Service on Its AI Agents

Bret Taylor's AI agent startup Sierra raised $950 million led by Tiger Global and Google's GV at a $15.8 billion post-money valuation, up from $10 billion last fall. The company says more than 40% of the Fortune 50 now use its agents for tasks like mortgage refinancing, insurance claims and returns, and that it crossed $150 million in ARR in just eight quarters. The round is a reality check for the 'thin GPT wrapper' narrative: the most valuable layer of the AI stack right now may be neither the foundation models nor the chips, but the boring operational glue that gets agents to work in real enterprise workflows.