AI News Briefing — May 7, 2026

7 Top Stories

1. Helen Toner’s Deposition Reveals Deep OpenAI Governance Failures

Former OpenAI board member Helen Toner delivered explosive testimony in the Musk v. Altman trial, revealing that she first learned about ChatGPT’s existence through Twitter screenshots. Toner described a board kept systematically uninformed and said Altman’s lack of candor — a “pattern of behavior” rather than any single incident — was the root cause of his 2023 ouster. She characterized AI model safety testing as “more like alchemy than chemistry,” criticizing the field’s lack of rigorous, reproducible methods for evaluating model risk.

2. Shivon Zilis Testifies: Altman Concerns, Microsoft Deal, and Tesla AI Ambitions

Shivon Zilis, an executive at both Tesla and xAI, took the stand in the Musk v. Altman trial and revealed she had raised serious concerns about Altman with the OpenAI board. She said the board was not notified before ChatGPT’s massive public release, and that a proposed deal with Helion (a nuclear energy startup in which both Altman and co-founder Greg Brockman had investments) “felt super out of left field.” Zilis also described Musk’s plan to build a world-class AI lab at Tesla, even suggesting recruiting DeepMind’s Demis Hassabis to bring more “humanity” focus to Tesla’s AI work.

3. Anthropic Launches “Dreams” — Claude Agents That Self-Improve Overnight

Anthropic unveiled a research-preview feature called Dreams that allows Claude managed agents to review their past sessions, identify patterns, and reorganize their memory stores — effectively “sleeping on it.” Powered by Claude Opus 4.7, Dreams takes a memory store and up to 100 session transcripts as input, then produces a cleaned-up output store with deduplicated entries, resolved contradictions, and surfaced insights. The input store is never modified, so developers can review and approve the results. This represents a significant step toward AI agents that autonomously improve over time.

4. Anthropic Strikes 300MW Compute Deal with SpaceX, Doubles Claude Code Limits

Anthropic announced a major compute partnership with SpaceX that gives it access to the full capacity of SpaceX’s Colossus 1 data center in Memphis — over 300 megawatts and 220,000+ NVIDIA GPUs, coming online within the month. As an immediate result, Anthropic is doubling five-hour rate limits for Claude Code across Pro, Max, Team, and Enterprise plans, removing peak-hours restrictions for Pro and Max, and significantly raising API rate limits for Claude Opus models. The deal is part of Anthropic’s broader compute expansion that includes agreements with Amazon (5 GW), Google (5 GW), and Microsoft ($30 billion in Azure capacity). Anthropic also expressed interest in partnering with SpaceX on orbital AI compute infrastructure.

5. OpenAI Releases MRC — Open Standard for AI Supercomputer Networking

OpenAI published the Multipath Reliable Connection (MRC) protocol specification through the Open Compute Project, a networking standard co-developed with AMD, Broadcom, Intel, Microsoft, and NVIDIA over two years. MRC extends RDMA over Converged Ethernet (RoCE) and enables multi-plane high-speed networks that can connect 130,000+ GPUs using only two switch tiers — versus three or four in conventional designs. It uses adaptive packet spraying to eliminate core congestion and SRv6-based source routing to bypass failures in microseconds. MRC is already deployed across OpenAI’s largest NVIDIA GB200 supercomputers used for training frontier models.

6. xAI Officially Becomes SpaceXAI as Musk Consolidates Empire

The company formerly known as xAI now refers to itself as SpaceXAI following SpaceX’s acquisition of xAI (which also owns X/Twitter). Elon Musk confirmed that “xAI will be dissolved as a separate company” and its AI products will be absorbed into SpaceX. The combined entity is valued at $1.25 trillion, and Musk has filed with the FCC for approval to launch up to 1 million data-center satellites into orbit, arguing that space-based AI compute will be the lowest-cost option within 2-3 years. Tesla separately disclosed a $2 billion investment in xAI as part of its recent earnings report.

7. Snap Confirms Perplexity AI Search Partnership Has Ended

Snap told investors in its Q1 2026 letter that its relationship with Perplexity ended “amicably,” and analysts should not expect any revenue contribution from the partnership going forward. Perplexity was originally slated to power AI search within Snapchat, but the deal never materialized into a revenue-generating product. Snap also hinted at more news about its “intelligent eyewear” Specs product coming in June, keeping its hardware AI ambitions alive despite the Perplexity setback.

Trend Watch

What to Watch

• Musk v. Altman closing arguments expected next Thursday — the trial has produced a treasure trove of internal emails and testimony about OpenAI’s founding, governance, and the events leading to Altman’s firing and return. A verdict could reshape the future of OpenAI’s structure and its relationship with Microsoft.
• Anthropic’s orbital compute ambitions — the company’s expressed interest in space-based AI data centers alongside SpaceX’s FCC filing for 1 million compute satellites could signal a radical new direction for how and where AI models are trained.
• Snap’s “intelligent eyewear” reveal in June — after the Perplexity partnership fell through, Snap is doubling down on its own hardware AI strategy. The upcoming Specs announcement could signal a new battleground for consumer AI beyond chatbots.