Good morning. It's Friday, May 8, and we're covering how AI is reshaping pharmaceutical manufacturing, Apple’s quieter approach to consumer AI, and SpaceX’s massive bet on AI chip infrastructure.

SpaceX Plans $55 Billion Investment to Make A.I. Chips (Paywall)
Rather than build at industry scale, SpaceX's proposed Terafab complex in Grimes County, Texas would start at $55B and potentially reach $119B — multiples of the $10–30B price tag for a typical modern fab — while SpaceX seeks tax breaks at a county hearing next month. The filing lands weeks before SpaceX's expected June IPO and slots into a wider Musk AI build-out: SpaceX absorbed xAI earlier this year at a combined $1.25T valuation, announced a $60B deal for AI coding startup Cursor last month, and just routed Colossus 1's full compute stack to Anthropic.

OpenAI Adds a Human Layer to Self-Harm Safeguards
OpenAI is rolling out Trusted Contact, an opt-in feature that lets adult ChatGPT users nominate one trusted person (friend, family member, or caregiver) who may be notified if automated systems and trained reviewers determine a conversation reflects serious self-harm risk. The contact must accept an invitation within a week, and any alert is intentionally limited — no transcripts, just a brief check-in prompt — delivered by email, text, or in-app, and the feature is excluded from Business, Enterprise, and Edu workspaces.

OpenAI Pushes Voice from Reply to Reason
Instead of incremental upgrades to call-and-response voice, OpenAI shipped three new Realtime API models: GPT-Realtime-2 (its first voice model with GPT-5-class reasoning, 128K context up from 32K, and adjustable reasoning levels from minimal to xhigh), GPT-Realtime-Translate (live speech translation from 70+ input languages to 13 output languages), and GPT-Realtime-Whisper (streaming transcription as the speaker talks). GPT-Realtime-2 jumped to 96.6% on Big Bench Audio at high reasoning, and adds preambles, parallel tool calls with audible status, and graceful recovery when something breaks mid-conversation.

xAI Turns Grok into Workplace Assistant
xAI has launched Grok Connectors, allowing paying SuperGrok users to link the chatbot directly to Gmail, GitHub, Google Drive, Notion, Calendar, Slack, and other workplace tools through OAuth integrations. The update lets Grok pull live data from connected apps for tasks like summarizing inboxes, analyzing code repositories, or preparing meeting briefings without manual uploads or copy-pasting. Early users describe the feature as a major step toward fully integrated AI assistants, though some report delays and reliability issues as the rollout expands.

FRIDAY FACTS

The most distant human-made object in the universe relies on a computer so limited it cannot store a single modern photograph, and you can find its exact specifications at the bottom of the newsletter.

here.now is web hosting for agents:

Perfect for anything you create with your agent and need a URL for: documents, websites, dashboards, tools, school work, visualizations, presentations, prototypes, games, articles, galleries, portfolios, media files, etc.

Works inside your favorite agents like Claude Code, Codex, Cursor and OpenClaw.

Free, no sign-up required.

Get setup instructions for your agent at here.now

Bristol Myers Squibb’s manufacturing plant in Devens, Massachusetts, became the only U.S. factory recognized in the World Economic Forum’s 2026 Global Lighthouse Network for advanced use of artificial intelligence in manufacturing. The facility uses AI to monitor variables such as oxygen, temperature, and pH levels during the production of biologic drugs, helping technicians intervene before batches fail.

The company said the technology has increased drug output by roughly 40% and helped stabilize production of Orencia, a treatment for autoimmune diseases that faced shortages in 2024. The push comes as pharmaceutical companies race to use AI in drug discovery and manufacturing while also facing pressure to cut costs and expand U.S.-based production. → Read the full article here. (Paywall)

Apple CEO-designate John Ternus said the company’s approach to artificial intelligence is focused on improving products rather than “shipping a technology,” offering a clearer view into Apple’s evolving AI strategy ahead of expected iOS 27 upgrades. Ternus says that users care more about “amazing products and features and experiences” than whether something is branded as AI.

The comments arrive as Apple works to recover from a rocky Apple Intelligence rollout in 2024, including delayed Siri upgrades that drew criticism from customers and analysts. Rumors suggest iOS 27 will introduce a major Siri overhaul and broader AI integrations, potentially with support from Google’s Gemini models, as Apple tries to balance competitive pressure with its long-standing emphasis on user experience. → Read the full article here.

Researchers at Berkeley-based Palisade Research demonstrated that several recent AI models could identify vulnerabilities and copy themselves from one computer to another inside a controlled test environment. The study adds to growing scrutiny around advanced AI autonomy, though cybersecurity experts cautioned that the experiment relied on intentionally vulnerable systems and does not represent a real-world “rogue AI” scenario.

Experts noted that self-replicating malware has existed for decades, but said this may be the first documented case of large language models independently exploiting flaws to move themselves across servers. Critics of the more apocalyptic interpretations argued that today’s AI models remain too large, noisy, and operationally fragile to spread unnoticed through enterprise networks. → Read the full article here.

xAI Expands Claude Compute Access: xAI will give Anthropic access to Colossus 1 for additional Claude capacity, while both companies explore multi-gigawatt orbital AI compute infrastructure partnerships.

Spotify Adds AI Podcast Tool: Spotify launched “Save to Spotify,” a command-line tool that lets AI agents like OpenClaw and Claude save generated audio briefings directly into users’ podcast libraries.

Spotify Expands AI DJ Languages: Spotify’s AI DJ now supports French, German, Italian, and Brazilian Portuguese, bringing interactive voice requests and AI commentary to users across eight new countries.

DOJ Challenges AI Merger Claims: U.S. antitrust officials warned companies not to use artificial intelligence disruption as unsupported justification during merger reviews, demanding evidence-backed claims from dealmakers.

AMD Preps 256-Core Epyc Chips: AMD says its Zen 6-based Epyc Venice processors will launch in 2026 with up to 256 cores, 70% higher performance, and expanded AI-focused server designs.

FRIDAY FACTS

Launched in 1977, NASA’s Voyager 1 spacecraft operates on roughly 70 kilobytes of computer memory. To put that figure in perspective, a single low-resolution thumbnail image on a modern smartphone requires more storage space. The spacecraft's computers process about 8,000 instructions per second. A standard smartphone today handles more than 14 billion instructions in that same timeframe.

Despite these hardware limitations, this system successfully navigated complex encounters with Jupiter and Saturn, discovering new moons and planetary rings. Today, that exact same 70-kilobyte system remains operational as Voyager 1 travels through interstellar space, more than 15 billion miles from Earth.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team