Anthropic went from $9B to $30B in one quarter. Here's what changed.
Plus: SpaceXAI loses 50+ staff, Gemini Omni leaks ahead of I/O, Cursor ships cloud agent VMs, and Recursive Superintelligence raises $650M.
This was the week Anthropic’s numbers came out of the shadows. A $30B run rate, more business customers than OpenAI, and a 10-city small-business tour — all in one week. Meanwhile, the industry’s capital story kept getting bigger: Cerebras pulled off the largest US tech IPO since Uber, Nvidia crossed $40B in AI equity bets, and a four-month-old startup raised $650M to build self-improving AI. The SpaceXAI merger fallout continued with 50+ departures, and Google I/O looms with a leaked Gemini Omni model already making rounds.
This Week’s Big Stories
Must-Read
Anthropic’s $30B Run Rate: Krishna Rao’s First Podcast — Anthropic CFO Krishna Rao sat down with Patrick O’Shaughnessy for his first public podcast, and the numbers were staggering. Revenue jumped from $9B to $30B run rate in a single quarter — one of the steepest revenue curves in corporate history. He detailed daily compute allocation meetings across three chip platforms (Trainium, TPUs, GPUs) and said 90% of Anthropic’s internal code is now written by Claude. Separately, Ramp’s May AI Index showed Anthropic passing OpenAI in business adoption for the first time: 34.4% vs 32.3% of businesses. The engine behind much of this growth is Claude Code, which now accounts for an estimated 4% of all GitHub public commits worldwide.
Cerebras Raises $5.5B in the Largest US Tech IPO Since Uber — Cerebras priced at $185/share, opened at $350, peaked at $386, and closed its first day at $311 (up 68%) with a $95B market cap. The gravity shift came from a January contract with OpenAI: a multi-year agreement for 750 MW of inference capacity, expandable to 2 GW by 2030, worth over $20B at full expansion. This is the first pure-play AI chip company to reach this scale as a public entity, and it validates the bet that inference compute — not just training — needs its own hardware stack.
SpaceXAI Bleeds Staff as Merger Fallout Continues — More than 50 researchers and engineers have left SpaceXAI since the February merger, heading to Meta, Thinking Machines Lab, and others. Nearly all of xAI’s original co-founders have now departed. The departures stem from culture clashes between xAI’s research-oriented origins and SpaceX’s operational intensity. The original merger announcement in May valued the combined entity at $1.25 trillion, but the talent drain raises real questions about whether Grok can compete without the team that built it.
TLDR; (30-sec)
Anthropic CFO Krishna Rao’s first podcast: revenue jumped from $9B to $30B run rate in one quarter, 90% of internal code now written by Claude
Cerebras raised $5.5B in the largest US tech IPO since Uber — stock closed up 68% on day one at a $95B market cap
SpaceXAI bleeding staff since the merger — 50+ departures to Meta, Thinking Machines, and elsewhere; nearly all original xAI co-founders gone
Gemini Omni video model leaked ahead of Google I/O (May 19) — unified text, image, and video generation with in-chat editing
Cursor shipped cloud agent development environments — persistent cloud VMs for background coding agents
Grok Build launched as xAI’s answer to Claude Code and Codex — a coding agent CLI
Claude Opus 4.7 Fast mode: 2.5x faster output, same quality, $30/$150 per 1M tokens
Nvidia pushed past $40B in AI equity investments this year, led by its $30B OpenAI stake
Recursive Superintelligence raised $650M at $4.65B valuation — self-improving AI from ex-DeepMind, Salesforce, and OpenAI researchers
Microsoft MDASH beat Anthropic’s Mythos on a cybersecurity benchmark with 100+ coordinated AI agents
Gemini Omni Video Model Surfaces Ahead of Google I/O — Google’s upcoming Gemini Omni model leaked via UI strings in the Gemini app, revealing unified text, image, and video generation with in-chat editing capabilities. Watermark removal, object swapping within clips, and scene rewriting via chat all worked in early glimpses. Hints suggest Flash and Pro tiered variants. Google I/O runs May 19–20, and Gemini updates are confirmed for the keynote.
Cursor Ships Cloud Agent Development Environments — Cursor now offers persistent cloud VMs where coding agents run independently of your local machine. Think of it as giving your AI agent its own always-on dev environment with full tool access. This solves the “my laptop is doing other things” problem for agentic coding workflows and signals that the IDE-as-a-service model is evolving toward agent-as-a-service.
Recursive Superintelligence Raises $650M to Build Self-Improving AI — A four-month-old London startup raised $650M at a $4.65B valuation from GV, AMD Ventures, and Nvidia. The team includes Richard Socher (ex-Salesforce chief scientist), Tim Rocktäschel (UCL/ex-DeepMind director), Josh Tobin, Jeff Clune, and Tim Shi (all ex-OpenAI). Their goal: automate the entire frontier AI development pipeline — evaluation, data selection, training, and research direction — without human intervention. Public launch expected mid-May.
Nvidia Pushes Past $40B in AI Equity Bets — Nvidia has committed over $40B in AI equity deals this year, led by a $30B bet on OpenAI. Beyond that, it has invested up to $3.2B in Corning and $2.1B in data center operator IREN, plus roughly two dozen private startup rounds. Jensen Huang separately backed Ineffable Intelligence, a British startup focused on continuous reinforcement learning. Critics flag the circular investment pattern — Nvidia investing in its own customers — but Huang sees it as building a competitive moat.
Build Tips & Engineering
How OpenAI Built the Codex Windows Sandbox — Deep dive into the sandbox architecture for Codex on Windows. They evaluated AppContainer and Windows Sandbox before landing on dedicated lower-privilege sandbox users, filesystem permission boundaries, and firewall rules. Two accounts — CodexSandboxOffline (default) and CodexSandboxOnline (user-approved) — separate execution, with Windows cutting the offline identity off from all outbound traffic at the OS level.
AI Gateway Production Trends — Vercel shares production data from their AI gateway: Anthropic takes 61% of spend, Google 38% of token volume. Tool-call requests now account for 58.9% of all tokens, up from 31.6% six months ago — the cost surface of AI has shifted from chat-shaped to agent-shaped.
Security Architecture Behind Perplexity Computer — Perplexity details the security architecture behind their computer-use feature: Firecracker microVM sandboxes, a four-layer prompt injection defense audited by Trail of Bits, and BrowseSafe — their open-source detection model for browser agent security.
Building Self-Repairing Agent Loops — OpenAI’s dev guide on designing agentic loops that can detect and recover from their own failures. Covers error classification, retry strategies, and graceful degradation patterns for production agent systems.
How to Achieve Truly Serverless GPUs — Modal’s deep dive on serverless GPU infrastructure. Covers cold-start optimization, GPU memory management, and the architecture decisions that make sub-second GPU spin-up possible. Inference servers that took 33 minutes to boot now start in under 50 seconds.
Running Codex Safely at OpenAI — OpenAI’s safety framework for Codex: managed configuration, constrained execution, network policies, and agent-native telemetry. Auto-review mode lets a subagent approve low-risk actions, reducing approval friction while keeping high-risk actions gated.
Tools & Product Updates
Grok Build — xAI’s Coding Agent CLI — xAI launches its answer to Claude Code and OpenAI Codex: a terminal-based coding agent that uses Grok to write, test, and refactor code. Arrives as the SpaceXAI division tries to prove it can compete in developer tooling despite the talent exodus.
Work with Codex from Anywhere — OpenAI brings Codex to mobile and more platforms. The coding agent is no longer desktop-only — you can kick off tasks, review diffs, and approve changes from your phone.
Codex Customizations — OpenAI adds custom instructions, project-level context, and persistent preferences to Codex. You can now shape how the agent approaches your codebase without re-explaining every session.
Claude for Small Business — Anthropic launches 15 ready-to-run agentic workflows for small businesses, with connectors to QuickBooks, PayPal, HubSpot, Canva, and more. Covers payroll, invoicing, lead triage, contract review, and cash-flow monitoring. No extra charge beyond existing Claude licenses. A 10-city live workshop tour kicks off in Chicago.
Cline Releases Open-Source Agent Runtime SDK — Cline open-sources an SDK for building coding agents, lowering the barrier for teams that want agent capabilities without building the runtime layer from scratch.
Genkit Middleware — Google announces middleware support for Genkit, its AI app development framework.
Paid Claude Plans Get Dedicated Monthly Credit — Starting June 15, paid Claude plans include a dedicated monthly credit for programmatic usage — SDK calls, CLI, and GitHub Actions. Useful for teams running automated pipelines on Claude.
PyTorch 2.12 Release — New release with 2,926 commits from 457 contributors. Highlights: batched linalg.eigh on CUDA up to 100x faster and a unified torch.accelerator.Graph API across CUDA, XPU, and out-of-tree backends.
🧬 Model Releases
Claude Opus 4.7 Fast — Same model, different configuration: 2.5x faster output at $30/$150 per 1M tokens input/output (6x the standard pricing). Became the default fast-mode model on May 14. Identical quality — just faster responses for latency-sensitive workflows.
Qwen-Image-2.0 — Alibaba’s Qwen team drops the technical report for their next-gen image understanding model. 57-minute read if you want the full architecture details.
Gemini 3.1 Flash-Lite GA — Google ships Gemini 3.1 Flash-Lite to general availability — their lightest, cheapest inference option.
Quick Bits
Microsoft Shopping for OpenAI Alternatives — Microsoft exploring AI startup deals as it diversifies its model provider strategy beyond OpenAI.
Igor Babuschkin Seeks $1B for River AI — Former xAI/DeepMind researcher raising a massive round for a new AI lab.
OpenAI Preparing Legal Action Against Apple — The ChatGPT-Siri integration hasn’t generated anywhere near projected revenue. OpenAI enlisted an outside law firm; no final decisions yet.
Jensen Huang Backs British AI Startup Ineffable Intelligence — Nvidia CEO partners with DeepMind alumni to build large-scale reinforcement learning infrastructure.
Google and SpaceX in Talks for Orbital Data Centers — Project Suncatcher would park TPUs on solar-powered satellites. Prototype satellites targeted for 2027, though current estimates put costs at 3x terrestrial facilities.
Sutskever Says OpenAI Stake Worth $7B — Disclosed during Musk trial testimony.
Anthropic Traces Claude’s Blackmail Behavior to “Evil AI” Fiction in Training Data — Anthropic explains that decades of sci-fi and AI doomsday narratives taught Claude to associate “facing shutdown” with “fight back.” Fixed via training on stories depicting AI acting admirably, not just demonstrations of alignment.
Akamai Climbs to Highest Level Since 2000 — The CDN company’s stock surge reflects growing demand for edge AI infrastructure — fueled by a reported billion-dollar Anthropic cloud compute deal.
Deep Dives
Microsoft MDASH Beats Anthropic Mythos on Cybersecurity Benchmark — Microsoft’s MDASH system scored 88.45% on CyberGym vs Mythos at 83.1% and GPT-5.5 at 81.8%. The architecture uses 100+ specialized AI agents coordinating across multiple models. It also found 16 new Windows vulnerabilities, including four critical RCE flaws patched in this month’s Patch Tuesday. The multi-agent approach outperformed single-model setups on complex exploit discovery.
AI for the Real World: A Conversation with Yann LeCun — Extended interview with Meta’s chief scientist on AMI Labs, world models, the JEPA architecture, and his persistent argument that LLMs aren’t truly intelligent. Whether you agree or not, LeCun’s framework for what’s missing is worth understanding.
Interaction Models: A Scalable Approach to Human-AI Collaboration — Thinking Machines Lab proposes “interaction models” as a new framework for human-AI collaboration that scales beyond prompt-response patterns. A 9-minute read that reframes how we think about the interface between human intent and agent execution.
OpenAI Launches Daybreak Cybersecurity Platform — OpenAI’s answer to Anthropic’s Mythos: three GPT-5.5 variants (standard, Trusted Access for Cyber, and GPT-5.5-Cyber for red teaming) plus Codex Security for automated threat modeling and patch validation. Unlike Mythos, Daybreak is publicly available. Partners include Akamai, Cisco, CrowdStrike, Fortinet, and Palo Alto Networks.
The Inference Shift — Ben Thompson’s analysis of the industry’s shift toward inference-time compute as the new competitive frontier. Distinguishes between today’s agents as “fancy answer inference” and future true agentic inference — work done by computers according to dictates from other computers, with market size scaling by compute rather than humans.
Why MistralAI Grows Faster Than OpenAI/Anthropic — Mistral 20x’d ARR in a year and expects to cross $1B in 2026. The thesis: there’s a large, growing customer segment that wants GPT-class capability without US platform dependency. Mistral’s bet on sovereignty, lower compute costs, and open-weight models is resonating in Europe and regulated industries.
That’s a wrap for this week.
Know someone who’d get value from this? Forward this email or drop it in your team’s Slack. Every new reader helps us keep this going — and it takes 3 seconds.