In partnership with
Hello from The AI Night,
Today in AI:
Anthropic Announces $30B Run-Rate Revenue and Google Compute Deal
OpenAI Launches a Safety Fellowship for External Researchers
Google Launches Gemma 4 for On-Device AI
Here's the deal: Anthropic announced a new partnership with Google and Broadcom for multiple gigawatts of next-generation TPU capacity, expected to come online starting in 2027. The company also disclosed that its run-rate revenue now exceeds $30 billion, up from roughly $9 billion at the end of 2025.
The Breakdown:
Enterprise customers spending over $1 million annually doubled from 500+ to over 1,000 in under two months.
The majority of new compute will be located in the United States, expanding Anthropic's $50 billion U.S. infrastructure commitment from November 2025.
The deal builds on existing Google Cloud TPU capacity announced last October.
Anthropic trains and runs Claude across AWS Trainium, Google TPUs, and NVIDIA GPUs, matching workloads to optimal hardware.
Claude remains the only frontier model available on all three major clouds: AWS Bedrock, Google Vertex AI, and Microsoft Azure Foundry.
The bigger picture: A 3x revenue jump in roughly four months signals that enterprise AI adoption is accelerating faster than most forecasts assumed. The multi-cloud, multi-chip strategy gives Anthropic unusual infrastructure flexibility as compute demand scales.
Here's the deal: OpenAI announced the OpenAI Safety Fellowship, a program inviting external researchers, engineers, and practitioners to conduct safety and alignment research on advanced AI systems. The fellowship runs from September 14, 2026 through February 5, 2027.
The Breakdown:
Priority research areas include safety evaluation, robustness, scalable mitigations, agentic oversight, privacy-preserving safety methods, and high-severity misuse domains.
Fellows work with OpenAI mentors and a peer cohort, with workspace available in Berkeley at Constellation or remotely.
Each fellow is expected to produce a substantial output such as a paper, benchmark, or dataset.
The program includes a monthly stipend, compute support, API credits, and mentorship, but no internal system access.
Backgrounds welcomed range from computer science to social science, cybersecurity, privacy, and HCI.
Applications close May 3, 2025, with decisions by July 25.
The bigger picture: This gives independent researchers structured access to OpenAI resources for safety work without requiring employment. It signals a push to expand safety research beyond internal teams, which could strengthen external oversight capacity across the field.
Here's the deal: Google released Gemma 4 on April 2, 2026, under the Apache 2.0 open-source license. The model family is built to run locally on everyday devices, delivering Gemini 3-level performance without requiring an internet connection.
The Breakdown:
Models range from 2.5 GB (mobile) to 31B parameters (desktop), targeting on-device deployment across hardware tiers.
Supports a 256,000-token context window, enabling long-document and multi-turn tasks locally.
Capabilities include reasoning, image analysis, multilingual queries, and agentic tasks like sleep tracking and trend logging.
When connected, Gemma 4 can also make API calls, extending its utility beyond offline use.
Available now through the Google AI Edge App on iOS and Android.
Early developer feedback highlights strong privacy and reliability, though some report hallucinations and elevated battery consumption.
The bigger picture: Gemma 4 makes capable, agentic AI available offline and on-device at zero API cost. For developers building privacy sensitive or connectivity-limited applications, this removes the dependency on cloud infrastructure entirely.
The fastest-growing repo on GitHub is a one person team!
OpenClaw went from 9K to 185K GitHub stars in 60 days — the fastest-growing repo in history.
Their docs? One person, plus Claude. They scaled to the top 1% of all Mintlify sites, shipping 24 documentation updates a day.
What else you need to know:
Cursor introduced "warp decode," a new MoE inference method that flips parallelism from experts to outputs, achieving 1.84x faster throughput and 1.4x better accuracy on NVIDIA B200 GPUs.
Telegram's latest update adds a fully private on-device AI message editor, 12 new poll features, Live Photos support, and a one-tap system letting AI bots create other bots for users.
GitHub user GitFrog1111's open-source VS Code extension BadClaude, which overlays a whip graphic and sends cheeky prompts to speed up slow Claude Code sessions, went viral with millions of views.
Developer Farza released Clicky, an AI screen companion powered by Claude and ElevenLabs that watches your screen via hotkey-triggered screenshots, talks to you, and visually points at elements to teach software skills in real time.
X announced its API will stay free for apps delivering emergency alerts and public service updates, with a new Public Utility developer program offering expanded access and clearer guidelines coming soon.
That’s it for today’s edition of The AI Night.
Our goal is to cut through the noise, surface what actually changed, and explain why it matters.
3 ways to support us:
Forward this to your AI-curious friend → https://www.theainight.com
Sponsor The AI Night and reach 500+ AI builders daily → passionfroot.me/theainight
Reply to this email — I read every response