The Meltdown — April 7–13, 2026
Anthropic week. Everyone else was also there.
01 — Anthropic's Revenue Run Rate Just Hit $30 Billion
Anthropic disclosed a $30 billion annual revenue run rate, including a 58% surge in March alone, putting it ahead of OpenAI's $25 billion run rate from February. The announcement came bundled with a compute expansion: Google and Broadcom are building out gigawatts of TPU capacity earmarked for training and serving future Claude models, with the hardware coming online in 2027. Two years ago Anthropic was the scrappy safety-first underdog in a race it wasn't supposed to win commercially. Now it's out-growing the most-funded startup in history, signing infrastructure deals in gigawatts, and still insisting the mission is safety. Hard not to wonder what that tension looks like at $100 billion ARR.
02 — Anthropic's New Model Found Zero-Days in Every Major OS and Browser
Also on the Anthropic agenda this week: Project Glasswing, a coalition including Apple, Microsoft, Google, AWS, NVIDIA, CrowdStrike, and more than 40 other organizations, all given early access to a preview of Mythos, its not-yet-released frontier model. Mythos Preview spent the past few weeks finding thousands of zero-day vulnerabilities in every major operating system and every major web browser. Anthropic is committing $100 million in usage credits to the effort and $4 million to open-source security organizations. The framing is that powerful AI will accelerate cyberattacks, so the best response is using that same AI to find vulnerabilities first. It's a coherent argument, and also a company acknowledging it's building something that will make the world less safe while positioning itself as the solution. Both things are true.
03 — OpenAI, Anthropic, and Google Formed a Club Against China
Three companies that compete fiercely for the same talent, benchmarks, and enterprise dollars announced they're cooperating to stop Chinese AI companies from extracting outputs from their frontier models, a practice called model distillation. The initiative involves technical and policy coordination to detect and clamp down on unauthorized copying. This is a rare thing in the AI industry, competitors setting aside market positioning to address a shared problem. Whether it holds under commercial pressure is worth watching. The skeptical read is that the coalition formed because regulatory and national security pressure made forming it easier than not forming it. That doesn't make it meaningless; it just means the motivation is complicated.
04 — Meta Released Its First Model Under Alexandr Wang
Meta debuted Muse Spark on April 8, the first model out of Meta Superintelligence Labs, the unit run by Chief AI Officer Alexandr Wang, whose hire came with a $14.3 billion investment in Scale AI. The model accepts voice, text, and image inputs, uses parallel reasoning agents to compete with the extreme reasoning modes of GPT and Gemini, and lands, by Meta's own benchmarks, as competitive with frontier models but not ahead of them. The bigger story is what Muse Spark isn't: open-source. Meta's long-running commitment to open models is getting softer, with only vague promises about open-sourcing future versions. The company that built Llama is now hedging, and that shift matters more than any benchmark.
05 — Google Shipped Gemma 4, Its Best Open Model Yet
Google DeepMind released Gemma 4 under an Apache 2.0 license, four model variants spanning smartphones to a 31-billion-parameter server version, all with multimodal inputs, 256K context windows, and a focus on agentic reasoning. Gemma models have been downloaded more than 400 million times across all generations. Open-weight models have quietly become the workhorse of production AI, and Google releasing near-frontier capability for free keeps the flywheel spinning, even as Meta's commitment to the same idea wobbles. If open models end up being where most teams build, Google wins regardless of which closed model tops the leaderboard.