Anthropic grew 80-fold in one quarter. Its own infrastructure was not ready for that.

article · anthropic · may 08

Why for you: The headline is the bluntest single-quarter growth figure in recent AI memory: Anthropic's revenue and usage grew 80-fold on an annualized basis from Q4 to Q1, a number Dario Amodei described as "just crazy" and "too hard to handle." The result was an infrastructure scramble that led Anthropic to rent compute from Elon Musk's xAI Colossus 1 cluster — hundreds of thousands of H100s sitting on the other side of a bitter industry rivalry. The real signal is organizational: they'd planned for 10x. Being off by 8x changes what you build, who you hire, and which compute contracts you sign. This is the context for every Anthropic partnership announcement you'll see this quarter.

analysis · stratechery · may 11

Why for you: Thompson's latest is nominally about Cerebras Systems' IPO pricing surge, but it's a clean structural analysis of why the GPU-dominated compute market of the training era is about to fracture. Training is highly parallelizable — GPUs are ideal. Inference is not: prefill, decode, and KV-cache-read steps have very different hardware demands. That heterogeneity means the winner-take-all dynamics that gave Nvidia its moat are unlikely to persist at the inference layer. For anyone thinking about which AI infrastructure bets will compound over the next three years, this is the mental model to internalize before the next wave of chip company announcements.

article · openai · may 12

Why for you: Daybreak combines GPT-5.5, Codex Security, and a partner network including Cloudflare, Cisco, CrowdStrike, and Palo Alto Networks to deliver AI-powered vulnerability scanning, threat modeling, and patch validation. It runs on three model tiers: standard GPT-5.5, a Trusted Access for Cyber variant for authorized defensive work, and a GPT-5.5-Cyber tier for red teaming. The timing is direct: Anthropic's Mythos launched to a restricted enterprise audience in January and is now in government hands. OpenAI is competing in cybersecurity as hard as it competes on general capabilities — and the economics here reward fast-patching, not frontier reasoning, which means this race has very different dynamics than the chatbot market.

article · fintech / agentic commerce · may 11

Why for you: Circle's Agent Stack — Agent Wallets, Agent Marketplace, a CLI, and a Nanopayments protocol enabling USDC transfers as small as $0.000001 at machine speed — positions Circle as the settlement layer beneath the agentic commerce flow. CEO Jeremy Allaire's framing says it directly: "Financial infrastructure has historically been built for people, with manual onboarding, approvals, and payment flows that were never designed for software acting on its own." This is the complement to what Stripe announced at Sessions — Circle is solving the hold-assets-and-pay side while Stripe solved the storefront and checkout side. Together they're building the stack for agents as economic participants.

article · google deepmind · may 07

Why for you: One year after its introduction, Google DeepMind is reporting production impact from AlphaEvolve, its Gemini-powered algorithm-design agent. The standout number: a grid optimization metric improved from 14% to 88% feasibility. AlphaEvolve-generated algorithms are now running across critical Google infrastructure, and the non-Google deployments span semiconductor manufacturing, financial services, logistics, health research, and drug discovery. Worth reading as a concrete answer to "what can agentic AI actually do in production today" — not benchmark comparisons but deployed systems making measurable improvements to real infrastructure.

essay · simon willison · may 11

Why for you: GitLab is restructuring for the agentic era: cutting its presence in up to 30% of the countries where it has small teams, eliminating up to three management layers in certain divisions, and reorganizing R&D from roughly 30 to roughly 60 smaller autonomous teams. The company's stated reasoning is that software demand will expand dramatically as production costs fall — so they're betting on an engineering-first structure that can move faster with fewer coordination layers. Willison adds the skeptical frame: GitLab's stock has halved from $52 to $26 over the past year. The restructuring may be as much cost discipline as strategic vision. Either way, this is a concrete case study in what "organizing for the agentic era" looks like operationally, not in theory.

article · openai · may 12

Why for you: Republican members of the House Oversight Committee and a coalition of state attorneys general are questioning whether Sam Altman's personal investments — in ventures that have since done business with OpenAI — blur the fiduciary line before a public offering that could value the company near $1 trillion. The more interesting detail is the source of the pressure: it's coming from Republicans, not Democrats, which complicates any clean narrative about unconditional administration support for frontier AI companies. The governance questions about Altman's dual roles have been building since his reinstatement in 2023; the IPO is the forcing function that makes them consequential rather than theoretical.

article · ai policy · may 05

Why for you: CAISI — the Commerce Department's AI Standards and Innovation center, housed at NIST — is expanding its classified model evaluation program from OpenAI and Anthropic to now include Google DeepMind, Microsoft, and xAI. The evaluations cover security properties prior to deployment and are conducted in classified environments under the Trump administration's AI Action plan. This is the emerging shape of US AI governance: not mandatory public red-teaming, but classified evaluation agreements with the largest labs — effectively a model-classification regime without calling it one. The direction of travel matters even when the current framework is formally voluntary.

essay · simon willison · may 06

Why for you: Willison has spent two years carefully distinguishing "vibe coding" (handing the wheel to AI without understanding the output) from "agentic engineering" (using AI as a tool while maintaining comprehension). His latest post, a podcast highlight reel from the High Leverage show, lands on an admission: the distinction is starting to collapse in his own practice. It's not that he's become a vibe coder — it's that the agents are good enough that maintaining full comprehension is starting to cost more than it returns in more contexts. From a staff-engineering perspective, that's the threshold worth marking: not when a benchmark is crossed, but when the people who've been most careful about the distinction notice they're working differently.