Anthropic puts $200B on Google's tab. Agents finally get their own wallet.

article · ai infra · may 05

Why for you: The Information's number that sent Alphabet up 2% after-hours: 5 GW of TPU capacity over five years, starting 2027. The single commitment is roughly 40% of the cloud backlog Google disclosed last week. The picture you were already assembling — Google as the de-facto Anthropic landlord, Anthropic as the de-facto Google Cloud tenant of record — is now priced. Useful to read alongside Ben Thompson's earlier "Anthropic-Google Alliance" piece you saved in April; the dollar figure makes the thesis concrete.

article · agentic commerce · may 07

Why for you: Bedrock now ships an agent-payments primitive — pick a Coinbase wallet or a Stripe Privy wallet, end users top up with stablecoins or a debit card, agents pay for APIs, MCP servers, content, and other agents mid-task. Built on Coinbase's x402 HTTP payment protocol. This is the wallet half of agentic commerce; Stripe's Sessions announcements were the merchant half. The three-way Stripe/Coinbase/AWS partnership is the news — quietly, the agent-economy stack just got an opinionated default.

Stripe newsroom: partnering with AWS to power AgentCore payments via Privy

article · openai · may 05

Why for you: Quietly the most consequential consumer-AI launch this week. Replaces 5.3 Instant as the default; OpenAI claims meaningfully lower hallucination on legal/medical/financial queries while keeping latency. The new variant can also reach back into prior conversations, files, and Gmail to personalize answers (Plus/Pro on web). The plumbing — sub-second answers grounded in your own corpus — is what most people will end up calling "ChatGPT" for the next quarter.

article · simon willison · may 05

Why for you: Andon Labs follow-up to the SF retail-store experiment: this time the AI ("Mona") runs a Stockholm cafe. Highlights from the Hall of Shame include 120 eggs to a stoveless cafe, 6,000 napkins, 3,000 nitrile gloves, and 9 liters of coconut milk. Willison's framing — that long-running embodied-agent failure modes are the actual eval surface, not benchmarks — is the part to keep. Worth five minutes; this is what "vibe coding for the physical world" looks like in 2026.

article · stratechery · may 07

Why for you: Stern's new book is the rare AI title written for someone who isn't already in the bubble — Ben loves it, and the conversation covers using an LLM for a career change, mammograms-and-AI, and where consumer LLMs still fail (the preying-mantis-pregnancy story is real). A useful counterweight to the heavy stack of infra/labs reading this week — what does this stuff look like to a smart Wall Street Journal columnist who isn't pre-loaded on the discourse?

Stratechery weekly roundup: Earning & Spending (May 8)

article · coding agents · may 08

Why for you: One-stop tour of the week's coding-agent moves: Coder ships a self-hosted native agent architecture for enterprises that won't put code in someone else's cloud; Snyk integrates Claude into its security platform; Opsera embeds DevSecOps agents into Cursor; Anthropic doubles Claude Code's five-hour rate limits and removes peak-hour throttling for Pro/Max. Pattern: the coding-agent layer is finally getting the boring enterprise plumbing — security, governance, self-hosting — that determines who actually wins the F500.

article · enterprise ai · may 08

Why for you: Read this as a stake in the ground for governance-as-distribution. ServiceNow stops trying to own the IDE and instead pushes its workflow/governance skills as connectors that ride inside whatever editor the developer already opened. If you're thinking about how staff/principal eng work changes when half the org runs coding agents under SOX/HIPAA, this is the operating model: the agent is wherever the engineer is, the policy lives in the platform of record.

article · derek thompson · may 05

Why for you: Adapted from David Epstein's Inside the Box — the case that constraints, not freedom, make better thinking, plus the long Malthus-to-natural-selection thread about how reading the same thing produces different ideas in different heads. Pairs neatly with how staff-eng work feels in 2026: the binding constraint moved from "what can we build" to "what should we cut," and the people doing the best work are the ones with the sharpest "no" muscle.

article · ruby on rails · may 08

Why for you: A new Rails Foundation Accessibility Guide is being drafted (semantic HTML, ARIA, Turbo, forms, testing — all together). Codebase highlights: Rails.cache.read now supports an atomic delete: true via Redis GETDEL (single-use values like OTPs); ActionController::Parameters#fetch_values mirrors Hash; Active Record Encryption can read its keys from Rails.app.creds; uniqueness errors carry :existing_id in errors.details. The :existing_id one is the small quality-of-life win you'll actually use in js-notes-style apps.