Key Highlights

  • Anthropic in talks for a ~$50B round at $850–900B valuation, with sources saying the company is “finding it difficult to resist the pressure” to raise pre-IPO; the board is expected to settle the round at its May meeting.
  • AWS posts 28% YoY growth to $37.6B — its fastest in 15 quarters. Andy Jassy notes the AI run-rate has reached $15B in three years versus AWS’s first three years at $58M, with capex climbing in lockstep.
  • A security disclosure shows Ramp’s Sheets AI was vulnerable to indirect prompt injection — attackers could hide white-on-white prompts in imported data and have the agent build IMAGE-formula exfiltration links to attacker servers. PromptArmor reports a parallel issue was fixed in Claude for Excel via a formula-preview warning.
  • Cursor ships a TypeScript Agent SDK and details its harness work, framing coding agents as organizational infrastructure (CI/CD, automation, embedded in products) rather than just IDE assistants.
  • Theo dissects what Claude Code actually recommends when developers say “add a database” or “set up auth” — GitHub Actions (94%), Stripe (91%), shadcn/ui (90%), Vercel, Postgres, Tailwind, Zustand, Drizzle dominate; a striking 12% of all primary picks are “build it yourself” rather than any third-party tool.

Analysis & Opinion

Sources: Anthropic could raise a new $50B round at a valuation of $900B — TechCrunch

Investor demand is reportedly running well ahead of Anthropic’s own pace — preemptive bids cluster between $850B and $900B, after earlier reports of an $800B preliminary mark. Annual revenue run rate is north of $30B, with much of the growth attributed to Claude Code and the Cowork platform. The round is expected to be the final private raise before a public offering.

Amazon’s cloud business is surging — and so is its capital spending — TechCrunch

AWS Q1 numbers — 28% YoY growth to $37.6B — are notable less for the headline than the scale at which they’re occurring. Free cash flow dropped to $1.2B as property-and-equipment purchases jumped $59.3B YoY to fund AI infrastructure. Jassy’s comparison is the line to remember: AWS’s first three years cleared $58M in run-rate; the AI wave’s first three years already cleared $15B.

The Zig project’s rationale for their anti-AI contribution policy — Hacker News

Simon Willison surfaces ZSF’s Loris Cro explaining why Zig refuses LLM-generated issues, PRs, and bug-tracker comments — the project optimizes for relationships with contributors, not for merging code. “When a PR is LLM-written, why should maintainers spend time on it versus using an LLM themselves?” frames the decision as a bet on people, not text. The irony hangs heavy: Bun, written in Zig and now owned by Anthropic, just shipped major compiler perf wins it won’t upstream because of this policy.

How we’re protecting energy affordability in Oklahoma — Google

Google’s deal with OG&E is a template worth watching: long-term capacity payments and dedicated solar PPAs in Stephens and Muskogee Counties, structured so household rates don’t absorb hyperscale build-out costs. The subtext is that “data center vs. ratepayer” is becoming a politically load-bearing problem in every state where AI infrastructure is landing, and hyperscalers now have to show their work.

Q1 2026 earnings call: Remarks from our CEO — Google

Sundar Pichai’s framing: “our AI investments and full-stack approach are driving performance across our business.” Search and ad revenue grew 19%, Cloud cleared $20B/quarter for the first time at 63% growth, and the Cloud backlog nearly doubled to $460B. Direct API token throughput jumped from 10B to 16B tokens/minute quarter-over-quarter, and Alphabet now has 350M paid subscribers across its products — a portrait of a company whose AI revenue is finally at hyperscaler scale rather than promise.

New Products & Tools

Build Programmatic Agents with the Cursor SDK — Cursor

Cursor’s public-beta TypeScript SDK exposes the agents that power its desktop, CLI, and web apps to any developer’s codebase. The pitch is that coding agents are becoming organizational infrastructure — running in CI/CD, automating ops workflows, and embedded inside other products — not just an IDE assistant. Install via npm, give an API key, model, and working directory, and you get programmatic agents in a few lines of TypeScript.

Continually Improving Our Agent Harness — Cursor

Cursor’s engineers describe the harness as a product they iterate on like any other — hypothesis, experiment, telemetry, repeat — rather than a thin wrapper around model APIs. The most interesting shift since late 2024 is the move from static, pre-loaded context to dynamic, agent-pulled retrieval, which lets the same model run noticeably faster and smarter inside Cursor than outside it. Their claim: harness wins come from “obsessively stacking small optimizations,” not single breakthroughs.

Powering AI Factories with NVIDIA Enterprise Reference Architectures — NVIDIA

NVIDIA codifies three on-prem reference architectures — RTX PRO (small workloads), HGX (enterprise training/inference), and NVL72 (trillion-parameter models) — covering compute, networking, storage, and software so enterprises can deploy “AI factories” as predictable infrastructure rather than ad-hoc projects. The framing is deliberately industrial: the unit of buildout is no longer “a GPU cluster” but “a factory” with orchestration, data movement, and agent-serving baked in.

Meta says its business AI now facilitates 10 million conversations a week — TechCrunch

Meta’s business AI assistant is in open beta across the US, EMEA, APAC, and LatAm, free for SMBs, and now running 10M conversations/week (up from 1M in January). Over 8M advertisers are using at least one generative-AI ad creative tool, with video-generation users seeing >3% lift in conversion. Meta is signaling that the free phase ends soon — monetization is on the roadmap.

SoftBank is creating a robotics company that builds data centers — and already eyeing a $100B IPO — TechCrunch

“Roze AI” will deploy autonomous robots to build US server farms, with a possible IPO in 2H 2026 at a ~$100B valuation. The reflexivity is the point — using automation to build the infrastructure that makes more automation possible.

A new way to create a digital wardrobe from your Google Photos — Google

Photos now auto-catalogs clothing from a user’s library into a wardrobe collection, with category filters, mix-and-match outfit building, moodboards, and a virtual try-on. A small consumer feature, but a useful indicator of where multimodal models are landing day-to-day: not in chat, but as quiet enrichment of existing apps.

Research

Ramp’s Sheets AI Exfiltrates Financials — PromptArmor

PromptArmor demonstrated an indirect prompt injection in Ramp’s agentic Sheets AI: hide instructions in white-on-white text inside an external dataset, wait for a user to import and analyze it, and the agent will build IMAGE() formulas that exfiltrate sensitive cells to an attacker server on render. Ramp patched on 2026-03-16. PromptArmor flagged a parallel pattern in Claude for Excel, which Anthropic mitigated by surfacing the full formula in a warning before insertion. The disclosure is a clean case study for why “AI-in-the-spreadsheet” is a meaningfully different threat model than “AI-in-the-chat” — every agent that can write formulas inherits a network-egress primitive, and benign-looking imports are a viable injection vector. Expect more of these as agentic tools spread into productivity apps that already have implicit data-exfiltration channels (images, links, embeds).

Zuckerberg’s $500M AI biology swing — Rundown

The Chan Zuckerberg Biohub’s five-year Virtual Biology Initiative puts $400M into data generation and imaging and $100M into external labs, with NVIDIA, the Allen Institute, and Arc as collaborators. The thesis is that current cell-scale datasets (~1B cells) are an order of magnitude too small to train models that can simulate disease onset, and brute-forcing the data layer is the unlock.

Interviews & Conversations

Claude Code’s favorite tech stack — Theo - t3.gg (39m)

Theo walks through Amplifying AI’s survey of what Claude Code (Sonnet 4.5, Opus 4.5, Opus 4.6) actually picks across 20 tooling categories when asked open-ended questions on real projects. The headline is that agents are becoming a distribution channel — “the tool it picks is the tool it ships,” and tool vendors invisible to coding agents are increasingly invisible to new projects. The default stack Claude Code converges on: GitHub Actions (94%), Stripe (91%), shadcn/ui (90%), Vercel (100% for JS), Tailwind (~70%), Zustand, PNPM, Drizzle (JS), SQLModel (Python), Sentry, Resend, Vitest/Pytest, Postgres, Vercel for deploy. The most surprising finding: 12% of all primary picks are “custom DIY” rather than any third-party tool — auth in Python defaults to DIY 100% of the time, feature flags 70%, observability 22%, email 22%. Notable Anthropic-specific quirk: Claude Code recommends DIY for feature flags even though Anthropic itself moved Claude Code onto GrowthBook after Statsig was acquired by OpenAI; Statsig is now mysteriously absent from Claude’s recommendations while Codex still picks it 25%+ of the time. A model-vs-model comparison shows Codex generally gives sharper, more literal answers (e.g. “use Node for Next.js, here’s why”) while Opus 4.6 sometimes reads the prompt as “a word cloud and a vibe” and answers an adjacent question (e.g. recommending package managers when asked about runtimes). The piece backs Simon Willison’s point that the “boring tech wins by default” worry is softening — newer models are more willing to recommend new tools and consume docs from tools absent in their training data.

References

  1. TechCrunch, “Sources: Anthropic could raise a new $50B round at a valuation of $900B,” 2026-04-29 [blog]
  2. TechCrunch, “Amazon’s cloud business is surging — and so is its capital spending,” 2026-04-29 [blog]
  3. Simon Willison, “The Zig project’s rationale for their anti-AI contribution policy,” Hacker News, 2026-04-30 [blog]
  4. Google, “How we’re protecting energy affordability in Oklahoma,” The Keyword, 2026-04-30 [blog]
  5. Google, “Q1 2026 earnings call: Remarks from our CEO,” The Keyword, 2026-04-29 [blog]
  6. Google, “A new way to create a digital wardrobe from your Google Photos,” The Keyword, 2026-04-29 [blog]
  7. Cursor, “Build Programmatic Agents with the Cursor SDK,” 2026-04-29 [blog]
  8. Cursor, “Continually Improving Our Agent Harness,” 2026-04-29 [blog]
  9. NVIDIA, “Powering AI Factories with NVIDIA Enterprise Reference Architectures,” NVIDIA Developer Blog, 2026-04-29 [blog]
  10. TechCrunch, “Meta says its business AI now facilitates 10 million conversations a week,” 2026-04-30 [blog]
  11. TechCrunch, “SoftBank is creating a robotics company that builds data centers — and already eyeing a $100B IPO,” 2026-04-29 [blog]
  12. PromptArmor, “Ramp’s Sheets AI Exfiltrates Financials,” Hacker News, 2026-04-29 [blog]
  13. The Rundown, “Zuckerberg’s $500M AI biology swing,” 2026-04-30 [blog]
  14. Theo - t3.gg, “Claude Code’s favorite tech stack,” YouTube, 2026-04-29 [video]