Alphabet Needs $80 Billion. Gemini's Agent Still Can't Do the Job.

Alphabet just told the world it needs $80 billion to keep the AI lights on.

That number landed Sunday night. Not as a boast, as a confession. Demand for Google's AI infrastructure is outrunning its physical ability to supply it. They're raising debt to build faster. Meanwhile, Gemini Spark, the "24/7 AI agent" Google showed off at I/O, just got its first independent hands-on review. The verdict: impressive in demo conditions, expensive in practice, and carrying privacy tradeoffs most users won't read the fine print on.

Two stories. Same company. Both worth your attention this week.

The shortlist also has OpenAI landing on AWS, Qwen 3.7 Plus hitting Vercel's AI Gateway, and Simon Willison's May roundup, which includes a line about Claude getting expensive that every operator should read. We'll get to it.

Drop count this week: 8 repos and 2 skills. One of them is the workflow platform every solo builder ignores until they're manually gluing APIs at 2am.

Let's go.

---

1. Alphabet plans to raise $80 billion for AI infrastructure, Google's parent company is tapping debt markets after confirming demand for its AI products is exceeding available supply. Enterprise appetite is driving the gap. TechCrunch

2. Gemini Spark gets a hands-on review, and a reality check, The Verge spent time with Google's new "24/7 AI agent" and found it genuinely capable in controlled conditions, but flagged real concerns about cost and privacy tradeoffs at scale. The Verge

3. OpenAI's frontier models and Codex are now on AWS, Enterprises with existing AWS procurement workflows can now build with OpenAI models without touching a separate account. Generally available as of this week. OpenAI Blog

4. Qwen 3.7 Plus hits Vercel's AI Gateway, Alibaba's latest multimodal model, vision plus language, GUI plus CLI operation, coding plus reasoning, is now routable through Vercel. One endpoint, less glue code. Vercel Blog

5. Simon Willison's May newsletter: "AI got expensive", Willison's monthly sponsor-only digest is out, and the headline note is that Claude costs have climbed fast. Worth reading if you're managing token spend at any volume. Simon Willison

---

8 repos and 2 skills this week.

[Repo] langgenius/dify, Open-source LLM app platform with a drag-and-drop RAG pipeline builder, a workflow editor, and self-host support. If you've been stitching together LangChain calls by hand, this is the platform that replaces the scaffolding, 80K+ stars, shipping fast, production-grade.

[Repo] n8n-io/n8n, Fair-code workflow automation with 400+ integrations and native AI agent nodes. The self-hosted option matters: you own the data, you own the logic, and you can wire Claude into any step without a SaaS middleman taking a cut.

[Repo] ggerganov/llama.cpp, Run quantized LLMs locally on CPU and GPU with zero Python overhead. Ggerganov's original port now supports a wide model zoo and stays updated within days of major releases, the reference implementation for local inference.

[Repo] microsoft/markitdown, Converts PDFs, Word docs, PowerPoints, and spreadsheets to clean markdown in one command. Pairs directly with any LLM pipeline that expects text input, no more writing your own doc-to-string preprocessing.

[Repo] BerriAI/litellm, One unified interface for 100+ LLM providers. Swap Claude for Qwen for GPT-4 with a single config change, add spend controls per key, and log everything to your own stack. Critical if you're running cost experiments right now.

[Repo] executeautomation/mcp-playwright, MCP server that gives Claude direct browser control via Playwright. Click, fill, screenshot, scrape, from inside a Claude conversation. The fastest path from "I need a browser agent" to a working one.

[Repo] anthropics/anthropic-cookbook, Anthropic's official collection of practical recipes: RAG patterns, tool use, vision, structured output, multi-agent coordination. Not vaporware, these are the patterns the Claude team uses internally, versioned and maintained.

[Repo] yamadashy/repomix, Packs an entire repo into a single LLM-ready file (XML, Markdown, or plain text). Paste your whole codebase into context. Works with remote repos too, npx repomix --remote owner/repo and you're done.

---

3 tools worth installing this week.

[MCP] mcp-playwright, Playwright as an MCP server, giving Claude a live browser. The non-obvious config: set headless: false during development so you can watch Claude navigate in real time and catch misclicks before you put it in production.

[Tool] LiteLLM Proxy, Run litellm as a local proxy server and route all your Claude, Qwen, and OpenAI calls through one endpoint with unified logging and per-key budget caps. With Willison's note about AI costs climbing, the spend-tracking alone is worth the setup.

[Tool] Vercel AI Gateway, If you're already on Vercel, the AI Gateway gives you model routing, fallbacks, and rate limits through the same dashboard as your deployments. Qwen 3.7 Plus just landed there, one place to test multimodal without adding another vendor.

---

One Claude setup move worth doing today.

Put your output schema in the system prompt, not the user turn.

Most people drop JSON schemas or format instructions at the bottom of the user message. Claude reads them, but its adherence is noticeably better when the schema lives in the system prompt as a named contract. The pattern:

SYSTEM: You are a contract risk analyzer. Always respond with valid JSON matching this schema: { "clauses": [{ "text": string, "risk": "high|medium|low", "reason": string, "counter": string }] } Never include prose outside the JSON object.

Then the user turn is just the raw contract text. No format reminders, no "respond only with JSON", the system prompt already owns that.

Anthropic covers structured output patterns in depth in the anthropic-cookbook, the tool_use and structured_output sections are the ones to pull from.

---

Alphabet needs $80 billion because the physical infrastructure of AI, chips, power, cooling, fiber, doesn't scale at the speed of a software product launch. Gemini Spark got a mixed review not because Google can't build agents, but because the gap between a polished demo and a product that earns trust at $20/month is still wide. That gap is exactly where solo operators have always lived: shipping the thing that actually works for one paying user before the platform companies finish their roadmap.

---

This Friday's kit is ClauseGuard, and the one-liner is: the clause letting your client keep your work unpaid is on page 4. Most freelancers sign client contracts they don't fully understand, and "I'll read it later" has cost people real money. This week we're asking the what: what does a Claude-powered contract scanner that flags every trap, explains it in plain English, and hands you a counter-clause to paste back actually look like as a product? Friday's drop has the full build, repo, prompts, schema, billing config.

That's the issue. The $80B number is the headline, but the Willison note on Claude costs is the one to sit with if you're building anything at volume this week.

See you tomorrow.