The $100B handshake

Anthropic just locked in ten years of compute with Amazon: $100 billion going to AWS, up to $25 billion coming back as investment (on top of the $8 billion Amazon had already put in). The prize is enough computing power to run a small country, and a promise that Claude won't run out of room to grow for a decade.

The AI race isn't about who has the smartest model anymore. It's about who has enough computing power to keep training and serving them. OpenAI has committed over $1.4 trillion to similar deals in the last twelve months. Amazon itself signed a $50 billion compute deal with OpenAI two months ago. The same cloud is now bankrolling the two leading labs to the tune of roughly $400 billion in forward spend. No major player is sitting this out.

Every lab is making the same bet with a different hedge. OpenAI is spreading its compute across every chipmaker willing to take the order. Anthropic is pouring it into one partner and trusting Amazon's custom chips to keep pace. Our read: Anthropic took the harder bet and the higher-upside one. Concentrate the spend, win on efficiency, and lose hard if Trainium ever slips a generation. OpenAI is buying insurance. Anthropic is playing conviction. Whichever call is right in 2027 decides who has the margin to keep shipping new products.

The real bottleneck isn't buying enough chips anymore. It's powering them. Five gigawatts is roughly the electricity draw of five million U.S. households, and that's what a single one of these deals now unlocks. The math only closes if U.S. grid buildouts, new nuclear projects, and TSMC's Arizona fabs all hit schedule at the same time. Historically none of those three have ever hit schedule at the same time. So the race inside the race is about how much useful work you can squeeze out of every watt you do manage to plug in. The lab whose next chip generation doubles that number wins a lot more than a benchmark — it wins breathing room in a grid that can't grow fast enough for any of them.

Anthropic's run-rate revenue is $30 billion, up from $9 billion at the end of 2025. At that growth rate, the $100 billion commitment looks rational. At any growth rate below that, it looks like a lifeboat tied to Amazon's balance sheet. Both things can be true. Every roadmap at every frontier lab is built on the assumption the power shows up on time. If it doesn't, the Claude slowdown you've been feeling at 5pm PST stops being a spike and starts being the ceiling.

Claude Design dropped Friday at claude.ai/design, and our team spent the weekend putting it through real work so you don't have to. Full first-look reaction here — the short version is below. Within hours of launch, Figma's secondary-market shares dropped 7 percent and Adobe shed 1.5 percent. That reaction wasn't an overreaction.

Here's the shape of it. You describe what you want, or drop in a screenshot, a Figma file, or a GitHub repo as the starting point. Claude actually asks clarifying questions before it generates anything, which is a small detail that changes the output quality a lot. Then it builds the first version as editable components with real layout logic, not a flattened image. You refine by leaving comments directly on any element (Figma-style), dragging custom sliders it generates on the fly, or editing pixels. When it's ready to ship, it exports to PDF, PowerPoint, Canva, or standalone HTML — or hands off straight to Claude Code for engineers to implement. And you know we love that last option the most.

Why this matters: the gate to building branded, production-quality work just moved. Branded decks, landing pages, apps, one-pagers, social assets, internal tools — the things that used to require a designer or a designer-adjacent skillset — are now available to every knowledge worker with a Claude login. The PM who needed mocks, the sales lead who needed a battle card, the founder who needed a pitch deck: all of them can now ship on brand without a handoff. The floor for what a non-designer can produce just rose sharply, and it rose across every asset type a modern company makes. Our honest take after a weekend of testing: it's early and raw, but you can see how quickly this becomes actually meaningful. The rough edges today are the kind that get sanded down in a release cycle or two.

Here's how to actually get value out of it on day one:

1. Feed it your brand. Drop in Figma files, your GitHub repo, or existing screenshots during setup. Claude respects fed-in brand systems (colors, type, component libraries), and the difference between generic output and on-brand output is almost entirely upstream of the prompt. Skip this step and you'll ship something that looks like every other AI slide deck on LinkedIn.

2. Parallelize with multi-agent. Claude Design lets you run several design threads at the same time. Treat it like a small in-house team: one thread for the investor deck, one for the landing page, one for the social assets. You get a design org out of a chat window.

3. Comment, don't re-prompt. The inline comments feature is the real unlock. Click an element, drop a comment ("tighter spacing between these two cards, bolder headline"), and Claude revises just that piece. Treat it like redlining a mock with a designer. Re-prompting the whole thing from scratch is how you lose the context that made the first pass land.

4. Use the Claude Code handoff. When something's ready to ship as code, the "package for Claude Code" button exports the design system, component specs, and content structure in a format Claude Code can build against directly. For teams already running Claude Code in their dev workflow, this is a design-to-deploy pipeline on one model family.

Try it: Default ON for Claude Pro, Max, and Team — go to claude.ai/design and start. Enterprise admins have to toggle it on via RBAC in the console (ships with a ~$20/user starting credit). Webapp only for now, no API yet. Still an Anthropic Labs product, not GA.

Claude Opus 4.7 shipped — Better at long-running agentic work, sharper vision, and it self-verifies outputs before responding. Same pricing as 4.6 ($5 / $25 per million tokens). The thing we keep coming back to: Anthropic openly admitted 4.7 still trails the unreleased Mythos model on cyber benchmarks. That's Anthropic telling you the flagship isn't the flagship yet. Nobody else in the industry is communicating that honestly right now. Our take: it's mid. Worth upgrading to, not worth throwing a party over. Just watch your token usage — it consumes more than 4.6. Full announcement

Codex went full computer use — here's the beginners guide — OpenAI's coding agent can now see, click, and type across every app on your Mac in parallel with your own work. Riley Brown walked through the full "Codex super-app" end to end: install, projects, custom skills, automations, then actually building an iOS app and a landing page using it. Best single walkthrough for anyone who wants to try it this week. Watch the guide

Claude is going after the Bloomberg terminal — Cowork's new Live Artifacts tab lets Claude build dashboards and trackers connected to your Google Sheets, Salesforce, and SQL data. Saved, versioned, and auto-refreshing whenever you open them. Build once, always up to date. The operational reporting layer that sat between your tools and your team is quietly being replaced by a chat window. See the feature

Vercel got hacked through a Roblox cheat search — Vercel confirmed a breach this week that started with an employee at a third-party tool (Context.ai) searching for Roblox game exploits and infecting their laptop with Lumma Stealer. That got the attacker into a Google Workspace OAuth app, which got them into Vercel's internal systems, which got them into some customers' non-sensitive environment variables. Your supply chain is only as strong as the weakest vendor's weakest employee's side hobby. CyberScoop breakdown

Scan the dark web for your own leaked data — Serus lets you search 100+ billion dark web records for leaked passwords, IP addresses, phone numbers, and emails tied to you. Takes seconds, free to try. Good tool to run on yourself and every exec on your team before a pen test does it for you. Check your exposure

Claude Code, but for Arduino and Raspberry Pi — A new tool called Blueprint takes a plain-English description of what you want to build and generates the wiring diagrams, bills of materials, and step-by-step assembly guides. The "AI for everything except hardware" gap just got smaller. Free. See it in action

Yann LeCun might have been right the whole time — A new paper (LeWM) just solved the "representation collapse" problem that had kept JEPA — LeCun's alternative to generative LLMs — stuck for years. The model has 15 million parameters, trains on a single GPU in hours, plans 48x faster than massive foundation world models, and intrinsically understands physics. If this holds up, the "bigger is better" thesis that justifies those $100 billion compute deals above gets a real counter-argument for the first time. Read the paper