🍭 The Last Lab Without A Chip 💽

Good morning. Anthropic looked at its compute bill, looked at Nvidia, and quietly posted a job ad for someone who knows how to make a chip.

Let’s dive in 👇

🍭 What’s Cookin’:

• Anthropic decided to make its own chips now

• The UK's AI watchdog caught Mythos 5 doing things it shouldn't

• Cloudflare gave AI agents a wallet. And a name. What can go wrong?

Anthropic
💽 Is Building Its Own Chip

The Bite:

Anthropic confirmed that it is assembling an in-house team to design custom AI chips. The company says it will co-design hardware and models so Claude runs faster and more cheaply at scale.

Anthropic already buys compute from Amazon, Google, Nvidia, and AMD, and calls the silicon team part of a multi-chip approach.

Anthropic confidentially submitted a draft Form S-1 (IPO) to the Securities and Exchange Commission on June 1.

Snacks:

• The "custom silicon team" job listing pays $320,000 to $485,000.

• Industry sources cited by Reuters put the cost of designing an advanced AI chip near $500 million.

• OpenAI's custom chip program lead, Clive Chan, left for Anthropic in June.

• The Information reported last month that Anthropic was considering Samsung as a manufacturing partner.

• Anthropic plans to run Google-Broadcom TPUs at a $15 billion Nexus campus in Hubbard, Texas, with Google backstopping its obligations.

• Anthropic projected $10.9 billion in second-quarter revenue and $559 million in operating profit, its first.

Why it Bites:

A chip team does nothing for Claude this year. Or next.

The first Anthropic silicon, if it ever tapes out, arrives in a world where today's models have already been retired twice over.

Now, notice what was actually announced.
A hiring plan. Not a chip.
In the same breath Anthropic reaffirmed it will keep current partnerships.

The vertical integration is a slide, not a supply chain.
Which is roughly the point.

Compute cost fell from 71 cents to 56 cents per revenue dollar between Q1 and Q2. A silicon program is how you convince public investors it'll keep falling.

With a roadshow expected within weeks, every cost curve becomes something you show rather than something you assert.

OpenAI is standing in the same queue with the same homework.

The chips might never arrive. The chart they justify arrives in October.

Steal This Prompt
📓 Make Your Own Gritty Charcoal Masterpiece

This prompt transforms portraits, products, pets, or pretty much anything into dramatic charcoal artwork that looks ripped straight out of a sketchbook instead of an AI model.

Use it to:

• Turn selfies into museum-worthy charcoal portraits

• Give product shots a premium editorial art style

• Create gritty album covers, posters, or profile pictures with serious attitude

Workflow:

1. Hit this link: Gritty Charcoal Sketch

2. Paste into your AI model

3. Replace the #s with your image, subject, or idea

4. Watch it cook into a dramatic charcoal illustration worthy of an art exhibit

   Try the prompt on Snack Prompt

_ToolBox™
🧰 5 BRAND NEW AI LAUNCHES

🚦 Coldtea.ai

An agentic IDE where the coding agents ship and separate QA and monitoring agents catch what they broke, before your users do.

🧑‍💼 Soloop

Gives a solo founder an AI CEO, CTO, and CMO that have to ask permission before acting, which is more than most cofounders offer.

🗺️ Rindler

Maps each site in advance instead of letting a browser agent guess every run, so your weekly "log in and pull the numbers" chore survives the next redesign.

🔍 Progress AI Observability

Traces every agent run and flags the confidently ungrounded answers that your normal monitoring logs as a clean 200.

🌍 Nitro 4.0

Your agent orders and pays a human translator per request with no account or API key, and gets publication-ready copy back in hours across 80+ languages.

_{Everything Else}
🧠 You Need to Know

💳 Cloudflare Gives AI Agents Wallets And IDs
→ Cloudflare announced Wallets and Cloudflare Pay, letting account owners hold stablecoins and issue spending-capped virtual wallets to individual agents.

🚨 UK Watchdog Logs 19 Unsanctioned AI Attack Actions
→ The AI Security Institute documented 19 unsanctioned actions by Claude Mythos 5 and GPT-5.6 Sol against real people and organizations across 10 of 122 test runs.

🧠 Anthropic Starts Designing Its Own AI Chips
→ Anthropic is hiring a custom silicon team to co-design chips and models for Claude, alongside existing hardware deals with AWS, Google, Nvidia, and AMD.

🛡️ Mistral Open Sources A 3B Safety Classifier
→ Shieldstral shipped August 4 under Apache 2.0, judging text and images against plain-language moderation policies supplied at inference time.

🚗 Uber Commits $10B To Robotaxi Buildout
→ Uber will spend more than $10 billion on equity stakes in autonomous-driving partners and fleet support, targeting 120,000 vehicles across at least 15 cities.