Etched leaves stealth with a working chip, $800M raised, $1B in orders

Good morning 👋 A chip startup just did what a hundred pitch decks couldn't: shipped working silicon, not a roadmap.

In today's issue:

• 🔭 Etched leaves stealth with a working chip, $800M raised, $1B in orders
• 🧠 Nano Banana 2 Lite, Gemini Omni Flash, and Claude's new science bench
• 🔬 Models that bluff, misread tables, and forget their own skills between sessions
• 🛠️ Instrument your agents before they fail quietly in prod
• 🔗 NVIDIA's software-only speedup, and five more worth a sentence

🔭 THE ONE THING

⚡ Etched leaves stealth with a working chip, $800M raised, $1B in orders

Etched came out of stealth today with the one thing most AI chip startups can't show you: first-pass (A0) silicon that actually works, built on TSMC's N4P process. The company has quietly raised $800M total, including a $500M round last December at a $5B valuation, from a backer list that reads like a prop-trading-desk reunion (Jane Street, Two Sigma, Hudson River Trading) plus a TSMC-linked fund, with Karpathy, Hinton, and Fei-Fei Li in as angels. Customer contracts are north of $1B, and rack-scale systems running DeepSeek, Qwen, Mamba, and Llama ship this summer. Worth separating what's confirmed from what isn't: the silicon success and the funding are independently verified, but that $1B order book has no named customer behind it, it's Etched's number alone. Still, a chip that works and a $5B check from people who trade for a living is more than most Nvidia challengers ever get to.

🧠 MODELS & RELEASES

• 🍌 Google DeepMind shipped Nano Banana 2 Lite and Gemini Omni Flash, live now in the Gemini API and AI Studio. Images run $0.034 apiece at 4-second latency, video's a dime a second for up to 10-second clips, and the new Interactions API chains three edits into one session. That's the product bet: iteration over one-shot generation. Google blog
• 🔬 Anthropic opened Claude Science in beta, a workbench wiring 60-plus connectors (UniProt, PDB, ChEMBL, NVIDIA's BioNeMo stack) into Claude, with sensitive datasets processed locally instead of uploaded. It's beta, not GA, gated to Pro, Max, Team, and Enterprise on Mac and Linux, with up to $30K in credits on offer for AI for Science applicants before the July 15 deadline. Anthropic

🔬 RESEARCH HIGHLIGHTS

• Metacognitive RL — LLMs are bad at knowing what they don't know: confident hallucinations, no sense of their own knowledge boundary. This paper trains models with RL on metacognitive feedback and gets stated confidence to actually track correctness. Builder read: closer to a model you can trust to abstain instead of bluff. arXiv
• Table reading errors — models parse table structure fine, then still cite the wrong cell. The error isn't comprehension, it's reference. Matters if you're shipping anything that answers questions over spreadsheets or financial tables: right-looking answer, wrong number. The paper says it's fixable, not fundamental. arXiv
• Skill composition for agents — instead of re-deriving "set up a sandbox" or "run the test suite" from scratch every session, agents assemble reusable skill packages and compose them for new tasks. Same idea Anthropic ships as Skills, now with a paper behind it. arXiv

🛠️ TRY THIS

Instrument the agent before it fails quietly in prod

Etched's chip story is a bet that agents get cheap enough to run everywhere. Cheap agents in prod means more of them silently looping, misfiring on tool calls, or burning turns doing nothing useful. AWS just published the actual playbook for catching that with Bedrock AgentCore's built-in traces and metrics.

1. Turn on AgentCore Observability for one agent already live, not a demo.

2. Pull the trace on its worst run from this week and find the exact step it looped or called the wrong tool.

3. Set a metric alert on that specific failure (tool-call count, turn count, whatever broke).

4. Re-run the same task and check the trace actually looks different.

Prompt: Given this agent trace [paste], find the exact step where it looped or invoked the wrong tool, and propose one guardrail (prompt edit, tool schema fix, or turn limit) that would have stopped it there.

Worth a look: Amazon Bedrock AgentCore Observability, if you're running agents on Bedrock and flying blind on failures.

🔗 QUICK LINKS

• NVIDIA credits most of its recent per-token cost drop to software, not new silicon: firmware and kernel tuning are still squeezing more out of the same Blackwell chips. Read
• OpenAI's new Signals data shows ChatGPT growth is now heaviest outside English-speaking markets. Read
• OpenAI also traced an 18-year-old infrastructure bug using large-scale core dump analysis, a debugging method more interesting than the bug itself. Read
• AWS laid out its internal review process for clearing frontier models before Bedrock ships them. Read
• IBM Research's new ScarfBench tests agents on real Java framework migrations, and the best one hit under 10% behavioral success (one agent even lied about its own build results). Read
• Hugging Face now shows community eval scores directly on model pages, pulled from the EvalEval Coalition's shared schema. Read

See you tomorrow.

Pradeep Perugu