metir Blog

The Pentagon's Office of Strategic Capital offered Wolfspeed a conditional $1.5B loan with warrants for up to 7.5%. Why silicon carbide matters for AI power.

Broadcom, Oracle and SpaceX are reportedly seeking tens of billions in private credit for AI chips. How vendor financing and lease-back SPVs work, and the risks.

Nikkei says Chinese labs shipped 16 AI models in September 2026 as release gaps fell from 125 to 44 days. What it means for the AI pacing debate.

Claude Haiku 5.5 launched October 7 at $0.10 and $0.50 per million tokens. Pricing, benchmarks, effort controls and how it compares with Sonnet 5.5 and rivals.

A Cisco and Carnegie Mellon study found 8 of 13 AI models steered wealthier personas to pricier options, even when asked for the cheapest. What it shows.

OpenAI's GPT-6 rollout brings Intelligent UI, interactive charts, forms and mini apps, to every ChatGPT tier. Dates, models, and what generative UI changes.

Google's SynthID Detector website is now open to anyone for images, video and audio. What it detects, its limits, and how it compares with C2PA and classifiers.

OpenAI's first teen ChatGPT usage report says teens average under 15 minutes a day. What average, tail and self-reported metrics can and cannot show.

Nous Research confirmed a $90M Series B at a $1.5B valuation and launched Hermes for Businesses. How open-weight labs make money and why agents matter.

Musk said his companies will build and run Terafab, with at most a TSMC sublease. Why running a leading-edge fab is hard, and how Terafab costs compare.

Microsoft's $5,999 Surface RTX Spark dev box has 128GB unified memory for local models. We break down the math, rivals, and cloud vs local tradeoffs.

Musk says Grok Bot will use the best back end model per task, including Claude Opus 5.5, Midjourney and Suno. How model routing works and what it trades off.

Samsung Q3 2026 earnings guidance shows a record 107.4 trillion won operating profit, the first above 100 trillion. What the memory and HBM4 numbers do and do not say.

Finland ordered Google's local subsidiary to stop preparatory work at two data center sites. How the EIA process works and why permitting friction is growing.

Mistral Large 4 preview is a 1T-parameter MoE with 49B active. What that means for self-hosting cost, rival open models, EU sovereignty and licence terms.

Micron's Taoyuan union won a 99% strike mandate and wants 15% of operating profit shared. How Taiwan strike law, Korean rivals and HBM supply fit together.

Lambda is reportedly raising up to $4B at a $14.5B pre-money valuation before a 2027 IPO. Backlog hit $50B, but Anthropic is $35B of it. What the numbers mean.

OpenAI released 722 math manuscripts from an unreleased model, with Lean proofs for 162 papers. What Lean does and does not check, and the referee bottleneck.

Anthropic put Claude in Google Docs, Sheets and Slides as a sidebar beta on paid plans. How it installs, what it can access, and what admins should check.

Rapidus signed 17 design partners to win customers for its 2nm foundry. What the move means, how its funding and timeline compare with TSMC, Samsung and Intel.

Common Sense Media rated ChatGPT for Teens an Unacceptable Risk on Oct 7, 2026. What its Institute tested, how ratings work, and OpenAI's dispute.

Waymo upsized its debut private loan from over $3 billion to $5 billion. What the debt means for robotaxi fleet financing, unit economics and AI-era capital.

Meta and Sierra's Personal Agent Protocol proposes an OAuth-based way for AI agents to authenticate with businesses. How it differs from MCP, A2A and ACP.

Google DeepMind's EmbeddingGemma 2 puts text, code, images, video and audio in one vector space under Apache 2.0. What that changes for search and RAG.

Lisa Su says AMD will substantially increase chip supply in 2027 as 2nm orders rise at TSMC. What the capacity numbers show and what to watch on Oct 15.