How High-Bandwidth Memory Works: Stacked DRAM, Wide Interfaces and Why It Is Hard to Make
High-bandwidth memory gets its speed from width, not clock rate: stacks of DRAM dies, joined by vertical connections, sit beside the processor and are read over a 1,024- to 2,048-bit interface. This post sets out how that works, what each generation changed and why more layers make it harder to build.
- JEDEC HBM4 interface2,048-bit
- Max stack (JEDEC)16-high, 64 GB
- Maker generations6
Latest
From official releases, filings and dataGoogle makes the Gemini API Computer Use tool generally available in the Interactions and GenerateContent APIs Gemini API changelog
How Computer-Use Agents See and Click: Claude, OpenAI and Gemini APIs ComparedGeminiGoogle
Epoch AI estimates new AI computing capacity is doubling every 9 months, about 2.5x a year since 2024 Epoch AI Data Insights
AI Accelerators With the Most HBM per Chip (October 2026)Epoch AI
Mistral Large 4 reaches Hugging Face: ~1T parameters, 49B active, 1M-token context, Mistral Ecosystem license Hugging Face Blog
Largest Open-Weight Models by Total Parameters (October 2026)Mistral AIHugging Face
NVIDIA and Oracle: Vera Rubin NVL72 reaches general availability on OCI; multi-gigawatt expansion is planned NVIDIA Newsroom
NVIDIA Vera Rubin by the Numbers: Specs, Claims and First ResultsOracleNVIDIA Rubin
OpenAI releases GPT-6.1 Terra in the API: a model for visual reasoning, computer use and writing OpenAI API changelog
Cheapest LLM APIs per 1M Output Tokens (October 2026)GPTOpenAI
Anthropic donates the Agent Skills standard to the Agentic AI Foundation; 70+ agent platforms have integrated it Anthropic
How the Model Context Protocol Works (2026-07-28 Revision)Agentic AI FoundationAnthropic
Recent posts
The whole stack →How Speech-to-Speech Voice Agents Work: Chained Pipelines, Native Audio and Full Duplex
A voice agent turns speech into action in one of three ways: a chain of speech recognition, a text agent and speech synthesis; one native audio model that listens, reasons and speaks; or a full-duplex voice model that keeps talking while a separate backend works. This explainer follows each design and how its timing is judged.
Highest-Elo Text-to-Video Models in the Artificial Analysis Arena (October 2026)
The 25 text-to-video models with the highest Elo on Artificial Analysis's AA-Video-T2V v2.1 arena, from blind human votes, with intervals, vote counts, API prices and open-weights status. Wan 3.0 leads at 1174, but the top ten sit within 46 points and their intervals overlap; a 62-point gap separates tenth from eleventh.
Anthropic's Compute Commitments, as Filed in Its S-1
Anthropic's preliminary S-1 describes compute arrangements with eight counterparties, from a $44.9 billion minimum commitment with Microsoft to about $161.2 billion of equipment-lease rent tied to Broadcom. Here is each one as the filing states it, with its type of figure, its term and its cancellation terms. Nothing here is a total, a ranking or a valuation.
China's 140 Trillion Daily Tokens and Google's 3.2 Quadrillion a Month: What Each Token Figure Measures
In 2026 a national agency, two companies and a router service have each published a token count: 140 trillion a day in China, 3.2 quadrillion a month at Google. The figures use different units, periods and scopes, and several don't say what they count. This feature sets them side by side without ranking or adding them.
How Agent Skills Work: SKILL.md, Progressive Loading and How They Differ from MCP
Agent Skills package instructions, scripts and reference files into a folder that an agent loads in stages, so it can keep many skills on hand at a small cost in context. On October 22, 2026 Anthropic donated the open standard to the Agentic AI Foundation. This explainer covers the format, the loading steps, the contrast with MCP and the security caveats.
Largest Open-Weight Models by Total Parameters (October 2026)
Twelve open-weight language models with 295 billion or more total parameters, ranked by the size each lab states on its model card, with active parameters, context window and license beside it. Kimi K3 is largest at 2.8 trillion, and nine of the twelve activate fewer than 50 billion parameters per token.
How SynthID and C2PA Content Credentials Work
Two different techniques are used to mark AI-generated content. C2PA Content Credentials attach a cryptographically signed provenance record to a file, while SynthID hides a watermark inside the image, audio, video or text itself. This explainer follows each mechanism and sets them side by side.
L08 · Chips & MemoryFeatureNVIDIA Vera Rubin by the Numbers: Specs, Claims and First Results
NVIDIA’s Vera Rubin NVL72 rack lists 2.3 times the dense 4-bit compute and 2.4 times the memory bandwidth of the GB300 NVL72 it follows, and NVIDIA says it cuts the cost per million tokens to a tenth of a GB200 rack’s. Here are its specs as NVIDIA states them, what NVIDIA claims it will do for the industry’s token costs, the first MLPerf and cloud results, and who is deploying it.
AI Agent Identity and Trust Standards: Who's Writing Them, and Where They Stand
The 15 specs, drafts and guidance papers that try to prove who an AI agent is and what it may do, from Web Bot Auth and AIMS to signed A2A cards and ERC-8004, with status and date.