FR

Market

Oracle opens its cloud to Google TPUs, Meta plans $95B of stock for AI: the October 28 roundup

Oracle will offer Google TPUs in 2027, Meta plans $95B of stock sales for AI, Google ships a CLI for coding agents and Liquid AI releases LFM3.

By ·

This roundup covers Monday evening, October 26, and Tuesday, October 27, 2026. On the infrastructure side, Oracle opened its AI World conference with chips from three suppliers, and Meta detailed how it plans to pay for its data centers. On the tools side, Google launched a command line built for coding agents, Codex, Claude Code and OpenCode shipped new versions, and Liquid AI released a family of small open-weight models. Unless stated otherwise, figures come from official announcements.

Oracle will offer Google TPUs in its cloud

Oracle and Google announced on October 27 that TPUs, the AI chips designed by Google, will be available on Oracle Cloud Infrastructure (OCI). It is the first time a major cloud other than Google Cloud offers them (Oracle press release). Oracle provides the power, cooling and networking, and the offer opens during calendar year 2027, ramping over several years. The release cites the eighth TPU generation, with one chip for training (TPU 8t) and one for inference (TPU 8i). According to Google, TPU 8t delivers nearly 3 times the compute per pod of the previous generation.

The same day, NVIDIA and Oracle announced general availability on OCI of Vera Rubin NVL72 systems (72 GPUs linked in one rack) and Vera CPUs, with a goal of expanding to "multi-gigawatt scale" (NVIDIA press release). Oracle also plans smaller HGX Rubin NVL8 servers: 8 Rubin GPUs, two 64-core Intel Xeon 6 CPUs and 4 TB of memory. The value of the deal was not disclosed. Oracle also plans to offer inference chips from Cerebras and SambaNova, according to Reuters.

Oracle and OpenAI: Codex inside Oracle's tools

OpenAI also expanded its partnership with Oracle (OpenAI announcement). Oracle applications and data are set to reach ChatGPT through an upcoming Oracle AI Workspace. Oracle App Composer will be used to build apps with Codex and other AI tools, then deploy them on OCI. In the US, Oracle customers will be able to buy the OpenAI API, Codex and ChatGPT Work on Oracle Marketplace with their Oracle credits. OpenAI says Oracle has 130,000 active ChatGPT users and more than 95,000 active Codex users internally.

Meta plans to raise $95 billion in stock

On October 26, Meta presented an equity financing plan of about $95 billion for its AI investments (Meta press release):

The next day, NVIDIA said it will invest $3 billion in the offering, and that Meta will deploy more Vera Rubin systems and Vera CPUs (NVIDIA press release). According to CNBC, Meta shares were down 1.6% on the morning of the announcement.

Google launches a Gemini API CLI for coding agents

On October 27, Google released gemini-api in beta, a zero-dependency, open-source command line designed to be called by coding agents, scripts and CI pipelines (Gemini API release notes). It does not replace Gemini CLI, the interactive agent: it is a stateless client that an agent such as Claude Code, Codex or Cursor can run to generate text, images, speech, music or video, analyze a file or launch a Google-managed agent such as Deep Research (CLI guide).

The design choices target agents: media are written to disk and only their path is returned, so they do not fill the context window; --dry-run previews the request without sending it; exit codes are typed (3 for an authentication problem, for example). Google warns that breaking changes can land between minor versions and recommends pinning the version in CI.

Codex, Claude Code and OpenCode: the new versions

Liquid AI releases LFM3, 2-billion-parameter models

On October 27, Liquid AI launched LFM3, a family of six open-weight models of about 2 billion parameters (text, vision, audio, omni and two decision models), all built from the same base model pre-trained on 22 trillion tokens (Liquid AI announcement). The text model, LFM3-1.7B, has a 262,144-token context window and fits in about 1 GB at 4-bit precision. It is released under Liquid's own LFM Open License v1.0 (Hugging Face model card).

Liquid AI reports 43.4% on SWE-bench Verified with the OpenCode harness, against 51.2% for Qwen3.5-9B, and acknowledges that larger models still lead in agentic coding. These are vendor figures. LFM3-1.7B joins our model comparison, without a score: our comparison does not track this benchmark yet.

Sources

Spotted an error in this article? Report it · Our corrections policy

Rekha icon: a house drawn in a single stroke

From the same author Self-promotion

An Eulerian path, drawn with your finger

Rekha, the iPhone game by OrigoTech's author: connect every dot in a single stroke without tracing the same link twice, across 500 drawings that are all solvable.

View on the App Store → Why it is a graph problem

The OrigoTech newsletter

A sourced weekly digest of AI tool news.

Read next