Oracle Cloud will offer Google TPUs from 2027, the first hyperscaler outside Google to do so
Google is putting its TPUs in Oracle Cloud Infrastructure from 2027, as Oracle adds Nvidia Vera Rubin, Cerebras and OpenAI deals at AI World. Supabase can now run agents inside your project, DeepMind's Genie 4 generates steerable worlds with sound, and Apple opens Keynote, Pages and Numbers to agents over MCP.
- Google will offer its TPUs on Oracle Cloud Infrastructure from 2027, making OCI the first hyperscaler outside Google to rent them out; Nvidia Vera Rubin NVL72 racks and Vera CPUs are generally available on OCI now.
- Supabase Compute, in private alpha, runs agents, background jobs and APIs inside a Supabase project for minutes or hours, beyond Edge Functions’ 400-second limit; it’s free to try and only in us-west-2.
- Google DeepMind’s Genie 4 world model generates audio with its worlds and can be steered in real time by text; it’s a research preview for trusted testers, with AI Ultra and Gemini API access to follow.
- Apple is adding MCP support to Keynote, Pages and Numbers on macOS 27 in November, so outside agents can read and edit documents, plus assistants built on Apple models and OpenAI.
- Wiz says it has seen no successful end-to-end PHP2Shell exploitation in its own telemetry, and exposed vulnerable machines fell about 67% in days; CISA still added both bugs to its exploited list on 26 October.
Oracle Cloud will rent out Google’s TPUs
Google and Oracle will bring Google TPUs to Oracle Cloud Infrastructure, with deployment starting in 2027. Google calls OCI the first hyperscaler outside Google to offer TPUs. The post doesn’t name the TPU generation, regions or prices. It points to TorchTPU, Google’s now open-source PyTorch backend, and support for vLLM and SGLang.
The deal led Oracle’s AI World announcements. Nvidia Vera Rubin NVL72 and Vera CPUs are generally available on OCI, with a “multi-gigawatt” expansion planned. Nvidia says OpenAI trained and served GPT-6 Astra on more than 100,000 Blackwell GPUs on OCI in Abilene. Cerebras inference is due on OCI in 2027, Converge Digest reports, and OpenAI said it is bringing Oracle’s apps to ChatGPT.
Why it matters
Teams that write for TPUs, including through TorchTPU and vLLM, get a second cloud to run them on, next to Oracle Database without cross-cloud transfers.
Google turns its chip into a product others sell
TPUs have been a reason to choose Google Cloud. Renting them through a rival suggests Google values TPU volume against Nvidia more than that lock-in. Oracle gets a hedge for customers who want to run on something other than Nvidia. What would change the call: TPU prices on OCI, still unpublished.
Supabase puts agent servers next to the database
Supabase Compute, in private alpha since 27 October, runs agents, background jobs and APIs inside a Supabase project. Services use Node, Deno or any Dockerfile on full Linux, in sizes of 1 vCPU with 2 GB or 2 vCPUs with 4 GB, and can run for minutes or hours. Edge Functions stop at 400 seconds. It’s free to try, meant for evaluation rather than production, limited to projects in us-west-2, and joined by waitlist. There’s no autoscaling, GPU or persistent disk, services can restart at any time, and every service gets all of the project’s secrets.
Why it matters
A data agent that makes many small queries gains most. Dreambase, a customer, reports a median Postgres query of 2.9 ms from Compute against 65 ms from another region.
The database host wants the agent runtime too
Supabase already hosts many AI builders’ databases; now it wants the long-running processes beside them, work teams now send to Fly.io, Railway or Render. Agent sandboxes aren’t part of the alpha, so it isn’t an E2B or Daytona rival yet; our agent sandbox comparison covers those, and our database-per-agent guide covers where agent data lives.
Genie 4 makes worlds with sound you can steer
Google DeepMind released Genie 4 on 27 October as a research preview for trusted testers. It is DeepMind’s first real-time world model to generate audio natively: music, ambient noise and dialogue. Users can change the weather, move the camera or add people and objects by typing, and Gemini 3.8 Flash runs in the loop to break goals into steps and check frames. Google AI Ultra subscribers in Project Genie get access first, then early-access partners in the Gemini API and AI Studio. Google published no resolution, frame rate, session length or price.
Why it matters
Game prototyping, previsualisation and robot simulation are the stated uses; DeepMind showed Apptronik’s Apollo humanoid playing checkers in a Genie world.
Still a demo until the API opens
Google admits object permanence drifts over long runs and after several edits. Until there’s a price and an API, compare it to rivals like World Labs on demos only.
Apple opens Keynote, Pages and Numbers to agents
Apple’s Creator Studio update adds MCP support to Keynote, Pages and Numbers, letting external agents read content, make edits and run actions on macOS 27. The apps and Logic Pro also get assistants in beta, in US English, built on Apple models and OpenAI. The update is free to subscribers in November; Creator Studio costs $12.99 a month or $129 a year. Apple also announced a 14-inch MacBook Pro with M6 from $1,999 with 16 GB or 32 GB, shipping 6 November; Apple claims up to 25% faster on-device LLM prompt processing than M5.
Why it matters
From November, on macOS 27, Claude, ChatGPT or a local agent will be able to build a deck or fill a spreadsheet in Apple’s apps through a standard interface instead of UI scripting.
Apple ships its own assistant and an open door
Apple is shipping both a built-in assistant, with OpenAI behind it, and a standard way for any agent to drive its apps. For people who already work through Claude or ChatGPT, the MCP door is likely the bigger change.
Quick hits
- Replit Cloud Agents now work on existing company repositories: anyone on a team describes a change, and the agent tests it in a cloud sandbox and opens a GitHub pull request or GitLab merge request. Available to Pro and Enterprise customers, with Slack for all.
- Wiz has seen no successful end-to-end PHP2Shell exploitation and says fewer than 53% of initially affected organisations were still vulnerable a week after disclosure. CISA added both CVEs to its exploited list on 26 October, which Wiz thinks a researcher’s compromised honeypot likely prompted; patch anyway.
- AMD will invest $1 billion in Meta’s equity offering and Meta will deploy a custom AMD Instinct GPU, a day after Nvidia committed $3 billion.
- Liquid AI released LFM3, six open-weight models of about 2B parameters for text, vision, audio, omni and decisions, all from one 1.7B base; the text model has 256K context and fits in about 1 GB at 4-bit.
- Gemini Robotics ER 2 is in public preview on Google Cloud through Model Garden, as a planner that calls robots’ control APIs as tools.
- Opera proposed WebAI, a web standard for running open-weight models such as Gemma locally in the browser, in early, opt-in preview in Opera One.
- Salesforce’s Agent Fabric added an agent kill switch and Model Wallet budget limits, generally available now; Agent Evals arrive in beta in November.