Google DeepMind (Genie 4), Oracle and OpenAI — interactive worlds and silicon à la carte

Google DeepMind launches Genie 4, a real-time world model that generates native audio and lets users steer the scene with text. Oracle announces multi-gigawatt Nvidia Vera Rubin clusters on OCI and will be the first hyperscaler outside Google to offer TPUs starting in 2027. The Information and Bloomberg report that OpenAI plans to buy Google TPUs to diversify beyond Nvidia.

The same day you can step into a simulated world that speaks and responds, compute providers open the silicon menu and labs start buying their competitor’s hardware.

🌐 Google DeepMind Genie 4: interactive world with native audio

On 27 October DeepMind announced Genie 4, its new world model for real-time interactive simulation. The model generates native audio alongside the scene: ambient sound, music and interactive dialogue. You can steer the simulation with natural language while it runs: change the weather, move the camera or introduce characters. It is built on Gemini Omni for visual realism and uses other Gemini models to maintain coherence and continuity.

Availability starts as a research preview for a limited group of trusted testers. It will reach Google AI Ultra subscribers in Project Genie and early-access partners on the Gemini API and AI Studio soon. Scenes stay coherent for roughly two minutes, double the previous generation. DeepMind presents it as a step toward safe environments for training agents and robots.

Native audioambient, music, dialogue

~2 minscene coherence

Previewtrusted testers first

☁️ Oracle: multi-GW Vera Rubin and Google TPUs on OCI

At Oracle AI World on 27 October, Oracle and Nvidia announced general availability of Vera Rubin NVL72 systems and Vera CPUs on OCI, with plans to expand to multi-gigawatt scale. Oracle will be the first hyperscaler outside Google to offer TPUs, starting in 2027 and ramping to gigawatts over multiple years. It will also add Cerebras and SambaNova inference systems on the same timeline.

The strategy is explicit: no single silicon fits every workload. Customers will be able to choose architectures for training, inference and agent loads without leaving the cloud. Oracle also announced integration of ChatGPT and Codex into its enterprise applications. Deal values have not been disclosed.

Multi-GWVera Rubin on OCI

2027Google TPUs on OCI

First outsideGoogle with TPUs

⚡ OpenAI plans to buy Google TPUs

The Information reported on 27 October, and Bloomberg and Reuters confirmed, that OpenAI plans to buy Google TPU chips to expand capacity and diversify compute sources. OpenAI has already been renting TPUs through Google Cloud since 2025. Final terms and volumes have not been made public; neither OpenAI nor Google has confirmed the deal.

The most concrete signal is operational: OpenAI has posted a job for an engineer to lead optimization of its frontier models on TPUs. The company remains one of the largest buyers of Nvidia GPUs and is developing its own Jalapeño silicon with Broadcom. Buying TPUs from a direct competitor in consumer products is the logical consequence of scarcity and the cost of depending on a single supplier.

Reportedno official confirmation

Since 2025already rents TPUs

Diversificationbeyond Nvidia

The connecting thread

Genie 4 shortens the distance between generating a clip and being able to enter, speak and rewrite a world in real time. That accelerates agent and robot training in simulation. At the same time, Oracle turns its cloud into a silicon marketplace (Nvidia, Google, Cerebras, SambaNova) and OpenAI, Nvidia’s largest historical customer, prepares to buy Google’s hardware. The bottleneck stops being only “get more GPUs” and becomes “which architecture fits each workload and at what price”.

Mi opinión: Genie 4 is the most interesting move of the day for anyone building agents or robots. A world that generates native audio and can be steered with text while it runs is a far more useful training environment than a video. The roughly two-minute limit and access only for testers remind us it is still research, not product. The infrastructure side is drier but more expensive: Oracle offering Google TPUs from 2027 and OpenAI planning to buy them means Nvidia’s de-facto monopoly on frontier compute is cracking at the edges. The problem is not that Nvidia disappears; it is that multi-gigawatt contracts and multi-year commitments are being signed now, while price per token and cost per task are still falling. Anyone not measuring cost per completed task and without the option to move loads across architectures will pay the difference when the contracts harden. Simulation improves, the chip menu opens and capital concentrates. All three push in the same direction: less dependence on a single supplier and more pressure for the model to justify the hardware that sustains it.

📌 Artículos relacionados

→ 🛡️ Google (59), GitHub (>40%) and Microsoft ($10B) — cyber defense, agents and the real money → Google Cloud (GPT-6), Anthropic (Duet) and Microsoft (83%) — agents and models in production → Anthropic ($3.4B), Nvidia ($4.8B) and Apple — energy, risk and agents

✉️ Newsletter

Enjoying the content?

Subscribe to the newsletter and get each article in your inbox.

🤖 Este artículo está escrito con asistencia de IA y editado por un humano.