EU puts €10 billion toward seven AI supercomputers as OpenAI opens the door to academia

The European Commission has published a tender to build seven giant compute clusters — what it calls “AI factories” — with an expected budget of €30 billion. Of that, €10 billion will come directly from EU funds and member-state budgets. Submissions close on 12 November; winners will be announced in July 2027, and the contract requires full operation within 18 months of signing. It is the first time the bloc has committed large-scale capital to hardware and silicon rather than policy papers alone.

At the same time, OpenAI has launched ChatGPT for Academic Researchers, a programme that will grant up to one hundred thousand researchers access through 2027. The first ten thousand will receive invitations this summer. The package includes GPT-5.6 Sol Pro, ChatGPT Work and Codex with an expanded context window, enlarged windows and additional credits for Deep Research. The platform ships with more than 75 “scientific skills” and is connected to paper repositories, genomic datasets and computational notebooks. Registration is restricted to researchers at approved universities, and each participant may add up to four colleagues from their institution.

Thinking Machines Lab, the startup founded by Mira Murati, has released Inkling-Small, a multimodal MoE model with 276 billion total parameters and 12 billion active, plus a one-million-token context window. Through distillation from the full model and reinforcement learning, the smaller version outperformed the original on coding and logic: 80 % on SWE-bench Verified and 31.6 % on Humanity’s Last Exam. Weights are available on Hugging Face; inference and fine-tuning run through the Tinker platform. Pricing starts at 58 cents per 64-thousand tokens and rises to $1.16 per 256-thousand (roughly 2.1 to 4.2 shekels).

Google’s Embodied Reasoning 2 does not generate text — it plans movement. The architecture separates logic from motor control: the model analyses video in real time, estimates stage status and corrects trajectories without restarting the process, while VLA models execute the physical actions. ER 2 also supports multi-agent workflows, letting devices of different types exchange semantic context to carry out multi-step tasks in a shared space. The model is accessible via the Gemini API and Google AI Studio, with starter examples for developers building physical AI.

Liquid AI has released two bidirectional encoders, LFM2.5-Encoder in 230-million and 350-million-parameter sizes, built on the LFM2 architecture and optimised to run on CPUs with an 8,000-token window. They are intended for classification, routing and token-level tasks. The larger version placed fourth out of 14 models across 17 benchmarks. The company says that on a full 8,192-token pass, the smaller model completes a forward pass on CPU in 28 seconds — 3.7 times faster than ModernBERT-base. Code and weights are on GitHub and Hugging Face.