Meta to Sell Surplus AI Compute Resources as a Cloud Service

Meta is moving ahead with plans to sell the surplus capacity of its massive AI computing infrastructure to outside customers as a cloud service, according to multiple outlets including TechCrunch and Quartz.

Meta has built enormous GPU clusters for training and running its AI models, but the compute capacity it secured as a buffer against demand spikes has grown larger than originally anticipated, and the company aims to monetize that surplus. Reports say the approach mirrors SpaceX selling its excess launch capacity to satellite operators.

The move could reshape the competitive landscape of the AI infrastructure market. If Meta enters the cloud computing business in earnest, incumbents such as AWS, Google Cloud, and Azure will face new competitive pressure. Details of the service offerings and pricing have not been announced at this point.

Venice AI Raises $65 Million, Becomes a Unicorn

Venice AI, which offers a privacy-focused AI platform, has closed a $65 million Series A round and reached unicorn status (a valuation of $1 billion or more), TechCrunch reports.

With its privacy-first approach of not storing or tracking user data, Venice AI provides an alternative to major services like ChatGPT. Amid growing concern about data privacy, investors appear to have recognized the market potential of privacy-centric AI.

There is clearly a segment of users who hesitate to use major AI services over privacy concerns, and this funding round stands out as a symbolic milestone demonstrating the growth potential of this space.

Google's Gemini Spark Comes to the Mac

Gemini Spark, Google's agentic AI assistant tool, is now available for the Mac, according to TechCrunch.

Gemini Spark is not just a chatbot — it is an "agentic" assistant AI that understands user intent and autonomously executes tasks. With Mac support, AI assistance can now be integrated more naturally into the desktop environment.

Agentic AI is one of the key trends of 2026, and companies are racing to integrate it into the desktop. With Google establishing a foothold on the Mac, competition with Apple Intelligence and Microsoft Copilot is likely to heat up further.

SWE-rebench Updated: Claude Opus 4.8 at 56.5%, GLM-5.2 at 51.1%

The leaderboard for SWE-rebench, a benchmark for software engineering tasks, has been updated with a large number of new models added. The update was reported in Reddit's LocalLLaMA community.

The key results are as follows:

Claude Opus 4.8 xhigh**: 56.5%

GLM-5.2**: 51.1%

Gemini 3.5 Flash**: 49.5%

DeepSeek-V4 Pro**: 42.7%

DeepSeek-V4 Flash**: 38.4%

Qwen3.6-27B**: 36.5% (notable as a local model)

Looking specifically at local and self-hosted models, Qwen3.6-27B posts a high score for its size, and Qwen3.6-35B-A3B and Gemma 4 31B were also added as comparison points. It is clear that open source models are steadily getting stronger.

Hugging Face × Cerebras: Real-Time Voice AI with Gemma 4

Hugging Face and Cerebras Systems have announced a partnership to deliver real-time voice AI built on Google's Gemma 4 models.

Combining the ultra-fast inference of Cerebras' large wafer-scale engine (WSE) with Gemma 4 is said to make it possible to build voice-based conversational AI with dramatically reduced latency. In voice AI, response delay is critical to UX, which gives this hardware-driven approach particular significance.

The LLM "Groupthink" Problem Comes Under Scrutiny

MIT Technology Review reports that today's LLMs are falling into the trap of "groupthink."

Because so many models rely on similar training data and methods, diverse perspectives and emergent behaviors are being lost, producing a phenomenon in which different models return similar answers. The article also covers startups working to solve this problem, and ensuring model diversity is emerging as a key challenge for future AI research.

---