Meta unveils MTIA 300 as Google releases Gemini Embedding 2

Key Takeaways

• Meta is deploying four custom AI chips (MTIA 300, 400, 450, 500) through 2027 to reduce reliance on Nvidia and AMD for AI workloads.
• Google AI released Gemini Embedding 2, a multimodal model processing text, images, video, and audio for improved RAG systems.
• Google Gemini app downloads exceeded 101 million in February 2026, surpassing ChatGPT, though ChatGPT maintains a lead with nearly 592 million Monthly Active Users.
• A new nonprofit, Radial, launched with at least $500 million to modernize scientific processes using AI.
• Canva introduced Magic Layers, allowing users to edit individual elements within AI-generated images.
• Wayfair integrated OpenAI models to improve product catalog accuracy and automate supplier support workflows.
• The Academic Senate is exploring AI tools for grading coursework, focusing on transparency and student privacy.
• A New York Times blind test found 54% of readers preferred AI-written content over human-written originals.
• Companies like Fish Audio and Resemble AI are advancing expressive text-to-speech technology with features like emotional control and zero-shot voice cloning.

Meta launches four custom AI chips to boost performance

Meta has released four custom-designed AI chips, with the first, MTIA 300, already deployed. These chips, including the upcoming MTIA 400, 450, and 500, are intended to improve AI tasks like content recommendations and generative AI. By creating its own silicon, Meta aims for greater control over its supply chain and to better manage costs associated with AI development. The company plans to release new chip generations approximately every six months to keep pace with rapidly evolving AI technology.

Meta expands AI chip line with four new MTIA generations

Meta is developing and deploying four new generations of its custom MTIA chips within the next two years to support AI workloads. These chips are designed for ranking, recommendations, and generative AI tasks, with a focus on inference. The company is using a portfolio approach, combining its custom silicon with chips from industry leaders. Meta's strategy prioritizes rapid development, an inference-first focus, and building on industry standards for easier integration into their data centers.

Meta unveils four new chips for AI and recommendations

Meta has developed four new computer chips called MTIA processors to power its AI and recommendation systems, aiming to reduce reliance on companies like Nvidia. The chips, developed with Broadcom and built on RISC-V architecture, are manufactured by TSMC. MTIA 300 is already in production, while MTIA 400, 450, and 500 are expected between early and late 2027. This rapid development cycle allows Meta to adapt quickly to changing AI needs and optimize performance for its specific workloads.

Meta's new AI chips challenge Nvidia and AMD

Meta Platforms is developing four new AI chips to power its metaverse ambitions and reduce its dependence on rivals like Nvidia and AMD. These custom silicon chips are designed to handle demanding AI tasks, including training large language models and powering recommendation systems. This move is part of Meta's long-term strategy to gain more control over its hardware infrastructure and optimize performance for its specific needs in the competitive AI landscape.

Meta deploys four new homegrown chips for AI tasks

Meta is preparing to deploy four new homegrown chips: MTIA 300, MTIA 400, MTIA 450, and MTIA 500. This effort aims to diversify its hardware sources, lower costs, and reduce reliance on external chipmakers amidst the competitive AI race. MTIA 300 is already in production for content ranking, while the others are slated for deployment through 2027. Meta is developing these chips in parallel to keep pace with the rapid evolution of AI workloads and ensure they remain state-of-the-art.

Google's Gemini Embedding 2 handles text, images, video, audio

Google AI has released Gemini Embedding 2, a new multimodal embedding model designed for AI developers. This model can process text, images, video, and audio, mapping them into a single vector space for easier cross-modal retrieval. It uses Matryoshka Representation Learning (MRL) to optimize for storage and compute efficiency, allowing for different levels of precision. Gemini Embedding 2 aims to improve Retrieval-Augmented Generation (RAG) systems by capturing richer semantic relationships between different data types.

Fish Audio S2 offers expressive text-to-speech with emotion control

Fish Audio has released S2-Pro, a new generation of expressive text-to-speech (TTS) technology. This model uses a Dual-AR framework for high-fidelity, low-latency speech synthesis and supports zero-shot voice cloning. S2-Pro allows for granular emotional control through in-context learning and inline tags, enabling dynamic emotional transitions within generated speech. The system is optimized for real-time applications, achieving sub-150ms latency.

Academic Senate discusses AI use in grading, student privacy

The Academic Senate is exploring the implementation of artificial intelligence tools for grading coursework, addressing faculty concerns about AI usage. A resolution was created to ensure transparent communication regarding faculty AI use and to mitigate potential risks, including student privacy. While some faculty see AI as a helpful tool, others worry about relying on technology for the human aspect of teaching. The Senate is gathering feedback to determine how to move forward with AI in the classroom.

New nonprofit Radial launches with $500 million for AI in science

A new nonprofit called Radial has launched with at least $500 million to modernize the scientific process for the AI era. Founded within the Astera Institute, Radial aims to build essential infrastructure and tools needed to fully realize the potential of AI in scientific research. The organization believes that updating current systems is crucial for unlocking the value of AI across various fields, including science and biotech.

Canva's Magic Layers makes AI images editable

Canva has introduced a new feature called Magic Layers that allows users to edit previously flat AI-generated images. This tool analyzes AI images and breaks them down into individual, editable elements like backgrounds and characters. Users can upload AI images created anywhere and transform them into editable Canva projects. Magic Layers aims to bridge the gap between static AI images and fully structured, editable designs, making AI image creation more practical for creators.

Gemini leads AI app downloads, ChatGPT leads active users

In February 2026, Google Gemini surpassed ChatGPT in app downloads with over 101 million, while ChatGPT had over 65 million. However, ChatGPT maintained a significant lead in Monthly Active Users (MAUs) with nearly 592 million, compared to Gemini's 111 million. This indicates that while Gemini is attracting new users, ChatGPT retains a much larger base of consistent, active users. Other AI apps like Microsoft Copilot, Grok, and Perplexity show considerably lower download and usage numbers.

Resemble AI offers realistic voice generation with advanced control

Resemble AI provides a professional platform for AI voice generation, focusing on realism and emotional control beyond basic text-to-speech. The platform features real-time API capabilities and a 'Fill' function for editing audio by typing. It supports cloning voices with high fidelity and allows granular control over emotions through inline tags and in-context learning. Resemble AI is designed for professionals like podcasters, marketers, and game developers who require scalable, high-quality audio.

NYT readers prefer AI writing in blind test

A blind test conducted by The New York Times revealed that 54% of readers preferred AI-written content over human-written content. The quiz involved comparing AI-generated versions of existing strong writing against the originals, with readers unaware of the author. This result challenges the notion that AI cannot be creative and suggests that AI writing is becoming indistinguishable from, and sometimes preferred over, human writing across various genres.

Wayfair uses OpenAI to improve catalog accuracy and support

Wayfair has integrated OpenAI models into its supplier support and catalog systems to enhance data accuracy and automate workflows for millions of products. The company focused on improving product attribute tagging, which is crucial for search and customer trust. By using a single, reusable AI architecture, Wayfair can now expand model coverage for new attributes at a much faster rate. This integration has led to significant improvements in SEO, clicks, and customer discovery of products.

Iowa businesses learn AI for productivity boost

Business owners in Iowa received an introduction to artificial intelligence on March 11, focusing on how the technology can improve their operations and productivity. The session, a partnership between the Homebuilders Association and Iowa State University's CIRAS, encouraged business owners to embrace AI as a tool for innovation. CIRAS has developed an AI training curriculum, delivered through community colleges across the state, to help businesses become more AI literate and fluent.

Sources

• Meta rolls out in-house AI chips weeks after massive Nvidia, AMD deals
• Expanding Meta’s Custom Silicon to Power Our AI Workloads
• Meta Developed Four New Chips to Power Its AI and Recommendation Systems
• Meta announces 4 new AI chips, raising competitive stakes with Nvidia, AMD
• Meta Preparing to Deploy Four New Homegrown Chips to Handle AI
• Google AI Introduces Gemini Embedding 2: A Multimodal Embedding Model that Lets Your Bring Text, Images, Video, Audio, and Docs into the Embedding Space
• Fish Audio Releases Fish Audio S2: A New Generation of Expressive Text-to-Speech (TTS) with Absurdly Controllable Emotion
• Artificial intelligence concerns raised by Academic Senate in coursework grading
• New nonprofit launches with at least $500 million to modernize scientific process for AI era
• Canva May Have Just Cracked One of the Biggest Issues With AI Images
• Gemini Leads ChatGPT In App Downloads In Feb, But ChatGPT Has Nearly 4x MAUs: SimilarWeb Data
• Analyse de Resemble AI (2026) : Génération de voix par IA réaliste ?
• NYT’s Blind Test With Readers Shows 54% Prefer AI-Written Content
• Wayfair boosts catalog accuracy and support speed with OpenAI
• Iowa business owners learn how artificial intelligence and improve productivity