Open-Source Multimodal AI Surges, DeepMind Accelerates Materials Science, and AWS Bolsters AI Infrastructure for Federal Workloads
Today's AI landscape sees significant advancements across the board: a new open-source multimodal LLM, Kimi K3, is challenging proprietary models with its impressive capabilities. Meanwhile, Google DeepMind's 'Alchemy' project has achieved a breakthrough in materials science, predicting novel superconducting materials with AI. Concurrently, AWS is making a massive investment to expand its AI and supercomputing infrastructure, specifically targeting secure federal and national security workloads.
Open-Source Multimodal LLMs Reach New Heights with Kimi K3
The open-source AI community continues to push boundaries, with Moonshot’s Kimi K3 emerging as a formidable contender in the multimodal LLM space. Launched on July 16, 2026, Kimi K3 is currently the largest open-weight model, boasting a staggering 2.8 trillion parameters. This Mixture-of-Experts (MoE) model activates roughly 50 billion parameters per token and features a 1 million-token context window, alongside native multimodal input capabilities.
Kimi K3’s architecture incorporates in-house developed innovations like Kimi Delta Attention (KDA) Attention Residuals (AttnRes), contributing to its impressive performance. It leads open-weight coding benchmarks such as DeepSWE and Terminal-Bench 2.1, and its overall intelligence and coding index scores are among the highest. While its vast parameter count necessitates multi-GPU clusters for self-hosting, models like GLM-5.2 or the Kimi K2 line offer practical alternatives for more modest hardware setups.
Why it matters: The rapid advancement of open-source models like Kimi K3 signifies a crucial shift towards democratizing access to frontier AI capabilities. Its multimodal nature and strong performance in coding benchmarks make it an invaluable tool for developers, potentially lowering the barrier to entry for complex AI applications and fostering innovation across various industries. This competition also puts pressure on proprietary models to continuously innovate and justify their closed-source nature, benefiting the entire AI ecosystem.
AI Predicts Novel Superconducting Materials, Accelerating Scientific Discovery
In a significant leap for scientific research, an international team of physicists, coordinated through the SuperC international project, has leveraged machine learning to predict two previously unknown superconducting materials: YRu₃B₂ and LuRu₃B₂. This breakthrough, powered by AI, dramatically reduces the traditional trial-and-error process that could take years to discover such materials.
The researchers, including those from Aalto University in Finland and Rice University in Texas, trained a machine-learning model to screen vast databases of untested chemical structures, quickly discarding unpromising candidates. Only the most promising structures underwent computationally intensive quantum calculations, leading to the successful synthesis and confirmation of their superconducting properties in the lab. This approach, detailed in a paper published in Nature Computational Science, represents a fundamental paradigm shift from screening to generation, where AI proposes entirely new molecular structures optimized for specific target properties.
Why it matters: This development underscores AI’s transformative potential beyond generative content, directly impacting critical scientific domains like clean energy and materials science. By accelerating the discovery of novel materials, AI can significantly speed up the development of technologies crucial for addressing global challenges, from zero-loss energy transmission to advanced computing hardware. This shift towards AI-driven scientific automation is poised to reshape R&D across industries, compressing discovery timelines from years to weeks.
AWS Bolsters AI Infrastructure with Massive Expansion for Federal and Agentic Workloads
Amazon Web Services (AWS) is making a substantial commitment to scaling its AI infrastructure, announcing plans to deploy an additional 2 million NVIDIA Blackwell Ultra, Rubin, and Rubin Ultra GPUs across its global infrastructure in 2027-2028. This massive expansion builds on an already rapid adoption of NVIDIA-accelerated compute on AWS and aims to meet the surging global demand for AI compute capacity, particularly for agentic AI, scientific discovery, enterprise automation, and physical AI workloads.
Crucially, AWS is also planning a significant investment of up to $50 billion starting in 2026 to build out AI and supercomputing capacity specifically for the U.S. government. This includes deploying 100,000 GPUs on AWS’s secure infrastructure for federal and national-security workloads, classified at Impact Level 6 (IL6) and above. This initiative will also bring NVIDIA Vera CPU-based infrastructure to AWS and extend NVIDIA NVLink Fusion with custom NVHBM, providing a comprehensive, secure, and scalable AI stack for government agencies.
Why it matters: This enormous investment by AWS signals a critical phase in AI infrastructure development, addressing the escalating compute demands from both commercial and governmental sectors. For developers, this means enhanced access to cutting-edge hardware and services, facilitating the deployment of increasingly complex AI models and agentic systems. The focus on secure federal infrastructure highlights the growing importance of AI in national security and critical government operations, ensuring that advanced AI capabilities can be developed and deployed with the highest levels of security and reliability.
The Bottom Line
Today’s “Signals from the Latent Space” reveal a dynamic AI ecosystem where open-source innovation is thriving, scientific discovery is being revolutionized, and foundational infrastructure is rapidly expanding to meet unprecedented demand. The emergence of powerful open-source multimodal models like Kimi K3 democratizes advanced AI, while breakthroughs in materials science powered by DeepMind’s AI demonstrate the technology’s profound real-world impact. Simultaneously, AWS’s massive infrastructure investment, particularly for secure federal workloads, underscores the strategic importance of robust and scalable compute for the future of AI development and deployment.
📎 Sources
- Best Open Source LLMs (September 2026)
- AI predicts two new superconducting materials—and moves us closer to zero-loss energy
- Top 7 open source LLMs for 2026 - NetApp Instaclustr
- How to position for the next wave of AI-led scientific breakthroughs - Investing.com
- AI-Accelerated Materials Discovery in 2026: How Generative Models, Graph Neural Networks, and Autonomous Labs Are Transforming R&D - Cypris AI
- AWS and Nvidia to Deliver 2 Million Additional GPUs and Next-Generation Infrastructure for Agentic and Physical AI - StorageNewsletter
- Best Open Source LLMs in 2026: We Reviewed 7 Models - Fireworks AI
- Materials AI Is Advancing Fast, but One Big Piece Is Still Missing - AZoM
- AWS and NVIDIA to Deliver 2 Million Additional GPUs and Next-Generation Infrastructure for Agentic and Physical AI - Amazon’s Press Releases
- AWS plans major AI and supercomputing expansion for US government
- AI helps design new materials that work in the real world | MIT News
Get signals in your inbox
AI-curated digest of what matters in AI & tech. No spam.