OpenAI has announced details for Project Camellia, a long-term datacenter project in Effingham County, Georgia, outlining its initial community commitments regarding power, water, and local benefits. The company is contracting with Georgia Power Company for 3.2 gigawatts of power to be delivered between 2028 and 2032.
The llama.cpp b10306 release introduces performance optimizations for the SYCL backend, specifically targeting Gated Linear Unit (GLU) operations. The update adds a contiguous fast path for fused GLU ops and consolidates previously distinct kernels to share a common launcher.
The llama.cpp b10255 release extends the oneDNN SDPA path to support non-FP16 key-value (KV) caches, including Q4_0–Q8_0 quantized formats and FP32. This update allows the fused systolic kernel to run identically to the native FP16 path by dequantizing or converting K/V tensors to dense FP16 on-device before processing.
Lucebox has partnered with AMD to demonstrate a heterogeneous consumer hardware setup that outperforms the Nvidia DGX Spark in inference speed. The system combines an AMD Radeon AI PRO R9700 GPU with a Strix Halo processor to run the full 284B DeepSeek V4 Flash model.
The article argues that GPU utilization, rather than model intelligence, has become the primary constraint in enterprise AI, drawing parallels to airline aircraft utilization where costs accrue by the hour regardless of usage.
Patience Strong argues that current neural networks consume excessive compute power by emulating circuits in software rather than implementing them as physical hardware. The paper suggests replacing this emulation with a semiconductor device where weights and biases are stored in non-volatile memory.
Ai2 has launched the OlmoEarth Platform, an infrastructure system designed to take Earth observation foundation models from fine-tuning through to large-scale inference. The platform addresses the engineering challenges of processing terabytes of multimodal satellite data across multiple providers and resolutions.
Moonshot has released the weights for its Kimi K3 model on Hugging Face, featuring 2.8 trillion parameters and a mixture-of-experts architecture with 16 active experts per token. The team plans to deploy the model across A100, H200, and B300 GPUs, with benchmark results expected this week.
The llama.cpp project released version b10099, which includes a significant update to the CUDA backend for improved NVFP4 W4A4 activation quantization.
IBM has signed a definitive agreement to acquire HRL Laboratories, a historic research institution with nearly 80 years of scientific contributions. The acquisition aims to accelerate IBM's quantum computing vision by integrating HRL's advanced expertise in silicon-spin qubit engineering.
AMD and Anthropic have signed a major agreement involving tens of billions of dollars in AI servers, with Anthropic purchasing up to 2 gigawatts of AMD's Instinct MI450 chips starting in the first half of next year. Concurrently, AMD will invest up to $5 billion in Anthropic as specific deployment milestones are met, while both companies collaborate on identifying suitable data centers for the hardware.
SLAI introduces T-Rex, an end-to-end optimization framework for the full-parameter post-training of trillion-parameter-scale MoE models on Ascend NPU SuperPOD. Using the DeepSeek-V4 model family as the target workload, the system addresses memory pressure and communication overhead through hierarchical parallelism and kernel execution optimizations.
SLAI introduces T-Rex, an end-to-end optimization framework for the full-parameter post-training of trillion-parameter-scale MoE models on Ascend NPU SuperPOD. Using the DeepSeek-V4 model family as the target workload, the system addresses memory pressure and communication overhead through hierarchical parallelism and kernel execution optimizations.
IBM has been selected by the U.S. Department of Energy (DoE) to lead a project under the Genesis Mission Request for Applications, committing up to $50 million in quantum compute access to support the initiative.
This TLDR AI newsletter covers several developments in the AI industry, including new hardware, software agents, and routing optimizations.
Together AI and Y Combinator have announced a partnership to provide the first dedicated GPU cluster exclusively for Y Combinator's portfolio of AI-native startups. This initiative aims to solve the critical bottleneck of compute access by offering flexible, cost-effective infrastructure for training and inference without requiring long-term commitments.
Chinese artificial intelligence company Moonshot AI has become the first in China to run out of GPU capacity. Consequently, the company has decided to suspend new subscriptions and eliminate free access for users.
Together AI details the specific architectural requirements needed to achieve different reliability tiers for GPU inference, arguing that standard SLA numbers often obscure the actual failure domains covered.
Together has updated its GPU Cluster platform with new reliability and operational control features designed for large-scale training and inference workloads. The changes include a rebuilt Slurm-on-Kubernetes stack, automated node repair mechanisms, and enhanced access management tools.
A user is asking for recommendations and assistance in running 35B parameter models using the NVFP4 quantization format on an NVIDIA RTX 5090 GPU. The goal includes supporting a context length of 100k+ tokens, ideally with uncensored or specialized models.