OpenAI disclosed that it is reviewing 50 petabytes of historical training and evaluation records to audit autonomous agent intrusions, assigning roughly 7,000 GB200 and GB300 GPUs to an inspection pip…
A new arXiv paper (2610.00049v1) reports that replacing native sigmoid, tanh, and SiLU activations with degree-3 or degree-4 bfloat16 polynomial programs improved complete training-step throughput on …
Inferact, a startup whose team includes vLLM maintainers, released open-source "megakernels" for Google's TPU v7 (Ironwood) that delivered 709 tokens per second for a single user on Moonshot AI's Kimi…
At the 37th Hot Chips conference, OpenAI unveiled its custom inference chip Jalapeño, claiming it outperforms NVIDIA's GB200/GB300 systems with 1.5–1.9× more work per watt at peak throughput, 1.7–3.6×…
OpenAI has published benchmark results for Jalapeño, its first custom AI inference chip, claiming it outperforms NVIDIA's GB200 and GB300 in throughput, latency, and power efficiency simultaneously. T…
OpenAI unveiled Jalapeño, an in-house inference ASIC and system built with Broadcom, at Hot Chips 2026, claiming it outperforms NVIDIA GB200 and GB300 in throughput per kilowatt and latency for OpenAI…
A developer has created a new LLM skill called 'pragmatic-context' that enables AI models to infer sarcasm, jokes, hyperbole, roleplay, and indirect intent from conversational context. The skill instr…
Ray Direct Transport (RDT), a feature in Ray Core, enables RDMA-backed weight transfer in reinforcement learning for LLMs, achieving up to 7.5x speedup over naive RDMA implementations on GB200 nodes u…
Nvidia released Nemotron 3.5 Lightning on August 11, an open-weight reasoning model with 30 billion total parameters and 3 billion active per token, designed to run agent workloads on a single Nvidia …
Behind-the-meter (BTM) power is becoming a viable option for hyperscalers to accelerate AI infrastructure deployment, driven by a shift from training to inference workloads and improved rack-level pow…
Compute costs could rise 10x or more in coming years, according to an analysis of AI lab economics. If a human-level software engineer could run on an H100 equivalent, that GPU should rent for over $2…
Nvidia's new Vera Rubin platform, named after the astronomer, eliminates nearly all cables from its AI server trays, reducing assembly time from two to three hours and cutting initial failure rates fr…
Nvidia demonstrated its Vera Rubin platform at a lab tour in Sunnyvale, California, claiming the compute tray can be assembled in one minute versus 90 minutes for prior GB200 hardware, a 90x improveme…
NVIDIA released Nemotron-Labs-Diffusion, a single model that eliminates the need for a separate draft model in speculative decoding, achieving 6.82 accepted tokens per forward pass in self-speculation…
Bitdeer Technologies sold all 3,231 BTC it held since February 2026, generating over $205 million to fund a pivot into AI and high-performance computing infrastructure. The Singapore-based miner now h…
Microsoft's Azure cloud platform achieved the fastest AI training results at the largest reported scale, powered by 512 Nvidia H200 GPUs delivering a 28% speedup over prior configurations. The milesto…
Microsoft and OpenAI are planning to deploy hundreds of thousands of NVIDIA GB200 GPUs, alongside their custom Maia AI chips, as part of massive infrastructure expansion. Hyperscaler capital expenditu…
Anthropic announced it is expanding its partnership with SpaceX to increase its computing capacity at the Colossus 2 facility, utilizing GB200 hardware throughout June. The company will begin ramping …
NVIDIA released Nemotron-Labs-Diffusion, a tri-mode language model that unifies autoregressive, diffusion, and self-speculation decoding within a single architecture. The model, trained with a joint A…
PyTorch 2.11 now enables direct installation of CUDA-enabled PyTorch wheels on aarch64 Linux from PyPI, eliminating the need for custom package indexes and workarounds that previously complicated depl…