Nvidia's H100 and B200 chips are increasingly treated as appreciating assets rather than depreciating capital expenditures, according to a commentary on the AI hardware market. This shift is driven by…
Local AI models will not win because datacenter inference is inherently cheaper and more powerful, according to an analysis by an unnamed author. The author argues that datacenter models benefit from …
Nvidia is shifting AI infrastructure financing from operational expenditure to capital expenditure, positioning itself as the anchor that makes these investments safe for Wall Street, according to a n…
NVIDIA's H100, H200, and B200 GPUs offer different trade-offs for AI workloads, and choosing among them depends on factors like model size, memory requirements, and whether the use case is training or…
Runware, the AI inference startup founded by Flaviu Radulescu and Ioana Hreninciuc, launched Sonic Inference Pods, modular data centers in 20-foot shipping containers that pack about 1,200 GPUs and on…
Pokee AI released Pokee-Isaac 28B, a 28B text-only foundation model with a 10M-token context window designed to run inside the customer boundary. The model scores 93.3% on RULER at 10M tokens, leads B…
Pokee AI released Pokee-Isaac 28B, a 28B-parameter proprietary non-decoder-only model claiming a 10M-token context that fits on a single RTX 4090 (24GB) in quantized form. Vendor-reported benchmarks i…
Pokee-Isaac 28B, the world's first real 10M-token context frontier-class agentic model, is now deployable on a single GPU starting from RTX 4090, according to its release announcement. The model achie…
RunInfra AI optimized DeepSeek-V4-Flash-0731, achieving a 2.98x speedup on 4x B200 GPUs, with median latency reduced from 9248 ms to 3095 ms and throughput increased from 113 to 363 tokens per second,…
Wafer, an AI infrastructure provider founded by Emilio Andere and Steven Arellano, reported that running Kimi K3 on eight AMD MI355X GPUs delivered 48 tokens per second per dollar, outperforming Nvidi…
AMD's MI355X GPU achieved 48 tokens/sec per dollar running the 2.8T-parameter Kimi K3 model, outperforming Nvidia's B300 (33 tokens/sec) and B200 (7 tokens/sec), according to benchmarks from Wafer. Th…
Wafer's benchmark shows AMD's MI355X GPUs deliver 952 tok/s/node on Kimi K3, outperforming NVIDIA's B300 on performance per dollar at 48 tok/s/$ versus 33 tok/s/$, despite B300's higher aggregate thro…
Atomarine proposes deploying small modular nuclear reactors on maritime vessels to power floating data centers, addressing energy bottlenecks for AI training clusters. A 4,000-GPU cluster requiring 40…
Semiconductor stocks are sliding across US and Asian markets as investors shift from the AI hype cycle to demanding proof of returns on massive capital expenditures. Big Tech is spending billions on N…
Apple Inc. reached a $5 trillion market capitalization milestone as investors rotate away from AI pure-play stocks, reflecting a market shift from AI training to inference and deployment. The move tow…
LLM inference profitability depends on the trade-off between batch size and GPU count, which determines token latency and cost per million tokens. Applying this model to Kimi K3, which requires at lea…
Nvidia's market strategy faces a hardware bottleneck as government trade restrictions force the company to create 'lite' versions of its H100 and B200 chips, such as the H20, to meet legal requirement…
Moonshot AI released the Kimi K3 open-weight model with 2.8 trillion parameters on Hugging Face on July 27, but only 1.8% of parameters (about 50 billion) are active per forward pass due to Mixture-of…
Google has developed a new custom AI chip designed specifically for its Gemini large language model, aiming to drastically reduce inference costs and latency by addressing the memory wall and power in…
NVIDIA CUDA 13.3 introduces the clmad PTX instruction for carryless multiplication on all Ampere and newer GPUs, enabling hardware-accelerated cryptographic workloads. On the NVIDIA B200, GHASH throug…