China is allowing small batches of Nvidia's H200 chips onto the mainland to help domestic AI firms compete with the US, with ByteDance and Tencent each receiving around 10,000 H200 processors, accordi…
China has approved ByteDance and Tencent to import 10,000 Nvidia H200 chips each, with Alibaba also winning approval, marking a reversal from Beijing's earlier push toward Huawei's Ascend chips. The U…
Nvidia's H200 processors have reached mainland China in limited quantities, with ByteDance and Tencent each receiving about 10,000 units in recent weeks, according to the Financial Times. U.S. regulat…
Modal, the cloud compute platform, announced a new feature that lets AI agents run inside full KVM virtual machines, providing kernel-level access and real GPU drivers, with hardware ranging from 1 vC…
The memory ceiling on local GPUs is the most reliable predictor of cloud inference cost trends, according to an analysis of 2026 hardware and pricing data. The article calculates that a 70B model requ…
Cloudflare released Unweight, a lossless compression system that reduces BF16 MLP weights in LLMs by 15–22% while producing bit-identical outputs, freeing roughly 3 GB of VRAM on Llama-3.1-8B running …
NVIDIA's data center GPU roadmap has advanced through four major architectures—Ampere, Hopper, Blackwell, and Vera Rubin—each delivering more memory, faster interconnects, and lower-precision compute …
An audit by computer science researchers finds that NVIDIA's Blackwell Ultra GPU (B300) nominally supports INT8 tensor-core compute at a 30:1 FP8-to-INT8 ratio, but the format is undeployable by defau…
Researchers from an unnamed institution released a paper on arXiv presenting IO Factory, a framework for simulating AI-enabled influence campaigns with up to 100,000 civilians and 10,000 operators, us…
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 …
Nvidia is collaborating with major Wall Street firms to orchestrate a financing initiative that could channel as much as $500 billion into AI infrastructure, according to a report from Inside AI. The …
Researchers introduced Data-Centric Parallel (DCP), a method that dynamically adjusts runtime settings based on each batch's sequence length to train deep learning models on variable long sequences, a…
Researchers introduced CubicQuant, a parametric non-uniform scalar format for large-language-model weight quantization that preserves dense integer code streams while adapting reconstruction levels wi…
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…
Cloudflare published an engineering post detailing how it quantizes models on Workers AI, including Moonshot's Kimi K2.6 with FP8 KV cache and Z.ai's GLM 5.2 with INT4 weights, achieving up to 41% hig…
Cloudflare's production benchmarks for Moonshot's Kimi K2.6 and Z.ai's GLM 5.2 show that quantizing the decode phase, not the prefill, yields the biggest cost and throughput gains for trillion-paramet…
Bloomberg News reported that Chinese AI startup Moonshot AI has access to approximately 20,000 Nvidia Hopper-generation chips through a computing agreement with Alibaba Group, providing key compute fo…
Nvidia's Jensen Huang announced at GTC that Vera Rubin delivers approximately 10x more inference throughput per watt compared to Blackwell (B200), but the comparison point matters because most organiz…
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…
Nvidia dominated the AI chip race in the first half of 2026, capturing 85% of the data center AI chip market, according to a report from Mercury Research. AMD held 10% and Intel 5%, with Nvidia's reve…