All NVIDIA items under Companies on AIFIRST News, newest first. 25 items.
A modified Nvidia RTX 5090 with 96 GB of VRAM has appeared on Alibaba, reportedly available for under $4,000. The modification triples VRAM over the reference card at roughly 65% of its cost, making it relevant for running large local models on a single GPU. Community reaction covers both the opportunity and the supply chain risks of grey-market hardware.
NVIDIA's DLSS 5 can now be applied to the entire desktop environment, allowing real-time AI upscaling of videos and images as they appear on screen. This extends AI upscaling beyond games to any visual content displayed on compatible hardware.
NVIDIA's Sol-Engine team released a fast inference method for MiniMax H3 that significantly improves generation speed compared to standard sampling. The technique is not yet available as a ComfyUI node, and the community is calling for an implementation.
AMD published ROCm 10.0, a major update to its open GPU compute stack marking a decade since the first release. The update focuses on agentic AI use cases with performance improvements for training and inference on Instinct GPUs. AMD positions ROCm 10 as its answer to CUDA for teams running large-scale AI pipelines on AMD hardware.
NVIDIA acquired Hugging Face for $12.93 billion, with the AI community noticing an easter egg embedded in the exact acquisition price. Hugging Face co-founder Julien Chaumond and prediction market Polymarket both commented publicly on the deal. The acquisition is one of the largest in AI history and marks a major consolidation of open-source AI infrastructure under NVIDIA's ownership.
Nvidia has announced a definitive agreement to acquire Hugging Face, the dominant platform for open AI model hosting and collaboration, for $12.9 billion. The deal gives Nvidia direct control over the largest community hub for AI models, datasets, and inference tooling. Community reaction focuses on concerns about platform neutrality and what this means for open-source AI distribution.
Following Nvidia's acquisition of Hugging Face, community members are highlighting ModelScope, Alibaba's model hosting platform, as a neutral alternative for downloading and sharing AI models. ModelScope hosts many of the same popular open-weight models and supports standard download tooling. The discussion reflects broader concern about concentration of AI infrastructure ownership.
A Stable Diffusion user tested NVIDIA DLSS 5 on hyper-detailed renders and documented the results, noting an initial color shift that confused the expected behavior. After trial and error, the author found the conditions under which DLSS 5 improves quality rather than introducing artifacts. The post is a practical guide for artists using DLSS with AI-generated imagery in their pipeline.
Asus quietly raised the price of the Ascent GX10 desktop AI workstation by \,000, with the 1TB model now listed at \,999 and the 4TB variant reaching \,999. The thread suggests this may signal a coming price increase for NVIDIA's DGX Spark as well. The moves come as desktop-class AI hardware demand remains strong despite supply and tariff pressures.
A developer released a standalone Windows application that applies NVIDIA's DLSS 5 neural rendering pipeline (Feature-18) to arbitrary images and video without requiring a DLSS-capable game or GPU-bound application. The tool wraps DLSS 5 for offline enhancement of existing content. Source code is available on GitHub.
Apple released updated Mac mini and Mac Studio models ahead of schedule, driven by unexpected enterprise demand for local AI hardware. Fulfillment challenges arose from global memory shortages and the lack of dedicated business and developer relations teams. Some orders shifted to Nvidia as a result.
A technical analysis examines Nvidia GPU architectural limitations for inference workloads and contrasts them with OpenAI's custom inference chip design, referred to as Jalapeño. The piece covers tradeoffs around memory bandwidth, compute density, and transformer-specific operations. It argues that purpose-built inference silicon can substantially outperform general-purpose GPUs for large language model serving.
Nvidia has agreed to acquire Hugging Face for $12.9 billion, bringing the largest open-source AI platform under hardware giant control. The deal would give Nvidia direct influence over the model repository and tooling that much of the AI ecosystem depends on. Community reaction has been largely skeptical, with concerns about open-source access and potential hardware lock-in.
In addition to the Hugging Face platform acquisition, Nvidia would effectively acquire the llama.cpp project and its core development team, which joined Hugging Face in February. llama.cpp is the dominant open-source runtime for running local LLMs on consumer hardware. This concentration raises questions about whether the project will remain hardware-neutral and freely licensed going forward.
Researchers achieved byte-identical output from a 72-billion-parameter LLM running on both AMD MI300X and Nvidia H100 hardware, directly contradicting the assumption that large-model results diverge across GPU vendors. The finding has concrete implications for reproducibility standards in model evaluation and hardware-independent benchmarking.
SemiAnalysis reports that OpenAI's custom AI chip, codenamed Jalapeño, outperforms Nvidia's Blackwell architecture on key inference metrics, a significant claim if accurate. The analysis suggests OpenAI could meaningfully reduce its dependence on external GPU supply for training and inference.
Xiaomi announced the AI Cube, a prototype three-chip AI inference system combining its Xuanjie O3, O100, and D100 processors in a single package with 1.2 TB/s of aggregate memory bandwidth. The system is positioned to compete with Nvidia-based inference appliances and Chinese domestic AI accelerators. Detailed production timelines and pricing have not been announced.
Nvidia has begun notifying customers of AI-related price increases exceeding 15% across its product lines. The hikes are linked to sustained demand for AI accelerators and constrained supply. The price increases affect both consumer and enterprise hardware buyers in the local AI ecosystem.
A developer hosted Kimi K3, Moonshot AI's 2.8 trillion parameter model, on 8 NVIDIA B300 GPUs via Modal, achieving 92 tokens per second at $190 per million tokens. Cold boot takes around 27 minutes to load the 1.56 TB model at a cost of roughly $56.79 per hour. The experiment shows frontier-scale inference is now accessible but remains expensive outside managed cloud services.
A technical comparison analyzes Etched's Sohu transformer ASIC against Nvidia GPUs for inference workloads. The Sohu chip is purpose-built for transformer architectures, offering high throughput per watt but less flexibility than general-purpose GPUs. The analysis covers cost-per-token, latency, and deployment trade-offs for production inference at scale.
energygraph v1.3 adds discrete GPU power monitoring for NVIDIA, Intel, and AMD cards, extending the tool beyond its previous CPU-only scope. The tool is open-source with zero build dependencies and displays live power consumption data directly in the terminal.
Fizgig v4.3.0 adds AMD Radeon support via ROCm, covering RDNA1 through RDNA4 GPUs, for LoRA training on Flux 2 Klein, Krea 2, and MiniMax H3. Previously NVIDIA-only, the free open-source tool now provides an AMD-native training path on Windows.
Nari Labs describes five inference optimizations applied to Qwen3-TTS on a single NVIDIA H100 SXM, reaching 10 requests per second and sub-50 ms p95 time to first audio. Throughput is about 630 characters per second at full load, at roughly $2 per 1 million characters. The post benchmarks the result against vLLM-Omni, SGLang-Omni, and VoxServe, and against the commercial services ElevenLabs and Cartesia, claiming lower latency at higher request volumes.
llama.cpp is a C/C++ implementation of large language model and vision language model inference, built on the ggml tensor library and MIT licensed. It ships a command line interface, an OpenAI-compatible API server, and a web chat UI, installable as an app, via Docker, from pre-built binaries, or from source. Backends cover Apple Silicon with ARM NEON, Accelerate and Metal, x86 with AVX through AVX512 and AMX, NVIDIA CUDA, AMD HIP, Vulkan and SYCL, plus hybrid CPU and GPU inference. The repository showed 124.9k stars and 22.0k forks with 778 open issues and roughly 1.4k open pull requests, and quantization spans 1.5-bit to 8-bit so output quality depends on the level chosen.
Nvidia released SoL-Pi, a standalone extension for Pi-based agent frameworks that reduces inference cost without sacrificing task completions. It integrates with Pi's existing AutoResearch loop and is available on GitHub with a companion blog post from NVLabs.
AIFIRST 
