GPU Infrastructure

Artificial Analysis released AgentPerf, the industry's first benchmark for agentic AI. Initial results show NVIDIA Blackwell Ultra NVL72 leading, running 20x more agents per megawatt than Hopper. The benchmark measures how many concurrent agentic tasks a platform can support under real-world coding agent workloads.

The article compares AI infrastructure to Roman legion camps, arguing that AI companies like Meta are adopting temporary, rapidly deployable structures to match the fast depreciation of chips and prioritize time-to-market over permanence. This strategy echoes historical frontier booms, marking a shift from permanent assets to time-sensitive investments.

The article discusses the emerging trend of designing loops to drive AI agents instead of manual prompting, covering key figures' insights, Anthropic's Fable 5 rollout controversy, automated research systems, data infrastructure bottlenecks, inference speed optimizations, and agent tooling developments.

This paper presents G-MAPP, a GPU-accelerated framework for reactive motion generation that achieves up to 5x speedup by parallelizing world modeling and planning on GPU, enabling real-time perception-action coupling in dynamic environments.

This paper presents a portable, low-power, battery-operated vision-based fall prediction and detection system using human pose estimation on an AMD Kria K26 SOM. The system uses an Intel RealSense D455 camera and a three-stage pipeline (quantized YOLOX, A2J, and CNN) to achieve real-time, privacy-preserving fall detection on the edge. Results show 4.5 FPS throughput with 75.85% classification accuracy.

Prometheus Inc., an AI startup co-led by Jeff Bezos, raised $12 billion in Series B funding at a $41 billion valuation. The company is developing AI tools to accelerate hardware development, focusing on prototyping and pre-production manufacturing. The funds will mainly be used for computing infrastructure.

As generative AI evolves into agentic AI, European enterprises face new challenges in data sovereignty, cost control, and infrastructure. This article argues that regional cloud providers like Vultr offer better compliance, performance, and cost efficiency than traditional hyperscalers for agentic workloads.

Funding from investors including Nvidia, Amazon and Qualcomm will support the vendor’s development of humanoid robots and physical AI.

NVIDIA's GeForce NOW summer sale offers up to $70 off a 12-month Ultimate membership and $35 off a Performance membership. The cloud gaming service eliminates hardware barriers, provides instant access to high-performance RTX gaming across devices, and announces Guild Wars 3 coming to the platform with exclusive rewards for current Guild Wars titles.

Cohere has released its first developer-facing coding model, North Mini Code, a 30B total parameter mixture-of-experts model with only 3B active parameters per token. It runs on a single H100 GPU, supports 256K context length, and is optimized for code generation, agentic software engineering, and terminal tasks. The weights are open under Apache 2.0.

This article is the second part of the PyTorch profiling series, delving into the internals of nn.Linear layers, including transpose operations, bias-fused epilogue techniques, and the impact of torch.compile on a single linear layer. It then dissects the performance characteristics of a Multilayer Perceptron (MLP) with GeGLU activation, showcasing the scheduling and execution of GPU kernels.

Google has released DiffusionGemma, a new open-weight model under Apache 2 license, available for free via NVIDIA's NIM cloud API. It delivers impressive generation speeds exceeding 500 tokens per second.

Google released DiffusionGemma, a 26-billion-parameter model that generates text via diffusion, achieving 1,000 tokens per second on an H100 GPU—four times faster than autoregressive models, but with lower quality. It's currently experimental.

As robotaxi services expand globally, NVIDIA introduces Halos OS—a comprehensive safety system integrating certified OS, standardized interfaces, AI guardrails, and a validation framework to ensure safety is built into autonomous vehicles from the ground up.

DiffusionGemma is Google DeepMind's experimental open text generation model that uses text diffusion instead of standard autoregressive decoding, achieving up to 4x faster generation on dedicated GPUs. The 26B MoE model (3.8B active parameters) is built on the Gemma 4 backbone, supports multimodal inputs (text, image, video), has a 256K context window, covers 140+ languages, and is released under Apache 2.0.

Google DeepMind released DiffusionGemma, an experimental open model for fast text generation using parallel token generation. NVIDIA optimized it to run faster on GeForce RTX, RTX PRO, and DGX Spark systems, achieving up to 1000 tokens/sec locally.

A new report from OpenAI details PRC-linked influence operations using AI to target U.S. tech debates, data center narratives, tariffs, and false claims about ChatGPT.

Researchers at the University of Twente have shown that by adjusting GPU clock frequencies at the per-kernel level, they can save up to 14% of the energy used in LLM training with minimal impact on speed.

Easybilling is an AI-native billing and monetization platform designed for AI SaaS, APIs, agents, and GPU platforms. It supports subscription, usage-based, and credit-driven pricing with real-time usage tracking, prepaid wallets, automated invoicing, and global payments, enabling AI companies to scale monetization without building complex billing infrastructure.

NVIDIA GPUs with Confidential Computing are now used for confidential inference in Apple's Private Cloud Compute (PCC), which is expanding from Apple's own data centers to Google Cloud. The technology provides a hardware-based security layer to protect user data during processing.

This post demonstrates how to train robot policies for the Unitree H1 humanoid using NVIDIA Isaac Lab on Amazon SageMaker AI. It covers two compute options: SageMaker HyperPod for resilient, persistent clusters, and SageMaker Training Jobs for ephemeral, on-demand training. The solution provides a unified Docker image, experiment tracking with MLflow, and detailed walkthroughs.

SpaceX wants to launch data centers into space, and Elon Musk is pitching it as a near-trivial engineering problem ahead of the company's IPO. A first AI satellite would match the output of a single Nvidia GB300 rack. But Google's own research suggests real AI training would require about 10,000 tightly coupled satellites.

Amazon has partnered with Corning for data center expansion, following Corning's previous AI infrastructure deals with Nvidia and Meta.

China plans to invest roughly $295 billion in a nationwide AI data center network over the next five years, with at least 80% of technology from domestic suppliers like Huawei. Meanwhile, Taiwan is considering criminalizing AI chip smuggling to China for the first time.

Apple debuted AI features at WWDC with a strong privacy focus, but reliance on Google and Nvidia servers due to its late start raises questions about whether the promise can be kept.

At WWDC 2026, Apple showed off a rebuilt version of Siri. The assistant runs on foundation models developed with Google. For complex queries, it taps Nvidia GPUs.

T4T is a decentralized marketplace for AI inference where GPU providers bid for prompts, payments in xBZZ on Gnosis, routing over Swarm, without API keys or middlemen.

This tutorial provides a hands-on workflow for NVIDIA cuTile Python, a tiled GPU programming interface. It covers environment setup, implementing tiled vector addition, matrix addition, and matrix multiplication, with a PyTorch fallback. Includes correctness checks and benchmark comparisons.

WWDC 2026 brought no new silicon, but a structural rebuild of how AI runs on Apple silicon: a new inference framework (Core AI), a new model format (.aimodel), a new generation of on-device models (AFM 3), and a changed posture toward the cloud including a partnership with Google and NVIDIA. The most surprising tell: Apple's flagship cloud model runs on NVIDIA GPUs in Google Cloud.

MuJoCo-Drones-Gym is an open-source Gymnasium-compatible multi-drone environment built on MuJoCo, supporting arbitrary Crazyflie 2.x nano-quadcopters with modular physics, action, and observation APIs, and including a PettingZoo wrapper for multi-agent RL and seven task environments.

This article explains how to migrate GitHub Actions CI to Hugging Face Jobs to overcome limitations of GitHub-hosted runners, such as slow speed and lack of GPU access. By setting up a dispatcher Space, a GitHub App, and modifying the runs-on label, CI jobs can run on Hugging Face infrastructure with CPU or GPU hardware, streaming logs in real-time. Trackio's experience shows a ~30% reduction in CPU job time.

Apple announced that developers with fewer than 2 million first-time App Store downloads can use its Foundation Models in Private Cloud Compute at no cost, aiming to attract smaller developers by lowering AI infrastructure barriers.

HPE announced the ProLiant Compute DL394 Gen12, a 2U server based on the NVIDIA Vera CPU, designed for agentic AI and data-intensive workloads. It integrates HPE's enterprise management and security stack, with a collaboration with NVIDIA and Redpanda; the NYSE is exploring it for its agentic AI infrastructure. The server features a monolithic architecture, LPDDR5X memory with up to 1.2TB/s bandwidth, and quantum-resistant cryptography. Availability is expected in fall 2026.

The deals come as Nvidia expands its presence in South Korea, spanning robotics, chip design and AI infrastructure.

Google has ordered more than three million AI chips from Intel for 2028. Nvidia is testing Intel's manufacturing tech for its upcoming Feynman architecture. Both moves come as TSMC can't keep up with AI chip demand. Intel's long-struggling foundry division is getting a rare second chance.

Xiaomi's MiMo team, with TileRT, released MiMo-V2.5-Pro-UltraSpeed, a serving mode for the MiMo-V2.5-Pro model. It decodes over 1000 tokens per second on a 1-trillion-parameter model using a single 8-GPU commodity node. The speedup comes from FP4 quantization, DFlash speculative decoding, and the TileRT runtime. API trial runs June 9–23, 2026.

One year after NVIDIA CEO Jensen Huang and UK PM Keir Starmer declared the UK would be an AI maker, NVIDIA and partners showcase progress at London Tech Week. Key developments include doubling of sovereign AI cloud providers, Isambard-AI supercomputer, Sovereign AI Fund backing startups, and enterprise AI deployments across sectors.

This new adage is a modern twist on the classic programmer quip "the cloud is someone else's computer," reflecting the reality of AI reliance on external infrastructure.

We present Accelerated Fourier SAT (AFSAT), a GPU-accelerated solver for pseudo-Boolean satisfiability based on continuous local search (CLS). AFSAT realises the proof-of-concept approach, FastFourierSAT, into a fully-engineered solver supporting any heterogeneous mixture of symmetric constraint types and lengths within a single problem instance. Using the JAX compiler, AFSAT leverages pure function composition, automatic vectorisation, automatic differentiation, and just-in-time (JIT) compilation to perform massively parallel CLS across batches of candidate assignments. We demonstrate substantially improved numerical stability, runtime performance, and memory efficiency over the proof-of-concept. We achieve this by way of identifying and addressing various limitations that arise from memory latency and floating-point representation, as well as leveraging automatic parallelisation and compact representations. The inherent representational and stability limitations of floating point are partially addressed by a tailored discrete Fourier transform implementation. We achieve near-linear throughput when scaling to multiple accelerators via JAX array sharding.

NVIDIA and LG Group are building an AI factory to accelerate LG's AI-driven businesses in robotics, autonomous driving, data center technologies, and GPU cloud services. The collaboration integrates NVIDIA's full-stack AI factory platform with LG's leadership in consumer electronics and robotics, aiming to create a unified workflow for physical AI systems.

OpenEnv is a tool for creating an agentic execution environment like terminals, browsers, or anything an agent can interact with. Today, we’re excited to announce that OpenEnv is becoming even more open, to make the future of training agents open source. Starting today, OpenEnv will be coordinated by a committee that so far includes Meta-PyTorch, Reflection, Unsloth, Modal, Prime Intellect, Nvidia, Mercor, Fleet AI, and Hugging Face. OpenEnv now lives at huggingface/OpenEnv. The project focuses on being an interoperability layer for RL environments, not a reward framework or trainer.

NVIDIA and Doosan Group are expanding their collaboration to advance new opportunities across physical AI, robotics and AI factory infrastructure, spanning Doosan Robotics, Doosan Bobcat, Doosan Enerbility and Doosan Corporation Electro-Materials BG.

The future of AI will be dominated by companies that control key physical infrastructure, such as chip manufacturing, data centers, and energy supply.

The AI infrastructure market is seeing a structure where giants invest in each other, buy each other's services, and generate each other's revenue, reminiscent of the dot-com bubble. SpaceX's pre-IPO AI computing leases with Google and Anthropic raise concerns about revenue authenticity. Investment focus should shift from GPU sales to bottlenecks like power and cooling.

Nyxclaw is an open-source project that gives AI agents a real-time face and voice. It runs locally without a GPU, offers self-hosted servers with end-to-end encryption, two voice pipelines (OpenAI Realtime and local CPU stack), and supports ARKit blendshapes at 30 FPS.

Microsoft's Surface Laptop Ultra, powered by Nvidia's RTX Spark SoC, offers up to 128GB unified memory, a 20-core CPU, and RTX 5070-level GPU. Hands-on impressions reveal a premium build, improved thermals, and repairability, but pricing and battery life remain uncertain.

Her is a tool that analyzes Claude Code session traces, reconstructing events in plain English, flagging risky moves (deploys, config changes, secrets), and showing token usage. It runs entirely on the local GPU, no third-party AI API is called, and includes an 'Ask Her' assistant to answer questions from the trace.

This article details the configuration and memory calibration required to run the Qwen 3.6 35B MoE model at a 450,000 token context window on a single 32GB VRAM GPU (NVIDIA RTX 5090) using llama.cpp with TurboQuant and YaRN scaling. It covers model selection, quantization trade-offs, KV cache quantization, RoPE scaling, multimodal setup, replication guide, VRAM lifecycle management, and performance evaluation.

At GTC Taipei at COMPUTEX last week, NVIDIA unveiled RTX Spark, the superchip that reinvents Windows PCs for the era of personal AI agents. On the heels of this announcement, NVIDIA founder and CEO Jensen Huang headed to South Korea, where he introduced RTX Spark to the nation’s passionate gaming community. Leading game developers — including Korea’s KRAFTON and NC — are already working to bring their titles to RTX Spark-powered systems.

This tutorial walks through NVIDIA garak as an end-to-end framework for defensive LLM red-teaming. It covers setup, plugin discovery, dry runs, real-model scans, multi-probe evaluations, report analysis, custom probe and detector creation, and AVID export. The workflow enables comprehensive LLM security testing and vulnerability reporting.