NVIDIA is quietly acquiring long-haul dark fiber across the United States, with fiber counts running up to 100 pairs, according to Wolfe Research. Using DWDM equipment and an assumed 800 Gbps transmission speed, the network’s theoretical capacity would scale to 7.6 petabits per second. The buildout is framed as a multi-year infrastructure effort, with Needham previously describing a telecom network project that could cost $5 billion to $10 billion over the coming 3 years. The company has also reported a 9.3 percent stake in the neocloud Nebius.

UBS raised its forecast for the sequential increase in DRAM prices in Q3 to 32% from 17%, and lifted its Q3 NAND price estimate to 30% from 17%. TrendForce also expects DRAM prices to rise 18% sequentially in Q3 and NAND to gain 15%. UBS said the supply-demand gap is widening, with 2027 demand projected to grow 36.2% versus supply growth of 19.3%, keeping the market constrained into Q2 2028.

Jefferies Equity Research expects DRAM and NAND prices to rise 40%–50% quarter-on-quarter in Q3 2026, followed by another 30%–40% increase in Q4 2026. The report also projects 2027 pricing to be up 40%–45% year-on-year. It attributes the tight market to constrained global supply, a surge in AI-driven demand and delayed new capacity, while saying Chinese suppliers are unlikely to ease shortages in the near term. Micron has signed 16 Strategic Customer Agreements and long-term contracts could reach 70%, further limiting supply for consumer devices.

Support for an NVIDIA GPU architecture labeled “BlackwellNext” has appeared in Linux 7.2 kernel patches. The change adds CXL DVSEC-based readiness polling in the nvgracegpu VFIO-PCI variant driver, according to Phoronix. The update is tied to AI and HPC data center platforms rather than any consumer GeForce launch, and it includes no shipment figures.

Mesa 26.2’s development branch has merged initial support for NVIDIA DLSS in the open-source NVK Vulkan driver, though it requires NVIDIA’s proprietary DLSS libraries. The change allows Linux users to enable AI upscaling in compatible Vulkan games, improving performance and image quality. Despite the closed-source dependency, it marks a key step toward narrowing the feature gap with NVIDIA’s proprietary Linux driver.

Foxconn disclosed that building an AI data center based on NVIDIA’s Vera Rubin architecture could cost as much as $47 billion per 1GW, with annual electricity costs reaching $1.3 billion. The company also said hardware depreciation is 6x higher than the power bill. Initial systems have been shipped to cloud providers for validation, and full-volume production is already underway.

AMD and Intel have released an updated specification for AI Compute Extensions (ACE), defining native matrix-multiply instructions, tile register state, and support for low-precision data formats for x86. The goal is to improve the energy efficiency and scalability of AI workloads running on general-purpose CPUs. The update is positioned as a coordinated x86 ecosystem upgrade rather than a product launch or revenue outlook, and it reinforces Intel’s technical roadmap credibility in edge-AI and hybrid computing scenarios.

ASML CEO Christophe Fouquet said in a Bloomberg interview that Elon Musk’s Terafab initiative resembles a “million-wafer” scale fab project, comparable to large DRAM lines in South Korea. He added that the AI-driven chip capacity buildout remains in the early stages, and that delivering the necessary equipment takes time. The comments reinforce investor attention on ASML’s role as a key supplier to AI infrastructure spending, particularly for EUV systems used in advanced-node expansions.

GMKtec发布EVO-X3迷你工作站，搭载AMD Ryzen AI MAX+ 395处理器、128GB内存，尺寸接近PS4，支持eGPU扩展，定位本地AI开发与边缘部署，明确对标NVIDIA DGX Spark。该产品以显著更低价格提供相近AI推理能力，并强调零云依赖、零token成本。目前未公布售价，但已开放6月22日早鸟预售。事件不涉及NVIDIA财务数据或 supply chain 变动，属竞争格局层面边际压力。