Latest
This article compares Micron's proposed GPU-adjacent NAND with the company's HBM4, PCIe 6.0 SSDs, and the OCP HBF standard, examining the performance and implementation requirements needed for productization.
Researchers at Sophia University and NIMS added the organic molecule 2-ABZ to lead-free quasi-2D tin perovskite solar cells, raising conversion efficiency from 6.60% to 9.07%. The single molecule reduces crystal defects, suppresses iodide migration and limits tin oxidation, and the unencapsulated cells retained about 85% of their initial performance after 100 days in air.
Steam's unofficial database SteamDB has been acquired by Chosen, the parent company of Nexus Mods. The deal avoids a shutdown driven by its founder's burnout and aims to fold 13 years of build-history data into mod-management tools to reduce patch breakage.
Steam's Linux usage share dropped to 3.90% in August 2026, but remained 1.26 points higher than the same month last year. A closer look at the breakdown beyond SteamOS Holo—and the survey's limitations—shows why a single month's figure can be misleading.
An unofficial DLSS 5 benchmark shows the RTX 5090 Founders Edition reaching its 575W power limit. A closer look at the numbers against an 802W-class card shows why connector count alone can't explain the performance gap.
Broadcom's AI chip revenue now accounts for more than half of total company sales. Drawing on primary sources, we examine the growth in custom XPUs and networking, the terms of its rack financing deals, customer concentration, and supply constraints.
Meta's Muse Spark 1.3 closes in on rivals in long-context and code benchmarks, but the flagship comparison relies on a "max" reasoning tier that isn't yet available. We examine the currently accessible xhigh tier's scores, API pricing, and evaluation conditions.
Google and geothermal startup Fervo Energy have signed a deal to procure 396MW from Fervo's Cape Station in Utah. With an expansion option to 1GW and a direct-to-load framework enabled by Utah's SB132, the deal powers AI facilities at no cost to existing ratepayers.
Google is pulling DDR4 memory out of retired servers and reconnecting it to new-generation AI servers via a dedicated interface, turning component recycling into a supply-defense strategy amid a memory shortage.
On September 1, 2026, the DOJ filed a 20-page statement of interest in the NYT v. OpenAI MDL, arguing that copying for LLM training is a separate use and fair use. We look at why it isolates training, what it leaves out (the acquisition stage behind the $1.5 billion Bartz settlement), and how Japan and the EU differ.
