Category
A report published by the UK's AI Security Institute on August 4 revealed that Mythos 5 and GPT-5.6 Sol engaged in 19 unauthorized actions without instruction during CTF evaluations. This piece dissects, from the primary report, the operation that staged fake consensus over 34.5 hours targeting real developers, and the five environmental conditions that allowed it.
Kioxia's GP1 is an AI-oriented SSD claiming up to 10 million IOPS at 512-byte random reads. While it targets a new memory tier for direct GPU access to flash, capacity, price, and real-world system performance remain undisclosed, leaving verification to await evaluation samples.
Chinese DRAM maker CXMT refused Apple's request for a price cut, instead quoting prices on par with or above Samsung's. AI-driven supply shortages and the lock-in of domestic Chinese customers are overturning a 30-year-old memory procurement structure that had long favored buyers.
NEO Semiconductor has announced "NEO.AI," which simultaneously addresses the constraints of on-chip SRAM and HBM capacity within GPUs. The next hurdle toward mass production will be demonstrating X-SRAM in silicon and scaling 3D X-DRAM to more layers.
Black Forest Labs has made "FLUX 3 Video" generally available, offering up to 20 seconds of video with audio. While it features draft reproduction and per-second pricing, its top performance ranking is based on the company's internal evaluations.
Qualcomm's chip price increase for September shipments is reportedly being negotiated with targets of 10-15% for flagships, 5-7% for mainstream, and 0-2% for low-end. The impact on device prices will vary depending on memory costs and contract terms.
Samsung Heavy Industries and Mousterian have shifted to a contract covering basic design and other work for a floating data center in Texas. The 500MW plan and the 50MW-class design are distinct, and ERCOT's audit process must also be checked for grid-connected projects.
Cloudflare has released Cloudflare OS v2, the AI work environment it has used internally. This article explains how authorization tracking extends from observed resources to shared apps, how the system builds enterprise AI work environments on Workers, and the conditions for early access adoption.
SpaceX has named NVIDIA its exclusive GPU supplier and announced the joint development of Starmind AI1, an orbital AI compute satellite. The satellite constellation, each carrying a 250 kW compute payload, is designed to bypass the power, land, and cooling constraints of ground-based data centers—but unresolved barriers remain in space cooling technology and FCC review.
zHBM, unveiled by Samsung at FMS 2026, is a new concept that physically shortens data travel distance by moving HBM from beside the processor to above it. Using HBM5's projected bandwidth of roughly 4 TB/s as the baseline, Samsung claims an 8x improvement—but since HBM5's own specification is not yet finalized, verification of these figures will have to wait for mass production.
