MiniMax released the weights for its omni-modal video model H3 on August 3, posting them on Hugging Face. The model can generate 2K, 24fps video with stereo audio, and it moved ahead of ByteDance’s Seedance 2.0 in the Artificial Analysis text-to-video ranking while taking first place in video editing. Pricing also came in lower: a 5-second clip costs about $0.70 for H3’s 2K output versus roughly $1.22 for Seedance 2.0’s 720p output. Community tests also showed the model running offline on consumer GPUs with 12GB of VRAM. The open-weight release, however, stops at 768p; 2K output still depends on a non-open API-side upsampling module, and the Community License carries regional and usage limits.

Kioxia rose 2.97% intraday on Aug. 4, according to Bitget market data, while the Nikkei 225 fell 0.53%. The move came as the company formally announced KIOXIA GP1, its first GP-series solid-state drive designed to support direct GPU access to high-speed flash storage. Kioxia said the product delivers 100M IOPS in random read performance. The company described the GP1 series as a product aimed at emerging AI storage architectures that bring high-speed flash media into the VRAM system. According to Kioxia, that setup allows AI systems to access larger datasets at a cost far below adding more HBM, while also improving GPU utilization.

BlockBeats on July 24 compiled a list of key developments ahead of the U.S. stock market open, with semiconductors and big tech at the center. Nvidia has sent price increase notices to its board partners covering GDDR6 and GDDR7 memory kits, while Apple is pressing OLED display suppliers to cut panel prices as soaring memory chip costs raise iPhone production expenses. Intel also stood out, with second-quarter 2026 revenue nearly 12% above analyst expectations and third-quarter guidance that came in above consensus across the full range. The roundup also highlighted comments from Kim Yong-beom, head of policy at South Korea’s presidential office, who said Samsung Electronics and SK Hynix are expected to sign major agreements involving “huge” sums with major U.S. technology companies during President Lee Jae-myung’s visit to Silicon Valley. Elsewhere, Oracle secured a $7 billion software contract from the U.S. Department of Defense, Amazon upgraded Alexa+ with broader AI assistant functions, and Nvidia signed a $1.5 billion agreement with Amkor Technology to expand advanced chip packaging capacity in Arizona. The report also noted upcoming earnings releases from SK Hynix, Samsung Electronics and Kioxia, as well as the latest court ruling in SK Group Chairman Chey Tae-won’s divorce case.

The author says NVIDIA’s $2,999 DGX Spark cut a roughly $1,900 monthly cloud GPU bill to about $10 in steady-state operating cost, keeping around $22,000 inside the business over the first year.

Odaily, citing Benchlife, reported that Nvidia has sent a price increase notice to its board partners covering GDDR6 and GDDR7 memory kits. Under Nvidia’s supply model, the company sells products to board partners as bundled kits that combine GPU chips with VRAM chips. Those partners then mount the components onto their own PCB designs to build custom graphics cards. With another increase in pricing for GDDR6 and GDDR7 memory kits said to be coming, AIB (Add-in Board) manufacturers are expected to pass the higher costs on by raising retail prices for graphics cards. The report did not disclose the size of the increase, but it said the adjustment applies to the memory kit portion of Nvidia’s supply package for partner-built cards.

Ethereum co-founder Vitalik Buterin has abandoned all cloud AI services, running a fully local and sandboxed AI system. He warns that roughly 15% of AI agent skills contain malicious instructions and details his private LLM setup using an Nvidia 5090 GPU and Qwen3.5 model.

Vitalik Buterin said DeepSeek V4’s low-memory release can run without centralized servers, opening the door to local AI use that could support Ethereum privacy and more automated smart contract audits.

Google DeepMind has released DiffusionGemma under Apache 2.0, saying it can generate text about four times faster than comparable autoregressive Gemma models, while public benchmark scores remain below standard Gemma 4.