U.S. stocks fell for a third straight session Tuesday as higher long-term bond yields put fresh pressure on richly valued technology names. The Nasdaq Composite dropped 1.33%, underperforming the Dow Jones Industrial Average and the S&P 500, while the 30-year U.S. Treasury yield briefly touched 5.338%, its highest level since 2007. The move was part of a wider global bond sell-off that also pushed long-dated yields higher in France, Germany, Japan and the U.K. Markets are increasingly focused on the growing debt burden tied to artificial intelligence expansion. According to figures cited in the report, AI-related bond issuance has reached $489 billion so far this year, well above an earlier full-year 2025 estimate of roughly $322 billion, while The Wall Street Journal reported that nine major technology companies have about $3 trillion in off-balance-sheet AI commitments. That backdrop hit semiconductors, memory, optical communications and AI cloud-service providers especially hard. Investors are also weighing fiscal deficits, oil-driven inflation risks tied to the Iran situation, and a heavy event calendar that includes U.S. tariffs on some Canadian products, a 20-year Treasury auction, Federal Reserve minutes and China’s one-year LPR decision.

Anthropic is in talks to acquire Israeli AI startup Decart for $6 billion, according to Bloomberg, in what would be the company’s largest acquisition to date if completed. The deal has not been finalized. Decart, founded in September 2023 by Dean Leitersdorf and Moshe Shalev, focuses on world models and cross-chip inference optimization. The company has about 100 employees, has raised roughly $450 million, and was last valued at about $4 billion after a $300 million round backed by investors including NVIDIA and Amazon. Decart said its optimization platform can deliver agentic inference speeds of up to 1,600 tokens per second, while its world-model system can generate video at 100 frames per second, though those figures have not been independently verified. In October 2024, the startup launched Oasis, which it described as the world’s first real-time interactive generative AI video model, and said the product passed 1 million users within three days. Israeli outlet Calcalist had earlier reported that NVIDIA was initially the leading buyer, before another international company entered the process. Possible buyers mentioned at the time included SpaceX, Amazon, and cloud provider Nebius, with a reported price range of $6 billion to $7 billion.

London’s King’s Cross has undergone a striking transformation from one of the city’s most notorious districts into one of the world’s leading centers for artificial intelligence. According to TechCrunch, the area now ranks behind only San Francisco and Beijing as a global AI hub, with companies including Google DeepMind, OpenAI, Anthropic, Meta, Isomorphic Labs, Cusp AI, and Synthesia clustered in and around the neighborhood. The shift traces back to redevelopment efforts in the early 2000s, helped by the rise of St Pancras as the Eurostar terminal and the presence of institutions such as University College London and the British Library. The local AI cluster accelerated after DeepMind moved into the area following its 2016 acquisition by Google. That growth is now pushing up real estate costs and intensifying hiring pressure. Knight Frank said AI-related companies have leased more than 1 million square feet of office space in London since June, while prime rents in King’s Cross have risen 18% over three years and vacancy for traditional office space has dropped to 0.9%. Dealroom data also shows London has about 3,600 AI startups, with $12.1 billion of the city’s $14.8 billion startup funding since late July going to AI companies. At the same time, local startups are facing a much tougher fight for talent as US tech firms expand their presence.

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.

Higgsfield has released the full production package for its 95-minute AI film Hell Grind, making public the project’s prompts, character models, scene assets, iteration history, a 19-minute behind-the-scenes breakdown, and a Claude Skill built around its prompting workflow. The company said the film was produced by a 15-person team in Almaty, Kazakhstan, over 14 days at a cost of nearly $500,000, with roughly $400,000 spent on GPU cloud compute and the rest on labor. CEO Alex Mashrabov described the process as feeling like a slot machine, with 253 usable shots selected from 16,181 generations, or about 64 to 1. Higgsfield also tied the release to promotion for its Higgsfield Global Film Festival, which carries a $1 million prize pool and a $500,000 top award. The open package details how the team locked character identity, structured prompts with anchors, coordinates, timestamps, and hard rules, and why it stopped using negative instructions such as “don’t” or “not game-like” after repeated failures.

Ethereum co-founder Vitalik Buterin said MiniMax H3 is the first open-source video model he has seen that outperforms Tencent’s HunyuanVideo 1.5. He added that Hunyuan staying ahead for a long period was itself an impressive achievement. In his view, leading models in image and video generation have become more closed than large language models, making H3 an important step toward reopening the field. Buterin also shared a video generated by H3 and said it took about 30 minutes to complete on his AMD laptop. The item was published by PANews on Aug. 7 and was presented as market information rather than investment advice.

MiniMax has become one of the market’s sharpest AI rerating stories in the span of a week. After releasing its multimodal generation model H3 on July 31 and open-sourcing it on Aug. 3, the company’s shares climbed 78.21% from the post-launch period through 10:53 a.m. on Aug. 7, including a 23.15% gain on the day. The source article argues that investors are no longer treating H3 as just another benchmark-driven model release. Instead, they are pricing in a broader thesis around video generation as a high-token-consumption workload, lower inference costs, and the ecosystem effects that follow an open-source strategy. According to MiniMax, H3 supports text, image, video and audio context, and can generate up to 15-second videos at 2K resolution, 24FPS, with native stereo audio. In Artificial Analysis blind testing, H3 scored 1242 Elo in text-to-video with audio, ranking second globally, while placing first in video editing and among the top three in image-to-video. The article also highlights price as a key factor, saying H3’s per-second 2K cost is less than one-third of flagship models and its 768P pricing is less than half of mainstream offerings. The strongest shift in sentiment came after open-sourcing. MiniMax said 16 chip vendors, multiple developer communities, cloud inference platforms and inference frameworks had already adapted or integrated H3, with more than 100 enterprises going live on Day 0. That has pushed the discussion beyond model rankings toward revenue growth, ecosystem distribution and strategic positioning.

Seedance 2.5 has rolled out in full, with the official Feishu handbook laying out a broader set of controls for AI video creation than the previous 2.0 version. The update raises native single-pass generation length to 30 seconds, allows repeated extensions to 60 seconds, and adds an ultra-long mode that supports up to 180 seconds. The documentation frames that shift as a move away from generating isolated clips and toward producing more complete narrative sequences in one workflow. The handbook places timestamp-based prompting at the center of the update. Users can specify second-level time windows to direct motion, camera movement, scene transitions, sound and dialogue, then make targeted edits to selected segments after generation while keeping continuity across the rest of the clip. It also expands multi-modal reference handling, with guidance to upload references by priority and explicitly assign each image, video or audio sample a role inside the prompt. Other changes listed in the manual include cleaner default outputs with fewer unwanted subtitles and background music, stronger physical realism, more consistent cuts, smoother handling of complex actions, support for mainstream foreign languages through direct prompting in the target language, and a recommended workflow built around planning, reference preparation, scripting, generation, localized edits and final post-production.