SemiAnalysis Founder: Most AI computing power will belong to only two companies by 2028.

Beating AI News Insight: Zhipu AI has confirmed that Ox Alpha is a new version of its GLM series, with the model’s weights set to be released publicly tonight. Earlier, the community had already identified Ox Alpha as Zhipu’s work via clues such as API error logs and tokenizer details, and this has now been officially verified. Ox Alpha anonymously launched on OpenRouter over the weekend, focusing on programming and long-duration agent tasks, while supporting inputs of text, images, and videos. Shortly after its launch, it rose to become the most-used model on OpenRouter, with usage volume twice that of DeepSeek. OpenRouter described this as the largest model release in the platform’s history. Multimodality is a notable shift here: Zhipu previously separated its core GLM models from GLM-V visual models – for instance, GLM-5 only supported text, while images and videos were handled by models like GLM-5V-Turbo. This time, Ox Alpha is confirmed as a new GLM version that can process both images and videos. If the final official model name drops the "V" suffix, it would mark Zhipu’s move to merge its two model lines. Ox Alpha will stay free for one week; post-period pricing has not been announced.

MINIMAX announced in a Hong Kong Stock Exchange (HKEX) filing that its revenue for the first half of 2026 reached $117 million, representing a 283.1% year-on-year surge. The company recorded a net loss of $358 million over the period, down 11% year-on-year, while its gross profit jumped 464.8% year-on-year to $20.813 million.

According to on-chain analyst Ai Yi (Twitter handle @ai_9684xtpa)’s monitoring, address 0xa69…49eAd has been steadily accumulating ETH since August 17, amassing a total of 6,000 ETH at a price range of roughly $1,894 to $2,505. The address’s current unrealized profit stands at approximately $130,000; it now holds ETH worth around $14.56 million, with an average entry price of $2,427.29. Its most recent accumulation occurred roughly 3 hours ago.

According to TradingBeats monitoring, an address starting with 0x7dac launched a non-deleveraging TWAP order for AAPL at 16:08 today, planning to buy 16,000 AAPL shares over 7,230 minutes (roughly 5 days). Calculated at a snapshot price of $309.51, the total planned value is approximately $4.952 million, with execution expected to run until 16:38 on August 31. As of 17:00, only 117.27 shares of this TWAP have been filled, with a trading volume of ~$36,300, an average price of ~$309.59, and a completion rate of about 0.73%. There are still around 15,880 shares pending execution, valued at ~$4.916 million, accounting for ~99.27% of the total plan. TradingBeats, an on-chain perpetual contract (perp) and address analysis tool, is now live, supporting real-time access to Hyperliquid data—from address tracing to whale operations, delivering in-depth analysis at a glance.

Renowned crypto KOL Ansem posted on pump.fun that the NetNet Capital protocol is an innovative DeFi mechanism on the Robinhood chain. “The protocol’s token NET is a reserve share backed by at least 1 USDG in the protocol’s on-chain asset portfolio. NetNet Capital accumulates productive assets such as stablecoins and stocks through multiple mechanisms. If the net asset value (NAV) is 1.75 times the base treasury, NET token stakers can earn a 1.2% daily return. Its current trading price is approximately 11 times the treasury, and the treasury’s growth rate is far outpacing the 1.2% NET issuance rate,” Ansem explained. Additionally, Ansem noted that the protocol’s founder has given several interviews, stating he has discussed future collaborations with the Robinhood team and previously worked on the NBA Topshot project. Transaction records show Ansem invested $57,600 on pump.fun to purchase NET tokens. As of press time, per GMGN market data, NET’s market cap stands at $51.47 million, with a 24-hour gain of 61.66%.

Beating AI Express News reports that Skild AI, a general-purpose robot brain company, has launched its next-generation robot foundation model S1. Its biggest breakthrough is that teaching robots new tasks no longer requires model retuning: users only need to show it a human demonstration video, and the robot can directly learn by example. The tasks demonstrated by the official lasted up to over 10 minutes, including previously untrained scenarios such as flipping pancakes, hand-brewing coffee, and repotting plants. S1 treats the demonstration video directly as a prompt to generate corresponding robot actions. In the repotting experiment, it took just 11 minutes from the start of recording the human demonstration to the robot initiating its own execution. Even if interrupted or encountering errors mid-task, the model adjusts based on the current environment instead of mechanically repeating actions from the video. In Skild’s internal evaluation, both models were pre-trained using the same 100,000 hours of data. When tackling unseen new tasks, S1 achieved an average single-step success rate of 66%, while traditional language-instructed Vision-Language-Action (VLA) models only hit 9%. The latter would need to collect roughly 380 additional task demonstrations and undergo fine-tuning to match S1’s performance after watching a single video. S1 has now been deployed to a small number of industrial clients, with plans to expand its client base in the coming months.