Moonshot AI is reportedly in talks with Microsoft, Amazon, and Google over revenue sharing for Kimi K3, with a maximum rate of up to 30%.

Beating AI Express: Surge in AI computing power demand continues to drive up storage chip prices, with upstream cost pressure starting to pass through to consumer electronics and hardware end products. Recently, Amazon has raised prices for multiple hardware products, including its best-selling entry-level Echo smart speaker, which jumped from $49.99 to $79.99 (a roughly 60% increase), and the 16GB base model e-reader, which rose from $109.99 to $149.99 (up around 36%). Before Amazon, tech firms including Apple, Microsoft, and Dell have already been responding to rising storage costs by raising product prices and cutting standard-equipment RAM in their hardware. Market research firm IDC previously warned that the global memory supply crunch could last until 2027, with the industry projecting storage chip prices may peak in 2028 before gradually declining.

Beating AI Express: SemiAnalysis founder Dylan Patel predicted in a recent podcast that by 2028, OpenAI and Anthropic could capture 70% to 80% of the world’s new AI computing power, with their combined capacity potentially exceeding 100GW. Patel noted the two companies currently account for roughly 30% of global annual new computing power, a share rising rapidly. He pointed out that leading AI labs’ business models are evolving, with AI computing power output efficiency improving significantly; Anthropic now generates around $50 million in revenue per megawatt of computing power, a figure that could climb to $100 million. This allows OpenAI and Anthropic to purchase or lease computing power at high prices of $25 million to $50 million per megawatt. More notably, Patel forecasts global AI-related capital expenditure will reach approximately $11 trillion between 2024 and 2029, with over $5 trillion needing to be financed via debt. Given AI infrastructure’s potential returns are far higher than those of traditional industries, tech giants may accept higher financing costs, which could push up overall credit rates and squeeze traditional asset valuations and highly indebted economies. Additionally, Patel believes future new computing power may not be primarily used for external model inference services, but will instead flow more toward internal R&D and model self-improvement at AI labs.

According to official announcements, OKX’s Flash Earn Lite will launch the AEON “Stake to Earn” campaign from 15:00 UTC+8 on August 31, 2026 to 15:00 UTC+8 on September 5, 2026. During the event, users who lock BTC, ETH, OKB, or AEON to participate in the subscription will be eligible to share the 4,100,000 AEON airdrop rewards. Notably, ETH borrowed via flexible loans will not count towards valid subscription quotas. Users can pre-subscribe starting immediately, with rewards calculated from the official campaign start. Participation is available via the campaign link or the “Flash Earn” section at the top of the OKX App’s Explore page.

Beating AI News reports that AI agent firm Einsia has launched the SWE Refactor Bench, a benchmark designed to test coding agents’ ability to fully migrate entire codebases independently. The 20 tasks are sourced from real-world projects including SQLite, zlib, libsodium, and GraphHopper, covering scenarios such as C-to-Rust conversion, Maven-to-Gradle migration, and SQLite porting to WASI. Each task grants agents 6 to 30 hours to complete. The evaluation process has three stages: first, verifying that old tech stacks are fully replaced to prevent agents from cutting corners by retaining old code; second, running over 130,000 static checks; finally, six independent coding agents spend one hour each to identify hidden bugs. A total of 520 runs were conducted across 8 leading models and 26 configurations. Of these, 340 successfully completed migration, 88 passed all pre-written test cases, yet 60 of those were found to have hidden bugs in the final stage. Only 28 runs fully passed all stages, resulting in a 5.4% pass rate; 13 out of the 20 tasks were never completed by any agent. Claude Opus 5’s xhigh configuration performed best, fully completing 5 of the 20 tasks. Current coding agents can perform large-scale code modifications, but they still have a long way to go to achieve error-free full-system refactoring.

Per monitoring by on-chain analytics platform Lookonchain, trader 0xD81a—who previously incurred roughly $12 million in cumulative losses from buying Ethereum at high prices and selling at low—has re-entered the market after nearly six months of inactivity. The address spent approximately $5.33 million to purchase 2,165 ETH at an average price of around $2,463. Notably, the trader had previously cut their ETH holdings at an average price of about $2,452 in their last trade, only to repurchase the asset at a higher price now, resulting in a "sell low, buy high" operation.

Beating AI Express News: The identity guessing game around the mysterious model Ox Alpha has clearly tilted toward Zhipu AI. On Polymarket, the probability assigned to Z.ai currently exceeds 90%, while second-place Google holds only around 5%. Ox Alpha is an anonymous inference model that recently emerged on OpenRouter, focusing on programming and long-duration agent tasks. The most concrete clues come from the backend: community developers intentionally sent incorrect parameters to Ox Alpha’s OpenCode interface, and the server returned com.wd.paas.api.domain.v4.chat.ChatCompletionRequest. The "paas/v4/chat" segment in this response matches Z.ai’s official API path exactly. When invalid "role" parameters were sent, Ox Alpha returned error code 1214: Incorrect role information, which aligns with Z.ai’s self-hosted GLM model. When the same GLM weights were hosted on DeepInfra, the error format changed. The model’s own fingerprint also matches. A public forensics test ran over 600 requests, using approximately 13.5 million input tokens, with all 44 tokenizer tests matching the GLM-5 generation.