Author: RadarAI Editorial
Editor: RadarAI Editorial
Last updated: 2026-10-20
Review status: Editorial review pending
Brief
速报
官方
AI动态
开源
The mismatch between AI capital expenditure and revenue growth has drawn market attention. Bank of America research shows AI capex this year beat expectations by 50% while revenue exceeded by only 20% [11], yet Morgan Stanley still forecasts that hyperscaler and neocloud AI spending will reach $1.58 trillion by 2027 [12]. On the technology side, ModelBest has for the first time achieved real-time, lossless on-device operation of a 35B-A3B-scale MoE on smartphones [10], while Marvell's photonic module allows HBM and GPUs to be separated by up to ...
Editorial standards and source policy: Editorial standards, Team. Content links to primary sources; see Methodology.
## 🔍 Core Insights
**The mismatch between AI capital expenditure** and revenue growth has drawn market attention. Bank of America research shows AI capex this year beat expectations by **50%** while revenue exceeded by only **20%** [11], yet Morgan Stanley still forecasts that hyperscaler and neocloud AI spending will reach **$1.58 trillion** by 2027 [12]. On the technology side, **ModelBest** has for the first time achieved real-time, lossless on-device operation of a **35B-A3B**-scale MoE on smartphones [10], while **Marvell**'s photonic module allows HBM and GPUs to be separated by up to **50 meters** [5]—breakthroughs in both on-device computing and interconnect architecture are advancing in parallel.
## 🚀 Key Developments
- **MemT 1.0 bets on memory-native design** [0]: Open-sourced with 340 billion tokens of training, it achieves a closed loop of long-cycle interaction experience, improving task success rates and reasoning efficiency.
- **OpenAI launches ChatGPT Slides** [9]: Supports generating presentations from conversations or files, with real-time collaborative editing and commenting for teams.
- **Kling releases Kling 4.0** [23]: Single videos up to 30 seconds, prompt understanding expanded to 8,000 tokens, supporting dual-channel audio and multilingual lip sync.
- **ModelBest's LES technology lands on-device** [10]: 35B-A3B MoE measured at over 20 tokens/s decoding on iPhone 18 Pro Max.
- **Marvell's photonic module breaks the memory wall** [5]: HBM can be up to 50 meters from GPUs with latency maintained at hundreds of nanoseconds, allowing AI clusters to share large-capacity memory.
- **Anthropic in talks for a second $45 billion compute deal** [14]: Reserved capacity at the West Virginia campus will double to 920 MW.
- **Qwen APP launches "Qwen Shopping Companion"** [3]: Screen-sharing voice shopping assistance supporting product review summarization, real-time price comparison and verification, and chat-based advice.
- **MiniCPM-V 4.7 gains vLLM v0.32.0 support** [7]: High-throughput, low-latency deployment becomes more convenient.
## 🔗 Sources
[0] Betting on memory-native design, MemT 1.0 breaks through in the second half of the nurturing-AI era — https://www.bestblogs.dev/article/0367f79510?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
[3] Qwen APP launches "Qwen Shopping Companion" screen-sharing voice shopping feature — https://aihot.news/items/tyfzlxplgvtt1i4ulajvuogho
[5] Marvell photonic module separates HBM from GPU by 50 meters — https://aihot.news/items/ptn8busvkfb6rp94vjau0bdt2
[7] MiniCPM-V 4.7 gains vLLM v0.32.0 support — https://aihot.news/items/fgx3ep8ja72k19wdagm8k4ep6
[9] OpenAI launches ChatGPT Slides presentation tool with real-time team collaboration — https://aihot.news/items/kjxkxat9nca4jdkyp49aevbr0
[10] ModelBest explains how forward-looking expert scheduling LES enables 35B MoE to run real-time and losslessly on smartphones — https://aihot.news/items/m2ylwtzb8x49c32spfl1wplqz
[11] Bank of America research says AI capex beats expectations
The mismatch between AI capital expenditure and revenue growth has drawn market attention. Bank of America research shows AI capex this year beat expectations by 50% while revenue exceeded by only 20% [11], yet Morgan Stanley still forecasts that hyperscaler and neocloud AI spending will reach $1.58 trillion by 2027 [12]. On the technology side, ModelBest has for the first time achieved real-time, lossless on-device operation of a 35B-A3B-scale MoE on smartphones [10], while Marvell's photonic module allows HBM and GPUs to be separated by up to 50 meters [5]—breakthroughs in both on-device computing and interconnect architecture are advancing in parallel.
🚀 Key Developments
- MemT 1.0 bets on memory-native design [0]: Open-sourced with 340 billion tokens of training, it achieves a closed loop of long-cycle interaction experience, improving task success rates and reasoning efficiency.
- OpenAI launches ChatGPT Slides [9]: Supports generating presentations from conversations or files, with real-time collaborative editing and commenting for teams.
- Kling releases Kling 4.0 [23]: Single videos up to 30 seconds, prompt understanding expanded to 8,000 tokens, supporting dual-channel audio and multilingual lip sync.
- ModelBest's LES technology lands on-device [10]: 35B-A3B MoE measured at over 20 tokens/s decoding on iPhone 18 Pro Max.
- Marvell's photonic module breaks the memory wall [5]: HBM can be up to 50 meters from GPUs with latency maintained at hundreds of nanoseconds, allowing AI clusters to share large-capacity memory.
- Anthropic in talks for a second $45 billion compute deal [14]: Reserved capacity at the West Virginia campus will double to 920 MW.
- Qwen APP launches "Qwen Shopping Companion" [3]: Screen-sharing voice shopping assistance supporting product review summarization, real-time price comparison and verification, and chat-based advice.
- MiniCPM-V 4.7 gains vLLM v0.32.0 support [7]: High-throughput, low-latency deployment becomes more convenient.
🔗 Sources
[0] Betting on memory-native design, MemT 1.0 breaks through in the second half of the nurturing-AI era — https://www.bestblogs.dev/article/0367f79510?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
[3] Qwen APP launches "Qwen Shopping Companion" screen-sharing voice shopping feature — https://aihot.news/items/tyfzlxplgvtt1i4ulajvuogho
[5] Marvell photonic module separates HBM from GPU by 50 meters — https://aihot.news/items/ptn8busvkfb6rp94vjau0bdt2
[7] MiniCPM-V 4.7 gains vLLM v0.32.0 support — https://aihot.news/items/fgx3ep8ja72k19wdagm8k4ep6
[9] OpenAI launches ChatGPT Slides presentation tool with real-time team collaboration — https://aihot.news/items/kjxkxat9nca4jdkyp49aevbr0
[10] ModelBest explains how forward-looking expert scheduling LES enables 35B MoE to run real-time and losslessly on smartphones — https://aihot.news/items/m2ylwtzb8x49c32spfl1wplqz
[11] Bank of America research says AI capex beats expectations
← Back to Updates