Daily AI RoundupsSunday, 16 August 2026

Anthropic is closing in on a roughly $7 billion acquisition of Israeli AI startup Decart, beating a higher offer from Nvidia ahead of its expected IPO, while Alibaba's open-weight Qwen models passed 3 billion global downloads in six months to eclipse Meta and Google. Dario Amodei argues that regulation and distributed power are a false choice for frontier AI, and Nvidia made news on three fronts: a potential $3 billion SB Energy investment for OpenAI's Ohio data center, its ambitious open-source Nemotron 4, and a 50% jump in one-year H100 rental prices as large clusters face 12- to 18-month waits. The roundup also covers Kimi Work's silent feedback-report attachments, llama.cpp's Kimi-K3 support, an NVFP4 Blackwell kernel that beats cuBLAS, and readings from Tim O'Reilly, Exponential View, Sebastian Raschka, and a token-economics essay on who buys AI's output.

• AI M&A
• open-weight models
• AI infrastructure financing
• AI governance
• AI security
• AI economics
• open-source AI

Stories in this roundup

1. 01
   Anthropic is closing in on a ~$7 billion Decart acquisition, winning the Israeli AI startup despite a higher offer from Nvidia, Calcalist reportsCalcalist
2. 02
   Alibaba's open-weight Qwen models pass 3 billion downloads in six months, eclipsing Meta and Google, Bloomberg reportsBloomberg
3. 03
   Dario Amodei argues regulation and distributed power are a false choice for frontier AIDario Amodei / X
4. 04
   Nvidia is in talks to invest as much as $3 billion in SB Energy for OpenAI's Ohio data center, The Information reportsThe Information
5. 05
   Nvidia is building Nemotron 4, its most ambitious open-source model, betting stronger open AI drives chip demandThe Information
6. 06
   The cost of renting Nvidia H100 GPUs on one-year contracts jumped 50% in six months, with 12- to 18-month waits for large clusters, The Information reportsThe Information
7. 07
   A reverse engineer finds Kimi Work's desktop app attaches five recent sessions to feedback reports without noticeRuntimeWire
8. 08
   llama.cpp adds Kimi-K3 text model support, covering its hybrid KDA and MLA architectureGitHub / llama.cpp
9. 09
   Pranjal Shankhdhar's Blackwell kernel outperforms cuBLAS on NVFP4, following his earlier H100 writeupPranjal Shankhdhar / Substack
10. 10
    Tim O'Reilly argues composability, not just open weights, will decide who controls the AI stackTim O'Reilly / Substack
11. 11
    Exponential View analyzes the economics of a $6 AI agentExponential View
12. 12
    Sebastian Raschka builds an AI text detector from scratch, from dataset construction to local deploymentSebastian Raschka / Substack
13. 13
    An essay betting that the models work asks who will buy all the tokens they produceHorse Energy