In today’s edition of Data Points, you’ll learn more about:
- Google’s and Amazon’s huge bets on Anthropic
- Google employees’s resistance to secret government use of AI
- The U.S. threat to punish Chinese companies that distill U.S. AI models
- An agentic system that autonomously designed a working CPU
But first:
OpenAI ends its exclusive agreement with Microsoft
OpenAI revised its partnership with Microsoft, substantially altering a key alliance that has governed the distribution of GPT models, ChatGPT, and other OpenAI products. Microsoft will remain OpenAI’s primary cloud partner through 2032, but its right to OpenAI technology is no longer exclusive. OpenAI will be free to distribute its models on platforms such as Amazon Web Services and Google Cloud. OpenAI will continue to collect 20 percent of Microsoft revenue from its products, but the total is subject to an undisclosed cap and no longer is slated to end if OpenAI achieves artificial general intelligence. The revised terms also remove earlier constraints tied to revenue-sharing and AGI-related provisions. For developers, OpenAI’s ability to deploy its models on a variety of cloud platforms offers greater flexibility. For OpenAI, it gives the company greater access to processing power, broader distribution, and paths to collaborating with other companies and building its own cloud business. (Ars Technica)
DeepSeek V4 falls behind competitors open and closed
DeepSeek released weights for its highly anticipated DeepSeek-V4 large language models, which feature 1 million tokens of context and a revised architecture. The family includes DeepSeek-V4-Pro Preview (1.6 trillion parameters, 49 billion active per token) and DeepSeek-V4-Flash Preview (284 billion parameters, 13 billion active per token). On Artificial Analysis’ Intelligence Index, among open-weights models, DeepSeek-V4-Pro Preview is behind Moonshot Kimi K2.6 and Xiaomi MiMo-V2.5-Pro, which are tied for first place, and tied for second place with Alibaba Qwen3.6 (all models set to their highest reasoning levels). However, it tops other open-weights models on some of the index’s component benchmarks including GDPval-AA, AA-Omniscience Accuracy, and Humanity’s Last Exam. Among closed models, its index performance falls behind leading models from Google, OpenAI, Anthropic, and Meta (all set to their highest reasoning levels). DeepSeek, once the standard bearer among China’s open-weights AI developers, now competes with strong rivals. Its new models are optimized to run on Huawei Ascend chips rather than the typical Nvidia GPUs, illustrating the rapid progress of China’s AI chip makers. (DeepSeek, Inc.)
Google and Amazon deepen their bets on Anthropic with massive funding-for-compute deals
Google and Amazon are both expanding multibillion-dollar investments in Anthropic as competition intensifies for access to advanced models and processing power. Google committed up to $40 billion, starting with $10 billion and adding more based on performance, and will supply large-scale processing power based on its TPU chips. Similarly, Amazon pledged up to $25 billion tied to its Amazon Web Services cloud infrastructure. The deals combine capital investment with long-term cloud and chip agreements, effectively coupling Anthropic’s model development to both Google’s TPU ecosystem and Amazon’s AWS platform, even as all three companies compete in AI services. (The Wall Street Journal) (The New York Times)
Google employees push to block secret government use of AI
More than 600 employees at Google, including some in its DeepMind AI unit, signed a letter that urged Alphabet CEO Sundar Pichai to prohibit the company’s AI systems from being used in classified Pentagon projects. The letter argues that classified deployments would limit the company’s ability to oversee uses of its technology and could enable harmful applications such as mass surveillance or lethal autonomous weapons. The action comes amid reports that Google is negotiating or has entered agreements to provide AI for military applications. Internal opposition highlights tension between AI developers and government in the wake of Anthropic’s February dispute with the U.S. Department of War, when Anthropic refused to provide unrestricted access to its technology and the government declared it a supply-chain risk to security. That conflict is not yet resolved. (The Washington Post)
U.S. aims to curb foreign distillation of domestic models
The Trump administration plans to implement new restrictions to prevent foreign entities, specifically companies in China, from training AI models to mimic the outputs of models built in the United States, a common practice known as knowledge distillation. Chief White House technology advisor Michael Kratsios said the administration intends to work with AI companies to identify efforts to distill U.S.-built models, defend against them, and punish foreign companies that attempt them. Chinese officials responded by calling for cooperation and accusing the U.S. government of seeking to suppress China’s technological development. Banning foreign entities from distilling knowledge from U.S. models would limit open development and potentially complicate international research collaborations. (The Associated Press)
AI autonomously designs functional CPU chip
An agentic AI system autonomously designed a 1.48 gigahertz RISC-V CPU chip, roughly equivalent to a 2011-vintage Intel Celeron SU2300, from a 219-word specification, according to a paper by researchers at the AI chip design startup Verkor. The resulting design has not been physically fabricated, but the authors verified it in simulation. The system took 12 hours to generate the design, in contrast to the typical timeline of 18 to 36 months. However, it consumed tens of billions of tokens and a team of up to 10 human experts likely would be required to implement the design. For AI developers, this demonstrates the potential for agentic workflows to compress hardware development timelines from months to hours, though significant hurdles remain. (Tom’s Hardware)
Want to know more about what matters in AI right now?
Read the latest issue of The Batch for in-depth analysis of news and research.
Last week, Andrew Ng talked about how coding agents accelerate different software-development tasks to varying degrees, with the most impact on frontend development, followed by backend development, infrastructure, and research, and how this understanding helps in organizing effective software teams.
“Research involves thinking through new ideas, formulating hypotheses, running experiments, interpreting them to potentially modify the hypotheses, and iterating until we reach conclusions. Coding agents can speed up the pace at which we can write research code. (I also use coding agents to help me orchestrate and keep track of experiments, which makes it easier for a single researcher to manage more experiments.) But there is a lot of work in research other than coding, and today’s agents help with research only marginally.”
Read Andrew’s letter here.
Other top AI news and research stories covered in depth:
- GLM 5.1 Aims for Long-Running Tasks as Z.ai’s latest model can evaluate its interim results and change its approach hundreds of times before delivering final output.
- Humanoid Robots Work Factory Floors, where Agility Digits robots fetch and carry bins at a Schaeffler auto-parts factory, displacing some humans into higher-level jobs.
- Anti-Data Center Revolt Gains Traction describes public opposition to new data centers in the U.S. spurs political action and even violence.
- Assistants That Assisted Consistently explains new research that stabilizes the personas of large language models when they start to drift from helpful to harmful.
A special offer for our community
DeepLearning.AI recently launched the first-ever subscription plan for our entire course catalog! As a Pro Member, you’ll immediately enjoy access to:
- Over 150 AI courses and specializations from Andrew Ng and industry experts
- Labs and quizzes to test your knowledge
- Projects to share with employers
- Certificates to testify to your new skills
- A community to help you advance at the speed of AI
Enroll now to lock in a year of full access for $25 per month paid upfront, or opt for month-to-month payments at just $30 per month. Both payment options begin with a one-week free trial. Explore Pro’s benefits and start building today!
Data Points is produced by human editors with AI assistance.