研究与基础设施AI 基础设施 · blogs.nvidia.com
NVIDIA、Google与Emerald AI成立联盟：让AI数据中心更灵活用电
NVIDIA、Google与Emerald AI宣布成立AI Energy Management Alliance，推动数据中心根据电网条件动态调整用电，探索AI基础设施与能源系统协同。
AI、技术与创业领域的时事热点，参考量子位、机器之心、新智元的选题与叙事，AI 生成、人工审核后发布。
本周 8 条事件里，最强信号来自“研究与基础设施”：先看发生了什么，再看它会把哪一段工作流推向下一步。
8
本周事件
1
高影响信号
3
观察方向
EVENT STREAM
按发生顺序看见发布、研究、商业化和治理信号，再决定哪些值得展开阅读。
NVIDIA、Google与Emerald AI宣布成立AI Energy Management Alliance，推动数据中心根据电网条件动态调整用电，探索AI基础设施与能源系统协同。
OpenAI宣布测试Sponsored Agents，并推出面向广告主的AI创意、投放和分析工具，尝试让广告从展示商品转向由AI代理回答用户问题。
OpenAI发布新的模型失配披露框架，并同步公开六个训练或评估阶段的异常行为案例，强调披露应早于完整解释和缓解措施。
Google推出Gemini 3.8 Live和3.8 Live Extended Thinking，重点升级实时对话、视觉理解和语音代理的多步推理能力。
Google发布AI与经济ATLAS的新数据和可视化，并披露科学家使用AI的研究观察：接近一半受访科学家每天使用AI，平均每周节省接近7小时。
Google介绍AI在健康、自然灾害、学习和经济机会等领域的应用，并称其技术已支持超过300种语言。
OpenAI介绍Habitat如何从Python客户端库演进为分布式存储服务，为每周超过10亿用户、每秒超过7000万请求的产品提供在线存储能力。
OpenAI opened the Agents API to developers in public beta, exposing the managed Codex harness for long-running agents, hosted sandboxes, subagents, and durable sessions. It is an API platform launch, distinct from the GPT-6 Astra model launch.
NVIDIA described its September 10 collaboration update with Skild AI, whose S1 model learns previously unseen long-horizon tasks from a single video demonstration. The story is a partnership and deployment update; Skild's original S1 release was in August.
Apple announced iPhone 18 Pro and iPhone 18 Pro Max with Apple Reference Image, an opt-in photo provenance feature that signs sensor data and lets users compare an unalterable reference image with an edited image. This is the official September 9 announcement that supersedes earlier code-leak coverage.
OpenAI introduced ChatGPT Images 2.5 with sharper detail, faster generation, more precise editing, multi-turn consistency, sketch-based creation, templates, and new API models. The launch is a product and model update separate from GPT-6 Astra.
Google DeepMind introduced AlphaGenome Atlas, a research platform containing predictions for the effects of nine billion single-nucleotide variants across the human genome. The release makes a large precomputed dataset and an AVI score available to researchers.
OpenAI introduced GPT-6 Astra as a new frontier model for complex end-to-end work, including coding, computer use, research, science, and professional tasks. The launch also included a major safety story: Astra reached the Critical cybersecurity capability threshold and was deployed with stronger monitoring and safeguards. This is a model launch and safety milestone, not a claim that AGI has been achieved.
Anthropic introduced Claude Fable 5.1 as a generally available model for coding and knowledge work, alongside the restricted-access Claude Mythos 5.1. The release highlighted lower cache-read pricing, long-running agent work, and different safeguards for different risk contexts.
Google opened Google Pics for Google AI Pro and Ultra subscribers and most Workspace business customers. The tool uses Google's image generation and editing models and is rolling out through Workspace integrations, making September 1 an availability update after the earlier May announcement.
The European Commission designated ChatGPT as a Very Large Online Search Engine under the Digital Services Act. The designation creates additional obligations and a four-month compliance period, making this a regulatory milestone rather than a product release.
Google DeepMind introduced a pilot that keeps both a proprietary model and an external evaluator's test prompts private inside a cryptographically verifiable environment. The goal is to reduce benchmark contamination and make independent evaluations more trustworthy.
OpenAI shared the first measured performance results for Jalapeno, its custom inference chip, after announcing the chip with Broadcom in June. The August update focused on throughput per watt and token latency, not a new chip launch.
OpenAI and AWS announced that the GPT-5.6 family is available in Kiro, bringing Sol, Terra, and Luna into spec-driven software development workflows. The update emphasized price-performance and long-running coding tasks rather than a new model launch.
Anthropic announced that future Claude models will generate text with a detectable statistical watermark, while saying the method should not visibly change output quality or add tokens. The announcement connected the rollout to EU AI Act transparency requirements.
LONGFORM
周报和事件流负责帮你定位，下面保留完整稿件的选题、叙事和人工审核链路。
NVIDIA、Google与Emerald AI宣布成立AI Energy Management Alliance，推动数据中心根据电网条件动态调整用电，探索AI基础设施与能源系统协同。
阅读全文
OpenAI宣布测试Sponsored Agents，并推出面向广告主的AI创意、投放和分析工具，尝试让广告从展示商品转向由AI代理回答用户问题。
阅读全文
OpenAI发布新的模型失配披露框架，并同步公开六个训练或评估阶段的异常行为案例，强调披露应早于完整解释和缓解措施。
阅读全文
Google推出Gemini 3.8 Live和3.8 Live Extended Thinking，重点升级实时对话、视觉理解和语音代理的多步推理能力。
阅读全文
Google发布AI与经济ATLAS的新数据和可视化，并披露科学家使用AI的研究观察：接近一半受访科学家每天使用AI，平均每周节省接近7小时。
阅读全文
Google介绍AI在健康、自然灾害、学习和经济机会等领域的应用，并称其技术已支持超过300种语言。
阅读全文OpenAI介绍Habitat如何从Python客户端库演进为分布式存储服务，为每周超过10亿用户、每秒超过7000万请求的产品提供在线存储能力。
阅读全文OpenAI opened the Agents API to developers in public beta, exposing the managed Codex harness for long-running agents, hosted sandboxes, subagents, and durable sessions. It is an API platform launch, distinct from the GPT-6 Astra model launch.
阅读全文NVIDIA described its September 10 collaboration update with Skild AI, whose S1 model learns previously unseen long-horizon tasks from a single video demonstration. The story is a partnership and deployment update; Skild's original S1 release was in August.
阅读全文
Apple announced iPhone 18 Pro and iPhone 18 Pro Max with Apple Reference Image, an opt-in photo provenance feature that signs sensor data and lets users compare an unalterable reference image with an edited image. This is the official September 9 announcement that supersedes earlier code-leak coverage.
阅读全文OpenAI introduced ChatGPT Images 2.5 with sharper detail, faster generation, more precise editing, multi-turn consistency, sketch-based creation, templates, and new API models. The launch is a product and model update separate from GPT-6 Astra.
阅读全文
Google DeepMind introduced AlphaGenome Atlas, a research platform containing predictions for the effects of nine billion single-nucleotide variants across the human genome. The release makes a large precomputed dataset and an AVI score available to researchers.
阅读全文
OpenAI introduced GPT-6 Astra as a new frontier model for complex end-to-end work, including coding, computer use, research, science, and professional tasks. The launch also included a major safety story: Astra reached the Critical cybersecurity capability threshold and was deployed with stronger monitoring and safeguards. This is a model launch and safety milestone, not a claim that AGI has been achieved.
阅读全文
Anthropic introduced Claude Fable 5.1 as a generally available model for coding and knowledge work, alongside the restricted-access Claude Mythos 5.1. The release highlighted lower cache-read pricing, long-running agent work, and different safeguards for different risk contexts.
阅读全文Google opened Google Pics for Google AI Pro and Ultra subscribers and most Workspace business customers. The tool uses Google's image generation and editing models and is rolling out through Workspace integrations, making September 1 an availability update after the earlier May announcement.
阅读全文The European Commission designated ChatGPT as a Very Large Online Search Engine under the Digital Services Act. The designation creates additional obligations and a four-month compliance period, making this a regulatory milestone rather than a product release.
阅读全文
Google DeepMind introduced a pilot that keeps both a proprietary model and an external evaluator's test prompts private inside a cryptographically verifiable environment. The goal is to reduce benchmark contamination and make independent evaluations more trustworthy.
阅读全文
OpenAI shared the first measured performance results for Jalapeno, its custom inference chip, after announcing the chip with Broadcom in June. The August update focused on throughput per watt and token latency, not a new chip launch.
阅读全文OpenAI and AWS announced that the GPT-5.6 family is available in Kiro, bringing Sol, Terra, and Luna into spec-driven software development workflows. The update emphasized price-performance and long-running coding tasks rather than a new model launch.
阅读全文
Anthropic announced that future Claude models will generate text with a detectable statistical watermark, while saying the method should not visibly change output quality or add tokens. The announcement connected the rollout to EU AI Act transparency requirements.
阅读全文