内容来源：https://techcrunch.com/2025/12/03/all-the-biggest-news-from-aws-big-tech-show-reinvent-2025/

内容总结：

亚马逊云科技年度技术大会re:Invent 2025于12月2日在美国拉斯维加斯开幕。在为期数天的议程中，人工智能成为贯穿始终的核心议题，尤其聚焦于能自主执行任务的“AI智能体”技术。

大会首日，亚马逊云科技CEO马特·加曼在主题演讲中强调，AI智能体正从辅助工具转向能够代表用户执行任务、实现自动化的生产力，这将为企业带来真正的投资回报。次日，亚马逊云科技Agentic AI副总裁斯瓦米·西瓦苏布拉马尼安进一步阐释，智能体能够理解自然语言指令、生成计划、编写代码并执行完整解决方案，从而大幅加速从构想到落地的过程。

本次大会发布了一系列围绕AI智能体的新产品与升级：

1. 推出新型“前沿智能体”：包括能编写代码、学习团队工作模式并自主运行数日的“Kiro自主智能体”，以及专注于代码安全审查和运维任务的两款智能体，目前均已开放预览。
2. 增强智能体开发平台：AgentCore平台新增策略管理功能，允许开发者为智能体设置行为边界，并新增用户行为记忆与日志记录能力，同时提供13种预置评估方案以帮助客户测试智能体性能。
3. 升级大模型开发工具：亚马逊云科技为其AI模型服务Amazon Bedrock和机器学习平台Amazon SageMaker推出新功能，包括SageMaker的无服务器模型定制，以及Bedrock的强化学习微调，旨在降低企业构建专属大模型的门槛。
4. 发布新一代自研AI芯片：推出训练芯片Trainium3及搭载该芯片的UltraServer系统。据称其训练和推理性能最高可达上一代的4倍，能耗降低40%。公司同时透露，正在开发能与英伟达芯片协同工作的下一代芯片Trainium4。
5. 扩展Nova模型家族：新增四款AI模型，其中三款专注于文本生成，一款支持图文生成。同时推出新服务Nova Forge，让企业能基于预训练模型，使用自有数据快速定制专属模型。
6. 推出“AI工厂”解决方案：与英伟达合作推出可部署于企业自有数据中心的AI系统，既支持英伟达GPU，也可选用Trainium3芯片，以满足政府和大型企业对数据主权与隐私的严格要求。

此外，亚马逊云科技也宣布了多项成本优化措施，包括数据库节省计划，承诺用量下年度成本最高可降35%；并向符合条件的早期初创企业免费提供一年的Kiro Pro+编程智能体服务额度。

客户案例成为展示技术成效的重要部分。网约车公司Lyft分享称，其通过Amazon Bedrock调用Anthropic的Claude模型构建的客服智能体，已将问题平均解决时间缩短87%，司机使用率今年提升70%。

亚马逊CEO安迪·贾西在社交媒体上透露，当前代次的自研AI芯片Trainium2已带来可观收入，并对新一代芯片的商业前景表示乐观。

本届re:Invent大会将持续至12月5日，预计将有更多产品动态发布。

中文翻译：

亚马逊云科技年度技术大会AWS re:Invent在又一轮产品发布、主题演讲与例行的客户成功案例分享中落下帷幕。贯穿始终的核心议题毫无悬念地聚焦于企业级人工智能。今年的升级重点在于赋予客户更强的定制化AI智能体控制能力，其中一款被亚马逊宣称能够自主学习并持续独立工作数日。

本届持续至12月5日的AWS re:Invent 2025以亚马逊云科技首席执行官马特·加曼的主题演讲拉开序幕。他强调AI智能体将释放人工智能的"真正价值"。"AI助手正逐渐让位于能代您执行任务并实现自动化的AI智能体，"他在12月2日的演讲中指出，"这正是企业AI投资开始获得实质性商业回报的关键转折点。"

12月3日的大会继续深化AI智能体主题，并深入探讨客户案例。亚马逊云科技智能体AI副总裁斯瓦米·西瓦苏布拉马尼安在主题演讲中展现出超越乐观的坚定信念。"我们正身处伟大变革的时代，"他表示，"这是人类历史上首次能够用自然语言描述目标，由智能体生成执行方案。它们编写代码、调用工具并实施完整解决方案，赋予开发者无限制的创造自由，极大加速从构想到产生重大影响的进程。"

虽然AI智能体相关消息贯穿AWS re:Invent 2025全程，大会还发布了其他重要公告。以下是我们关注的亮点汇总。TechCrunch将持续更新本文内容，最新动态将置顶显示，敬请持续关注。

强化大语言模型布局
亚马逊云科技宣布为企业客户提供更多自建模型工具，重点增强Amazon Bedrock和Amazon SageMaker AI的定制化大语言模型构建能力。例如在SageMaker中推出无服务器模型定制功能，开发者无需考虑计算资源即可开始建模，既可通过自主引导路径，也可通过AI智能体指令触发该功能。同时Bedrock平台新增强化学习微调功能，开发者可选择预设工作流或奖励机制，由平台自动完成全流程定制。

加入Disrupt 2026候补名单
立即登记Disrupt 2026候补名单，抢占早鸟票优先购买权。往届大会曾汇聚谷歌云、Netflix、微软、Box、Phia、a16z、ElevenLabs、Wayve、Hugging Face、埃拉德·吉尔与维诺德·科斯拉等250多位行业领袖，呈现200多场助力企业成长与竞争力提升的深度会议。更有数百家跨领域创新初创企业等待您的对接。

安迪·贾西披露关键数据
亚马逊首席执行官安迪·贾西在社交媒体平台X上阐释了马特·加曼的主题演讲，透露其英伟达竞品AI芯片Trainium2当前代次已创造可观收益。这番表态配合新一代Trainium3芯片的发布，预示该产品线将持续带来营收增长。

数据库成本优化方案落地
在众多公告中，一项已引发欢呼的更新是折扣计划。亚马逊云科技推出数据库节省计划，客户承诺一年期固定用量（美元/小时）最高可降低35%数据库成本。该优惠将自动适用于支持数据库服务的合格用量，超出承诺部分按需计费。Duckbill首席云经济学家科里·奎因在博客中精辟总结："长达六年的建议终见成效。"

免费策略抢占开发者心智
面对AI编程工具市场竞争，亚马逊试图通过全年免费额度推广其产品Kiro。本月底前申请的合格初创企业将获赠Kiro Pro+额度，但该优惠仅限特定国家的早期初创公司。

新一代AI训练芯片与英伟达兼容性
亚马逊云科技推出升级版AI训练芯片Trainium3及配套UltraServer AI系统。这款芯片在提升4倍AI训练推理性能的同时降低40%能耗。公司还预告正在研发的Trainium4将具备英伟达芯片协同工作能力。

扩展AgentCore功能
AgentCore AI智能体开发平台新增多项功能，其中策略管理功能让开发者更便捷设定AI智能体边界。平台还新增用户行为日志记忆功能，并推出13种预置评估系统帮助客户测试智能体。

永续型AI智能体工作单元
大会发布三款名为"前沿智能体"的新型AI智能体，包括可编写代码的"Kiro自主智能体"，它能学习团队工作模式并持续独立运行数小时至数日。另两款分别负责代码审查等安全流程与防止新代码上线事故的DevOps任务，预览版现已开放。

Nova模型家族扩容
Nova AI模型家族新增四款模型，其中三款专注文本生成，一款支持文图生成。同步推出的Nova Forge服务让云客户能获取预训练、半训练或后训练模型，并基于自有数据完成最终训练，强调灵活性与定制化优势。

Lyft的AI智能体实践案例
多家亚马逊云科技客户在大会分享成功实践，网约车平台Lyft通过Amazon Bedrock调用Anthropic的Claude模型创建AI智能体，处理司机与乘客咨询问题。该公司表示该智能体使平均解决时间缩短87%，今年司机使用率提升70%。

私有数据中心AI工厂方案
亚马逊推出"AI工厂"解决方案，允许大型企业与政府在自有数据中心部署AWS AI系统。这套与英伟达合作设计的系统兼容双方技术，客户既可配置英伟达GPU，也可选用亚马逊自研Trainium3芯片。该方案旨在满足数据主权需求，帮助政府和企业实现数据可控的AI部署。

赞助内容：观看re:Invent行业流媒体
直击拉斯维加斯亚马逊云科技旗舰大会最新动态，涵盖智能体AI、云基础设施、安全等全方位前沿发布。本视频由亚马逊云科技合作呈现。

英文来源：

Amazon Web Services’ annual tech conference AWS re:Invent has wrapped up another day with a deluge of product news and keynotes — plus the obligatory customer success stories.
The unsurprising theme is AI for the enterprise. This year it’s all about upgrades that give customers greater control to customize AI agents, including one that AWS claims can learn from you and then work independently for days.
AWS re:Invent 2025, which runs through December 5, started with a keynote from AWS CEO Matt Garman, who leaned into the idea that AI agents can unlock the “true value” of AI.
“AI assistants are starting to give way to AI agents that can perform tasks and automate on your behalf,” he said during the December 2 keynote. “This is where we’re starting to see material business returns from your AI investments.”
On December 3, the conference pressed on with its AI agents messaging, as well as deeper dives into customer stories. Swami Sivasubramanian, vice president of Agentic AI at AWS, gave one of the keynote talks. To say he was bullish is perhaps understating the vibe.
“We are living in times of great change,” Sivasubramanian said during the talk. “For the first time in history, we can describe what we want to accomplish in natural language, and agents generate the plan. They write the code, call the necessary tools, and execute the complete solution. Agents give you the freedom to build without limits, accelerating how quickly you can go from idea to impact in a big way.”
While AI agent news promises to be a persistent presence throughout AWS re:Invent 2025, there were other announcements, too. Here is a roundup of the ones that got our attention. TechCrunch will update this article, with the newest insights at the top, through the end of AWS re:Invent. Be sure to check back.
Doubling down on LLMs
AWS announced more tools for enterprise customers to create their own models. Specifically, AWS said it is adding new capabilities for both Amazon Bedrock and Amazon SageMaker AI to make building custom LLMs easier.
For instance, AWS is bringing serverless model customization to SageMaker, which allows developers to start building a model without needing to think about compute resources or infrastructure. The serverless model customization can be accessed through either a self-guided path or by prompting an AI agent.
AWS also announced Reinforcement Fine Tuning in Bedrock, which allows developers to choose a preset workflow or reward system and have Bedrock run their customization process automatically from start to finish.
Join the Disrupt 2026 Waitlist
Add yourself to the Disrupt 2026 waitlist to be first in line when Early Bird tickets drop. Past Disrupts have brought Google Cloud, Netflix, Microsoft, Box, Phia, a16z, ElevenLabs, Wayve, Hugging Face, Elad Gil, and Vinod Khosla to the stages — part of 250+ industry leaders driving 200+ sessions built to fuel your growth and sharpen your edge. Plus, meet the hundreds of startups innovating across every sector.
Join the Disrupt 2026 Waitlist
Add yourself to the Disrupt 2026 waitlist to be first in line when Early Bird tickets drop. Past Disrupts have brought Google Cloud, Netflix, Microsoft, Box, Phia, a16z, ElevenLabs, Wayve, Hugging Face, Elad Gil, and Vinod Khosla to the stages — part of 250+ industry leaders driving 200+ sessions built to fuel your growth and sharpen your edge. Plus, meet the hundreds of startups innovating across every sector.
Andy Jassy shares some numbers
Amazon CEO Andy Jassy took to social media platform X to expound on AWS chief Matt Garman’s keynote speech. The message: The current generation of its Nvidia-competitor AI chip Trainium2 is already bringing in loads of cash.
His comments were tied to the reveal of its next-generation chip, Trainium3, and meant to forecast a promising revenue future for the product.
Database savings arrives
Tucked among the dozens of announcements is one item that is already getting cheers: Discounts.
Specifically, AWS said it was launching Database Savings Plans, which help customers reduce database costs by up to 35% when they commit to a consistent amount of usage ($/hour) over a one-year term. The company said the savings will automatically apply each hour to eligible usage across supported database services, and any additional usage beyond the commitment is billed at on-demand rates.
Corey Quinn, chief cloud economist at Duckbill, summed it up well in his blog post, “Six years of complaining finally pays off.”
Can’t get a better deal than free, Amazon hopes
Is there any way for another AI coding tool to win the hearts of startup founders? Amazon hopes a year’s worth of credits, for free, will do the trick for its offering, Kiro. The company will be giving away credits to Kiro Pro+ to qualified startups that apply for the deal before the end of the month. However, only early-stage startups in certain countries are eligible.
An AI training chip and Nvidia compatibility
AWS introduced a new version of its AI training chip called Trainium3 along with an AI system called UltraServer that runs it. The TL;DR: This upgraded chip comes with some impressive specs, including a promise of up to 4x performance gains for both AI training and inference while lowering energy use by 40%.
AWS also provided a teaser. The company already has Trainium4 in development, which will be able to work with Nvidia’s chips.
Expanded AgentCore capabilities
AWS announced new features in its AgentCore AI agent building platform. One feature of note is Policy in AgentCore, which gives developers the ability to more easily set boundaries for AI agents.
AWS also announced that agents will now be able to log and remember things about their users. Plus it announced that it will help its customers evaluate agents through 13 prebuilt evaluation systems.
A nonstop AI agent worker bee
AWS announced three new AI agents (there is that term again) called “Frontier agents,” including one called “Kiro autonomous agent” that writes code and is designed to learn how a team likes to work so it can operate largely on its own for hours or days.
Another of these new agents handles security processes like code reviews, and the third does DevOps tasks such as preventing incidents when pushing new code live. Preview versions of the agents are available now.
New Nova models and services
AWS is rolling out four new AI models within its Nova AI model family — three of which are text generating and one that can create text and images.
The company also announced a new service called Nova Forge that allows AWS cloud customers to access pre-trained, mid-trained, or post-trained models that they can then top off by training on their own proprietary data. AWS’s big pitch is flexibility and customization.
Lyft’s argument for AI agents
The ride-hailing company was among many AWS customers that piped up during the event to share their success stories and evidence of how products affected their business. Lyft is using Anthropic’s Claude model via Amazon Bedrock to create an AI agent that handles driver and rider questions and issues.
The company said this AI agent has reduced average resolution time by 87%. Lyft also said it has seen a 70% increase in driver usage of the AI agent this year.
An AI Factory for the private data center
Amazon also announced “AI Factories” that allow big corporations and governments to run AWS AI systems in their own data centers.
The system was designed in partnership with Nvidia and includes both Nvidia’s tech and AWS’s. While companies that use it can stock it with Nvidia GPUs, they can also opt for Amazon’s newest homegrown AI chip, the Trainium3. The system is Amazon’s way of addressing data sovereignty, or the need of governments and many companies to control their data and not share it, even to use AI.
Sponsored: Watch re:Invent industry streams
Check out the latest reveals on everything from agentic AI and cloud infrastructure to security and much more from the flagship Amazon Web Services event in Las Vegas. This video is brought to you in partnership with AWS.