智东西（公众号：zhidxcom）
作者 | ZeR0
编辑 | 漠影
智东西10月23日报道，近日，在2026亚马逊云科技汽车与制造行业峰会上，亚马逊云科技汽车与制造全球总经理Ozgur Tohumcu分享了AI时代带给制造业的变化，以及AI如何跨过试点，进入规模化生产环境。会后，他与智东西等媒体围绕这些思考进一步交流。
Ozgur Tohumcu在演讲中着重谈道，现在越来越多的概念验证正转入生产环境，这极大改变了我们理解智能体系统的方式。随着智能体思维和相关企业应用原则的引入，这些系统正在走向执行层，不再只是从不同来源汇集数据、展示在仪表盘上，而是开始实际完成任务。
他也很看好物理AI的发展，物理AI在制造业有很多机会，例如机器人视觉引导、开展内部物流、人形机器人应用、质检、系统优化、自研模型等，其中人形机器人广泛应用仍需考虑安全性和经济性。
▲亚马逊云科技汽车与制造全球总经理Ozgur Tohumcu
亚马逊已打造全球最先进的机器人运营体系之一，在其运营网络中部署超过100万台机器人，这些机器人每天会处理数百万件包裹、与数十万名员工协同工作。亚马逊云科技正通过基础模型、仿真与合成数据、模型蒸馏与优化、边缘计算与本地部署、硬件与物理载体五大支柱，来助力物理AI的全链路发展。
一台机器人能搬货、会分拣，已经不稀奇。真正考验企业的，是把这样的能力复制到更多设备、更多工厂，同时让成本算得过来。这正是汽车与制造业面对的一道考题。
按照Ozgur Tohumcu的观察，全球汽车与制造行业有三大趋势：1）自动驾驶经过多年的AI训练与部署，已迈入大规模落地阶段；2）物理AI正深度渗透制造业与汽车行业；3）全球大型企业正在迈向智能体架构和智能体组织，贯穿销售、营销、采购、供应链管理等各环节。
其中，物理AI尤其受关注。它让AI的能力从屏幕走向现实，涉及机器如何感知环境、理解任务并采取行动。但机器人单机能力强，不等于整座工厂就能高效运转。设备之间的通信、机器人与人的协作、可复用的经验，才是规模化应用要解决的问题。
Ozgur Tohumcu认为，比单台设备的学习更有价值的，是机器人群体之间的经验共享，以及这些能力在全球运营网络中的复制。
从这个角度来看，工业AI的竞争，正在延伸到整个系统的协同能力。
“我们与客户的共同实践表明，真正的商业价值从不源于单一环节的突破，而是来自贯穿整条价值链的协同创新。”Ozgur Tohumcu说。
在推进AI规模化落地的过程中，他将真正决定部署速度的关键因素归纳为数据、治理和组织。
其中，数据问题往往最基础，也最难绕开。供应链、生产、销售各有一套系统，车间设备的运行数据与企业管理数据又常常彼此割裂。AI要协助制定生产计划，就需要同时理解订单、物料和产能。只把数据集中存放，还不足以支撑跨部门的业务执行。许多企业正在重新设计数据架构，以尽可能释放AI创造价值的潜力。
治理的要求也在提高。过去，AI提供建议，由人决定是否采用。如今，智能体开始调用工具、推进流程，企业就必须明确它能访问什么、能执行哪些操作，以及每一步如何记录和追溯。
组织同样决定速度。一家工厂验证有效的方案，推广到另一家工厂时，仍要面对不同的流程、管理习惯和人员信任。技术方案可以复制，组织接受它却需要过程。
亚马逊云科技行业集群总经理田锋也谈道，汽车和制造业是具备物理实体的行业，产业链条更复杂，研发环节也与其他行业截然不同，亚马逊云科技发现客户的真实诉求不仅在于技术支持，组织和人才层面的AI素养提升同样至关重要。为此亚马逊云科技推出了Amazon Quick、Kiro等产品，并已积累大量优秀的实践案例。
▲亚马逊云科技行业集群总经理田锋
这些问题，构成了亚马逊云科技的切入点。在客户部署物理AI时，亚马逊云科技致力于提供一站式平台服务：通过Amazon Bedrock提供前沿基础模型，通过Amazon Bedrock AgentCore支持智能体的统筹编排与生命周期管理，并在边缘侧提供底层基础设施与低延迟解决方案。
面向物理AI，还需要衔接仿真、模型蒸馏与优化，以及边缘部署，让模型适应车间对延迟和成本的要求。
选择模型也要根据任务来判断。复杂任务需要更强的能力，高频、明确的任务则要仔细权衡成本。试点阶段可以接受的调用费用，扩大到大量设备或车辆后，可能就会改变项目的商业可行性。
那么，企业会先从哪里获得回报？
Ozgur Tohumcu判断，在制造业和汽车行业，研发环节往往能最早、最快地为客户带来投资回报。从设计迭代、软件开发到仿真验证，AI可以参与多个环节。企业既有机会压缩研发成本，又有机会缩短产品上市周期。
供应链则蕴藏着另一类机会。数据分散在企业内部和供应商系统中，智能体有望帮助企业跨系统获取信息、衔接流程。
相比之下，售后服务的价值更多体现为客户满意度和留存，未必能迅速转化为直接收入。
这也让汽车与制造业的AI落地有了更清晰的次序，从业务问题出发，找到能够衡量收益的环节，再把数据、权限、流程和成本逐项理顺。
走进生产现场后，AI要接受的是更长久的检验，需解决如何持续缩短一次研发迭代、减少一段流程等待、把有效经验复制到更多工厂等现实问题。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Intellectual (WeChat Official Account:zhidxcom)
Author ZeR0
Edit |
On October 23, Ji-in, it was reported that in recent days, at the 2026 Amazon Cloud Science and Technology Industry Summit, Amazon Cloud Science and Technology Managing Director Ozgur Tohumcu shared the changes in manufacturing brought about by the AI era and how AI had crossed the pilot and entered the scale production environment. After the meeting, he shared these thoughts further with the media, including intellectuals.
In his presentation, Ozgur Tohumcu focused on the fact that more and more conceptual validation is now moving into the production environment, which has dramatically changed the way we understand the AI agent system. With the introduction of AI agent thinking and the relevant corporate application principles, these systems are moving to the executive level, no longer simply compiling data from different sources and showing them on dashboards, but beginning to fulfil their tasks.
He also appreciated the development of physical AI, which had many opportunities in manufacturing, such as robot visual guidance, in-house logistics, human robotic applications, quality checks, system optimization, self-study models, among others, where extensive human robotic applications still needed to consider safety and economics.
Ozgur Tohumcu, Managing Director of the Global Automobile and Manufacturing of Amazonian Cloud Technologies
The Amazon has built one of the world ' s most advanced robotics operating systems, deploying more than 1 million robots into its operating network, which handle millions of packages every day and work with hundreds of thousands of employees. Amazon cloud technology is contributing to the full-link development of Physical AI through basic models, simulation and synthesis data, model distillation and optimization, margin computing and local deployment, and hardware and physical carriers.
It's no surprise that a robot can move and sort. The real test for enterprises is to replicate such capabilities to more equipment, more factories and cost. This is exactly the same thing that cars and manufacturing face.
According to Ozgur Tohumcu, there are three major trends in the automobile and manufacturing industry worldwide: (1) auto-driving, after years of AI training and deployment, has reached a large-scale landing phase; (2) physical AI is deeply permeating the manufacturing and automobile industries; and (3) large global enterprises are moving towards AI agent structures and AI agent organizations, cutting across marketing, procurement and supply chain management.
Of these, Physical AI received particular attention. It moves AI ' s capabilities from screen to reality, and it concerns how machines perceive the environment, understand tasks and act. But robotics have a strong single capacity, which does not mean that the entire plant can operate efficiently. Communication between equipment, collaboration between robots and humans, and reusable experience are the issues to be addressed in scale applications.
Ozgur Tohumcu believes that learning is more valuable than learning from single-station equipment is the sharing of experience among robotic groups and the replication of these capabilities in global operating networks.
From this perspective, industrial AI competition is extending to system-wide synergy.
“Our common practice with clients shows that real business values never stem from a single-link breakthrough, but rather from synergistic innovations that cut across the value chain. "Ozguur Tohumcu says.
In moving forward with the scale-down of AI, he summarized the key factors determining the true pace of deployment as data, governance and organization.
Among them, data issues are often the most fundamental and most difficult to circumvent. There are systems for supply chains, production and marketing, and the operating data for workshop equipment and enterprise management data are often fragmented. To assist in developing production plans, AI needs to understand both orders, materials and capacity. The centralization of data does not support cross-sectoral operational implementation. Many enterprises are redesigning the data architecture to maximize the value-creating potential of AI.
Governance requirements are also increasing. In the past, AI has provided recommendations for adoption to be decided upon. Now that AI agent has started to call on tools, advance processes, it is important for the enterprise to know what it can access, what operations it can perform, and how each step can be recorded and traced.
The organization also determines speed. An effective programme of certification at one plant is still subject to different processes, management practices and personnel trust when it is extended to another. Technical programmes can be replicated, and organizations need processes to accept them.
The General Manager of the Amazon Clouds Industry Cluster also said that automobiles and manufacturing were the industries with physical entities, that the industrial chain was more complex and that the R&D chain was different from other industries, and that Amazon Clouds technology found that the true demand of clients was not just technical support, but that the organizational and human-level AI upgrading was also crucial. As a result, Amazon Quick and Kiro products have been launched and a large number of good practice cases have been accumulated.
Tansai, General Manager of the Amazon Industrial Cluster
These problems constitute entry points for Amazon cloud technology. When clients deploy Physical AI, Amazon Cloud Technology is committed to providing all-in-one platform services: providing front-line foundation models through Amazon Bedrock, supporting integrated AI agent organization and life-cycle management through Amazon Bedrock AgentCore, and providing bottom infrastructure and low delay solutions on the edge.
For physical AI, there is also a need to link simulations, model distillation and optimization, and peripheral deployments to adapt models to workshop requirements for delay and cost.
The selection of the model is based on the task. Complex missions require greater capacity, while high-frequency, well-defined missions are carefully weighed. The commercial viability of the project may be changed by extending the cost of mobility that is acceptable during the pilot phase to a large number of equipment or vehicles.
So where do companies get their return first?
Ozgur Tohumcu judges that in manufacturing and automobile industries, R&D tends to yield the earliest and fastest return on investment to clients. From design iterative, software development to simulation validation, AI can participate in multiple links. There are opportunities for enterprises to reduce R & D costs as well as to shorten product listing cycles.
Another type of opportunity lies in the supply chain. Data are scattered within the enterprise and in the vendor system, and AI agent is expected to help enterprises access information and connect processes across the system.
In contrast, the value of after-sales services is more reflected in client satisfaction and retention than in the rapid translation of direct income.
It also allows for a clearer sequence of the downside of the automobile and manufacturing industry, from the point of view of business problems, to find links that can measure the benefits, and then to rationalize data, permission, processes and costs.
After entering the production site, AI is subject to a longer-term test to address such practical issues as the continuous shortening of a research and development overlap, the reduction of a process waiting, and the replication of effective experience to more factories.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.