“企业把AI方案推向规模时，拖慢速度的通常不是技术。”亚马逊云科技汽车与制造全球总经理 Ozgur Tohumcu 给出了一个不太常见的判断。
过去一两个月，大模型行业甚至出现了要不要放慢前沿模型研发节奏的讨论，模型能力本身已经足够超前，平台服务也足够先进。真正卡住企业的是另外的要素，有没有撑得住AI服务扩张的数据基座，有没有为自主执行的智能体准备好护栏、可观测性与可追溯机制，以及组织是否还容得下试验。
Ozgur Tohumcu举例表示，大量车企的AI项目停留在概念验证阶段，没有真正走向量产与工业化落地，制造企业已经证明AI行得通，剩下的难题是怎么让它铺满整条价值链。
亚马逊云科技为此拿出了自身的运营经验。据其介绍，亚马逊运营网络中已部署超过100万台机器人。单台机器人的学习之外，更有价值的是集群层面的学习，以及这些经验如何在全球运营网络中复制。
从区域市场的差异，到研发和供应链的投入回报，再到数据、组织与安全，物理AI走出试点之后，如何成为一套可以持续运行、也值得持续投入的生产系统？
模型在美国，速度在中国，规则在欧洲
Ozgur Tohumcu在主题分享中列出了三个趋势：自动驾驶经过多年的训练和部署，开始进入大规模落地阶段；物理AI在工业场景中的应用不断扩大，制造业和汽车行业尤其明显；全球大型企业正在向智能体架构和智能体组织演进，应用范围覆盖销售、营销、采购、供应链管理和制造等环节。
面对这轮变化，不同地区企业的优势也有所不同。在他看来，前沿模型上美国领先，围绕前沿模型形成了一个巨大的重心，物理AI落地时仍要依赖这部分能力，美国因此保持领先。
中国、日本和韩国等地，尤其是中国，领先的是速度和规模。他用"令人惊叹"来形容中国的节奏，这里试验探索的氛围很浓，一旦某种做法被验证有效，企业会以极快的速度扩张。
在欧洲地区，企业在治理政策上很强，对主权的理解很深，部署物理AI系统时，用了什么数据、数据存在哪里、谁有权访问、数据性质如何，都需要在应用推广之前考虑清楚。
区域差异之外，还有一个变量在重塑车企的技术路线，合规政策持续收紧，数据主权要求叠加关税影响，企业的应对方式是，在保持全球化运营的同时把更多事情区域化，尽量把业务闭环留在同一个区域内。例如，亚马逊云科技在德国建了欧洲主权云，满足部分欧洲数据留在欧洲境内的要求。
落到亚马逊云科技自身，Ozgur Tohumcu给出的角色定位是一站式平台。
首先是模型接入，客户可以根据具体任务选择基础模型，有些任务并不需要使用成本最高的模型。模型蒸馏和定制是另一个重点，训练完成后，要让模型运行在工厂设备或机器人上，还需要重新考虑部署环境、响应速度和成本。
在这一体系中，Amazon Bedrock提供统一的模型接入途径，Amazon Bedrock AgentCore支持智能体的统筹管理。随着企业内部运行的智能体数量增加，对权限、运行过程和生命周期的管理也会变得更加重要。
工厂车间对延迟敏感，边缘侧的低延迟方案和底层数据基础设施，则支撑着模型从云端开发走向现场部署。
试点跑通之后，还要过规模与成本两道关
AI在汽车制造业的哪些环节最容易产生回报？
Ozgur Tohumcu表示，真正的效率突破很少来自单个环节，而是全链条打通带来的协同。具体到投入产出，研发通常最先也最快见到回报，汽车软件的设计、开发、验证、部署四个阶段都能吃到AI的红利，设计阶段迭代更快，开发阶段有 Kiro 这样的编程助手提升工程师效率，验证阶段可以快速生成仿真工况做高并发测试。
例如，宝马集团已经在亚马逊云科技上完成了超过10亿公里的自动驾驶仿真训练，欧洲传统OEM的研发周期过去在4到5年，现在压到3年左右，中国车企只需18到24个月。
供应链排在第二，重构难度大，但降本空间也最大。供应链数据散在很多系统里，智能体可以跨系统取数，把提炼价值和简化流程的用例搭起来。Ozgur Tohumcu认为，一家典型传统车企如果重塑供应链运营，能省下数十亿美元。
AI 的难点还在于从试点走到量产，第一是容不容易规模化，推广到生产环境要付出多少工作量，第二是规模化之后还赚不赚钱。
Ozgur Tohumcu见过的最典型情况是，德国某家工厂的试点效果极好，但要推广到140家工厂，每位厂长都在管自己的运营体系，亚马逊云科技需要证明这套体系对每家工厂都有作用。
财务这一关同样难过，一辆车上运行良好的方案，推广到大型车企数千万辆汽车的体量，成本可能显著增加，企业甚至需要重新选择基础模型，以降低长期运行费用。之前也存在一种情况，客户已经证明方案在技术上可以扩展，首席财务官却认为商业角度不划算。
这些问题最终又回到数据、安全和组织基础上。
在制造环境中，运营数据与IT数据往往相互割裂，如何把工厂设备产生的OT数据接入云端，并与企业管理系统中的IT数据互通，听起来基础，却仍是许多客户需要补上的一环。亚马逊云科技所强调的工业数据架构，就是要将制造执行系统（MES）、产品生命周期管理系统（PLM）等不同来源的数据连接起来，为后续应用提供基础。
设备连接越多，生产现场对安全管理的要求也越高，客户需要确认，既有的安全基础设施能否覆盖新增的智能设备，以及这些设备的运行过程能否得到有效管理。
Ozgur Tohumcu还提到了一个容易被忽略的问题，单台机器人已经可以展示很强的能力，但进入生产环境后，更重要的是机器人之间如何通信、如何协同，以及如何与人共同工作。
对于汽车座舱和自动驾驶中的模型幻觉风险，Ozgur Tohumcu则明确了职责边界，亚马逊云科技提供基础设施，具体自动驾驶系统的设计、训练、运营和验证，仍由客户负责。
组织和人才，是这条落地链路的另一端。亚马逊云科技行业集群总经理田锋表示，汽车和制造业是少数具备物理实体、产出物理产品的行业，链条长，研发环节与其他行业完全不同，客户的诉求不只在技术，组织与人才层面的AI素养同样关键。（本文首发于钛媒体APP，文 | TechPulse）
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。
“When an enterprise pushes an AI program to scale, it is usually not technology that slows down.” Ozgur Tohumcu, the global manager of Amazon Cloud Technology and Manufacturing, gave a rather unusual judgment.
In the past month or two, the LLM industry has even been discussing whether to slow down the pace of forward model development, and model capacity itself is sufficiently advanced and platform services are sufficiently advanced. There are other elements that are really stuck in the business, whether the data base for the expansion of AI services is holding up, whether the fences, observational and traceability mechanisms are available for the self-implemented AI agent, and whether the organization is open to testing.
Ozgur Tohumcu, for example, indicated that the AI project of a large number of cars was at a conceptual validation stage without a real move towards mass production and industrialization, that manufacturing firms had proven AI to work and that the remaining challenges were how to fill the entire value chain.
Amazonian cloud technology has brought its own operational experience to this end. According to them, more than 1 million robots have been deployed in the Amazon operating network. In addition to the learning of single robots, more valuable is the learning at the cluster level and how these experiences are replicated in global operating networks.
From regional market differences to returns on R & D and supply chain inputs, to data, organization and security, how can physical AI emerge from the pilot as a production system that is sustainable and worth sustained input?
Models in America, speed in China, rules in Europe.
Ozgur Tohumcu lists three trends in the thematic sharing: auto-driving, after years of training and deployment, begins to enter a large-scale landing phase; the growing use of physical AI in industrial scenes, particularly in manufacturing and automobile industries; and the evolution of large global enterprises to AI agent structures and AI agent organizations, with applications covering sales, marketing, procurement, supply chain management and manufacturing.
In the face of this change, the advantages of enterprises vary from region to region. In his view, the United States was leading on the front model, forming a huge focus around the front model and still relying on that capability when physical AI landed, so the United States retained the lead.
China, Japan and South Korea, in particular, are leading by speed and size. He described China's rhythm as "surprisingly" where the atmosphere of experimental exploration was strong, and once a practice was proven to be effective, firms would expand at an extraordinary rate.
In the European region, companies are strong in governance policies and have a strong understanding of sovereignty, and the data they use, where they exist, who has the right to access and the nature of the data need to be considered before their application can be replicated.
In addition to regional differences, there is a variable that is reshaping the technological path of a car company, continuing tightening compliance policies, data sovereignty requiring tariff overloading, and that enterprises respond by regionalizing more while maintaining global operations and keeping business closed to one region as much as possible. For example, Amazon cloud technology has built a European sovereignty cloud in Germany to meet the requirement that some European data remain in Europe.
Ozgur Tohumcu's character is the all-in-one platform.
The first is model access, whereby clients can select the underlying model based on specific tasks, some of which do not require the use of the most costly model. The distillation and customization of models is another priority, and once the training has been completed, the deployment environment, response speed and costs will need to be reconsidered in order for the models to operate on plant equipment or robots.
In this system, Amazon Bedrock provides a unified model access and Amazon Bedrock AgentCore supports the integrated management of AI agent. With the increase in the number of AI agentes operating within enterprises, the management of permissions, operating processes and life cycles will become more important.
Plant workshops are sensitive to delays, low-delayed programmes on the edges and the bottom data infrastructure, which support the evolution of the model from cloud to field deployment.
After the pilot runs, there's a double scale and cost.
Which parts of AI are most likely to generate returns in the automotive industry?
Ozgur Tohumcu states that real efficiency breakthroughs rarely come from individual links, but rather from synergies brought about by chain-to-chain connectivity. In terms of input output, R & D is usually the first and the fastest-to-see return, with four phases of design, development, validation and deployment of automobile software reaping AI dividends, a more iterative design phase, a programming assistant such as Kiro in development to improve the efficiency of engineers, and a rapid production of simulations for high-level and high-quality testing at the validation stage.
For example, the BMW Group has completed over 1 billion kilometres of auto-driving simulation training in Amazon cloud technology, and the European traditional OEM research and development cycle, which has been running for four to five years, is now on the order of three years, with only 18 to 24 months of driving in China.
The supply chain is second in rank and is difficult to reconstruct, but it is also the largest reduction space. Supply chain data are scattered in many systems, and AI agent can be counted across systems, combining refining values with the use of simplified processes. Ozgur Tohumcu argues that a typical traditional car company could save billions of dollars if it re-engineered the supply chain.
AI also faces difficulties in moving from pilot to production, first, in terms of size, the amount of work to be carried out in the production environment, and secondly, in terms of profit-making after scale.
The most typical case that Ozgur Tohumcu has seen is that a German plant has been very successful in piloting it, but to reach 140 factories, each director is running its own operating system, and Amazon cloud technology needs to prove that it works for each plant.
Finance is equally sad, with well-functioning programmes that extend to tens of millions of vehicles in large cars, with potentially significant cost increases, and businesses even need to re-select their basic models to reduce long-term operating costs. There was also an earlier situation where clients had demonstrated that the programme could be technically expanded, while the Chief Finance Officer considered the business perspective uneconomical.
These problems eventually return to a data, security and organizational basis.
In the manufacturing environment, operational and IT data tend to be fragmented, and how to connect the OT data generated by plant equipment to the cloud and with IT data in enterprise management systems sounds to be the basis, but still remains the link that many clients need to make up. The industrial data architecture highlighted by Amazon cloud technology is to link data from different sources, such as manufacturing implementation systems (MES), product life cycle management systems (PLM), to provide the basis for subsequent applications.
The more equipment is connected, the higher the security management requirements at the production site, the greater the client needs to ascertain whether the existing security infrastructure will cover the additional smart equipment and whether the operation of the equipment will be effectively managed.
Ozgur Tohumcu also referred to the easy question of how one robot can already demonstrate a strong capability, but when entering the production environment, it is more important how robots communicate, work together, and work with people.
With regard to the risk of model hallucinations in car compartments and auto-driving, Ozgur Tohumcu defines the line of responsibility, and Amazon cloud technology provides infrastructure, the design, training, operation and validation of specific auto-driving systems remains the responsibility of the client.
Organization and talent are the other end of this landline. The General Manager of the Amazon Science and Technology Cluster stated that automobiles and manufacturing were a small number of industries with physical entities, output physical products, long chains, and that the R & D chain was completely different from that of other industries, that client demands were not limited to technology, and that organizational and human-level AI literacy were as critical. (This post was launched on titanium medium AP, in | TechPulse)
Auto-fetched and translated by AI Atlas. All rights belong to the original author.