AWS's Ozgur Tohumcu on How Physical AI Can Cross the Scale Threshold

Smart Things reports on October 23 that recently, at the 2026 Amazon Web Services Automotive and Manufacturing Industry Summit, Ozgur Tohumcu, Global General Manager of Automotive and Manufacturing at Amazon Web Services, shared the changes AI is bringing to manufacturing in the AI era, and how AI can move beyond pilots into scaled production environments. After the event, he further exchanged views on these thoughts with Smart Things and other media.

In his speech, Ozgur Tohumcu emphasized that more and more proofs of concept are now moving into production environments, which greatly changes how we understand agentic systems. With the introduction of agentic thinking and related enterprise application principles, these systems are moving toward the execution layer, no longer just aggregating data from different sources and displaying it on dashboards, but beginning to actually complete tasks.

He is also very optimistic about the development of physical AI. Physical AI has many opportunities in manufacturing, such as robot vision guidance, internal logistics, humanoid robot applications, quality inspection, system optimization, and self-developed models. Among these, the widespread application of humanoid robots still requires consideration of safety and economics.

▲ Ozgur Tohumcu, Global General Manager of Automotive and Manufacturing at Amazon Web Services

Amazon has built one of the world's most advanced robot operating systems, deploying more than 1 million robots across its operations network. These robots handle millions of packages every day and work alongside hundreds of thousands of employees. Amazon Web Services is supporting the full-chain development of physical AI through five pillars: foundation models, simulation and synthetic data, model distillation and optimization, edge computing and local deployment, and hardware and physical carriers.

It is no longer surprising for a robot to move goods and sort items. What truly tests enterprises is replicating such capabilities across more devices and more factories while keeping costs manageable. This is exactly a challenge facing the automotive and manufacturing industries.

According to Ozgur Tohumcu's observations, the global automotive and manufacturing industries have three major trends: 1) after years of AI training and deployment, autonomous driving has entered the stage of large-scale implementation; 2) physical AI is deeply penetrating manufacturing and the automotive industry; 3) large global enterprises are moving toward agentic architectures and agentic organizations, spanning sales, marketing, procurement, supply chain management, and other links.

Among these, physical AI is particularly noteworthy. It takes AI capabilities from the screen into reality, involving how machines perceive the environment, understand tasks, and take action. But strong standalone robot capability does not mean an entire factory can operate efficiently. Communication between devices, collaboration between robots and people, and reusable experience are the problems that must be solved for scaled application.

Ozgur Tohumcu believes that more valuable than the learning of a single device is experience sharing among robot fleets and the replication of these capabilities across the global operations network.

"Our joint practice with customers shows that true business value never comes from a breakthrough in a single link, but from collaborative innovation across the entire value chain," said Ozgur Tohumcu.

Among these, data issues are often the most fundamental and the hardest to avoid. Supply chain, production, and sales each have their own systems, and the operating data of workshop equipment and enterprise management data are often separated from each other. For AI to assist in formulating production plans, it needs to understand orders, materials, and capacity at the same time. Simply storing data centrally is not enough to support cross-department business execution. Many enterprises are redesigning their data architectures to unleash as much of AI's value-creation potential as possible.

Governance requirements are also increasing. In the past, AI provided suggestions, and people decided whether to adopt them. Now, agents are beginning to call tools and advance processes, so enterprises must clarify what they can access, what operations they can perform, and how each step is recorded and traced.

Organization also determines speed. A solution proven effective in one factory still faces different processes, management habits, and personnel trust when promoted to another factory. Technical solutions can be copied, but organizational acceptance requires a process.

Tian Feng, General Manager of Industry Clusters at Amazon Web Services, also said that automotive and manufacturing are industries with physical entities, with more complex industrial chains, and their R&D links are completely different from those of other industries. Amazon Web Services has found that customers' real demands are not only for technical support; improving AI literacy at the organizational and talent levels is equally crucial. To this end, Amazon Web Services has launched products such as Amazon Quick and Kiro and has accumulated a large number of excellent practice cases.

▲ Tian Feng, General Manager of Industry Clusters at Amazon Web Services

These issues form the entry point for Amazon Web Services. When customers deploy physical AI, Amazon Web Services is committed to providing one-stop platform services: providing cutting-edge foundation models through Amazon Bedrock, supporting orchestration and lifecycle management of agents through Amazon Bedrock AgentCore, and providing underlying infrastructure and low-latency solutions at the edge.

Model selection should also be judged according to the task. Complex tasks require stronger capabilities, while high-frequency, well-defined tasks require careful cost trade-offs. Call costs acceptable during the pilot phase may change the commercial viability of a project once scaled to large numbers of devices or vehicles.

So where will enterprises see returns first?

Ozgur Tohumcu judges that in manufacturing and the automotive industry, the R&D link can often bring investment returns to customers earliest and fastest. From design iteration and software development to simulation verification, AI can participate in multiple links. Enterprises have the opportunity both to compress R&D costs and to shorten product time to market.

The supply chain holds another type of opportunity. Data is scattered within enterprises and in supplier systems, and agents are expected to help enterprises obtain information across systems and connect processes.

By comparison, the value of after-sales service is more reflected in customer satisfaction and retention, and may not quickly translate into direct revenue.