Amazon Web Services' Ozgur Tohumcu: What often determines the speed of physical AI's scaled deployment is not the technology itself

"When companies push AI solutions to scale, what usually slows them down is not technology." Ozgur Tohumcu, AWS Global General Manager for Automotive and Manufacturing, offered a somewhat uncommon judgment.

Over the past month or two, the large-model industry has even seen discussion about whether to slow the pace of frontier model development; model capabilities themselves are already sufficiently advanced, and platform services are sufficiently advanced as well. What is really holding companies back is other factors: whether they have a data foundation that can support the expansion of AI services, whether they have prepared guardrails, observability, and traceability mechanisms for autonomously executing agents, and whether the organization still has room for experimentation.

Ozgur Tohumcu gave an example, saying that a large number of automakers' AI projects remain at the proof-of-concept stage and have not truly moved toward mass production and industrial deployment. Manufacturers have already proven that AI works; the remaining challenge is how to spread it across the entire value chain.

AWS has therefore brought out its own operational experience. According to the company, more than 1 million robots have been deployed in Amazon's operations network. Beyond the learning of a single robot, what is more valuable is learning at the fleet level, and how that experience is replicated across the global operations network.

Models in the US, speed in China, rules in Europe

In his keynote, Ozgur Tohumcu listed three trends: after years of training and deployment, autonomous driving is beginning to enter a phase of large-scale rollout; physical AI is continuously expanding its application in industrial scenarios, especially in manufacturing and the automotive industry; and large global enterprises are evolving toward agent architectures and agent-based organizations, with applications covering sales, marketing, procurement, supply chain management, and manufacturing.

In places such as China, Japan, and South Korea, especially China, what leads is speed and scale. He used "amazing" to describe China's pace; the atmosphere of experimentation and exploration here is strong, and once a practice is proven effective, companies expand it extremely quickly.

Beyond regional differences, another variable is reshaping automakers' technology roadmaps: compliance policies continue to tighten, and data sovereignty requirements are compounded by tariff effects. Companies' response is to regionalize more activities while maintaining global operations, trying to keep business closed loops within the same region. For example, AWS built a European sovereign cloud in Germany to meet the requirement that some European data remain within Europe.

As for AWS itself, the role Ozgur Tohumcu described is that of a one-stop platform.

First is model access; customers can choose foundation models according to specific tasks, and some tasks do not require the most expensive model. Model distillation and customization are another focus; after training is complete, running the model on factory equipment or robots still requires rethinking the deployment environment, response speed, and cost.

In this system, Amazon Bedrock provides a unified path for model access, and Amazon Bedrock AgentCore supports the coordinated management of agents. As the number of agents operating within enterprises increases, management of permissions, runtime processes, and lifecycles will also become more important.

Factory floors are latency-sensitive, and low-latency edge solutions and underlying data infrastructure support the transition of models from cloud development to on-site deployment.

Ozgur Tohumcu said that true efficiency breakthroughs rarely come from a single link, but rather from the synergy created by connecting the entire chain. Specifically regarding input and output, R&D usually sees returns first and fastest; all four stages of automotive software—design, development, validation, and deployment—can benefit from AI. The design stage iterates faster, the development stage has coding assistants such as Kiro to improve engineer efficiency, and the validation stage can quickly generate simulation conditions for high-concurrency testing.

For example, the BMW Group has completed more than 1 billion kilometers of autonomous driving simulation training on AWS. The R&D cycle for traditional European OEMs used to be 4 to 5 years, and is now compressed to about 3 years, while Chinese automakers need only 18 to 24 months.

Supply chain ranks second; restructuring is difficult, but the room for cost reduction is also the largest. Supply chain data is scattered across many systems, and agents can retrieve data across systems to build use cases that extract value and simplify processes. Ozgur Tohumcu believes that if a typical traditional automaker reshapes its supply chain operations, it could save billions of dollars.

The most typical situation Ozgur Tohumcu has seen is that a pilot at a German factory works extremely well, but to roll it out to 140 factories, each plant manager is running their own operating system, and AWS needs to prove that this system works for every factory.

The finance hurdle is equally difficult. A solution that works well on one vehicle, when scaled to the volume of tens of millions of vehicles at a large automaker, may increase costs significantly, and companies may even need to reselect foundation models to reduce long-term operating expenses. There have also been cases where a customer had proven that the solution could scale technically, but the CFO considered it uneconomical from a business perspective.

These issues ultimately come back to data, security, and organizational foundations.

In manufacturing environments, operational data and IT data are often separated from each other. How to connect OT data generated by factory equipment to the cloud and make it interoperable with IT data in enterprise management systems sounds basic, yet it is still a gap many customers need to close. The industrial data architecture emphasized by AWS is intended to connect data from different sources such as manufacturing execution systems (MES) and product lifecycle management (PLM) systems, providing a foundation for subsequent applications.

The more devices are connected, the higher the safety management requirements at the production site. Customers need to confirm whether existing security infrastructure can cover newly added intelligent devices, and whether the operation of these devices can be effectively managed.

Ozgur Tohumcu also mentioned an easily overlooked issue: a single robot can already demonstrate strong capabilities, but after entering the production environment, what matters more is how robots communicate with each other, how they collaborate, and how they work together with humans.

Regarding the risk of model hallucinations in automotive cockpits and autonomous driving, Ozgur Tohumcu clarified the boundary of responsibilities: AWS provides infrastructure, while the design, training, operation, and validation of specific autonomous driving systems remain the customer's responsibility.

Organization and talent are the other end of this deployment chain. Tian Feng, AWS General Manager of Industry Clusters, said that automotive and manufacturing are among the few industries with physical entities that produce physical products; the chain is long, and the R&D process is completely different from other industries. Customers' demands are not only about technology; AI literacy at the organizational and talent levels is equally critical. (This article was first published on TMTPost APP, written by TechPulse)