At the recent Milken Institute Global Conference in Beverly Hills, leaders from across the AI ecosystem shared a surprisingly candid view of the industry’s biggest challenges. The conversation brought together executives working across chips, cloud infrastructure, AI agents, robotics, and alternative AI architectures.

What emerged was a much more grounded picture of the current AI boom. Behind the rapid product launches and trillion-dollar optimism, the industry is running into very real limits around hardware, energy, data, governance, and even the underlying assumptions powering modern AI systems.

The discussion included leaders from ASML, Google Cloud, Applied Intuition, Perplexity, and Logical Intelligence. Together, they highlighted where the AI economy is beginning to strain under its own scale.

The AI Boom Is Hitting Physical Limits

One of the clearest themes from the panel was that AI growth is no longer constrained primarily by ideas or software. It is constrained by physical infrastructure.

Christophe Fouquet, CEO of ASML, explained that the semiconductor industry is accelerating production aggressively, yet demand still far exceeds supply. Advanced AI systems depend on highly specialized chips, and the companies building massive AI platforms are competing for limited manufacturing capacity.

This matters because AI infrastructure cannot scale independently from semiconductor production. Without access to cutting-edge chips, even the most advanced models face performance and efficiency limitations.

At the same time, cloud providers are seeing unprecedented demand. Google Cloud revealed that its backlog of committed infrastructure revenue has expanded dramatically, signaling that companies are rushing to secure long-term AI compute capacity before shortages become even more severe.

The challenge becomes even more complex in physical AI systems.

Applied Intuition, which develops autonomy systems for vehicles, drones, defense systems, and industrial equipment, pointed to a different bottleneck entirely: real-world data.

Unlike purely digital AI systems, physical AI cannot rely entirely on synthetic simulations. Autonomous systems still need exposure to unpredictable real-world environments in order to learn safely and effectively. Simulation helps, but it cannot fully replace real operational data.

That distinction is increasingly important as industries move from generative AI interfaces toward AI systems that interact directly with the physical world.

Energy Is Becoming the Next Major Constraint

If compute availability is the first challenge, energy may become the defining infrastructure problem of the next decade.

Google Cloud confirmed during the discussion that orbital data centers are now being explored as a serious long-term possibility. The motivation is straightforward: AI systems require enormous amounts of electricity, and future demand may outpace terrestrial infrastructure expansion.

Even so, moving infrastructure into space introduces entirely new engineering problems, particularly around heat dissipation. Traditional cooling systems depend heavily on air and liquid cooling, while space environments rely primarily on radiation-based heat transfer, which is significantly more difficult to manage efficiently.

The broader message was clear: AI scalability is increasingly an energy problem as much as a computing problem.

This is also changing how companies think about infrastructure design. Google emphasized the advantage of vertically integrated AI systems, where custom chips, models, and infrastructure are co-designed together for greater efficiency.

That strategy reflects a growing realization across the industry: AI performance is no longer measured only by model quality, but also by operational efficiency, energy consumption, and long-term infrastructure sustainability.

Some Companies Are Questioning the Entire LLM Paradigm

While much of the AI industry continues to focus on scaling large language models, some researchers are beginning to question whether scale alone is the right direction.

Logical Intelligence introduced a very different approach based on Energy-Based Models (EBMs). Instead of predicting the next token in a sequence, these systems aim to model the underlying structure and rules governing data.

The company argues this approach more closely resembles human reasoning, particularly in environments where understanding physical systems matters more than generating language.

Their models are significantly smaller than modern LLMs yet reportedly operate much faster and adapt more dynamically to changing information.

The implications are important for industries such as robotics, engineering, industrial automation, and chip design, where AI systems need to reason about physical environments, not simply generate convincing text.

This reflects a broader shift now emerging inside the AI industry itself: growing interest in architectures optimized for reasoning, adaptability, and operational efficiency rather than scale alone.

AI Agents Are Introducing New Security and Governance Challenges

Another major focus of the conversation was the rapid evolution from AI assistants toward autonomous AI agents.

Perplexity described its vision of AI systems acting more like digital employees than traditional software tools. Instead of simply answering questions, these systems are increasingly designed to execute workflows, interact with enterprise systems, and perform operational tasks autonomously.

That transition creates obvious governance and security concerns.

One of the most important concepts discussed was permission granularity. In enterprise environments, AI agents require tightly controlled access rules that define not only which systems they can interact with, but also whether they can read, modify, or execute actions inside those systems.

This distinction becomes critical once AI systems begin operating directly inside production environments, financial systems, infrastructure platforms, or operational workflows.

The conversation reinforced a growing reality for enterprise AI adoption: trust, auditability, permissions, and operational control are becoming just as important as model capability itself.

For companies deploying AI internally, governance architecture is rapidly becoming a competitive advantage.

Physical AI Is Becoming a Sovereignty Issue

One of the strongest themes from the discussion involved the geopolitical implications of physical AI.

Unlike traditional software platforms, physical AI systems operate directly inside national infrastructure, transportation, agriculture, manufacturing, defense, and logistics environments.

That changes how governments think about control and dependency.

Applied Intuition argued that many countries are increasingly uncomfortable relying on foreign-controlled intelligence systems operating inside critical physical infrastructure. Autonomous systems introduce concerns around safety, operational authority, data ownership, and national resilience.

This dynamic may shape the next phase of AI competition globally.

At the same time, ASML highlighted how semiconductor supply chains remain one of the biggest strategic chokepoints in global AI development. Even highly advanced AI software ecosystems remain dependent on access to cutting-edge hardware manufacturing capabilities.

The result is an AI industry that is becoming deeply tied to industrial policy, energy strategy, semiconductor independence, and geopolitical alignment.

The Industry Is Moving Beyond “Bigger Models Solve Everything”

One of the most important takeaways from the panel was that the AI conversation is maturing.

The industry is starting to move beyond the assumption that scaling larger models alone will solve every problem.

Instead, the next phase of AI development appears increasingly shaped by:

Infrastructure availability
Energy efficiency
Real-world operational data
Governance and trust
Physical-world reasoning
Hardware access
Sovereignty and regulation
Human oversight and security controls

This shift matters for companies evaluating their own AI strategies.

The future competitive advantage may not come from simply adding AI features faster than everyone else. It may come from building systems that are operationally sustainable, trustworthy, efficient, maintainable, and deeply integrated into real business workflows.

That is a much harder engineering challenge, and one that looks increasingly closer to systems architecture than to model experimentation alone.

Source