Navigating the AI Frontier: Bottlenecks, Energy, and Foundational Shifts
A recent discussion at the Milken Institute Global Conference brought together key figures from across the artificial intelligence supply chain to address critical challenges and emerging paradigms. The panel, featuring leaders from semiconductor manufacturing, cloud infrastructure, AI simulation, AI-native search, and foundational AI research, highlighted the tangible constraints shaping the industry’s trajectory, alongside divergent approaches to intelligence itself.
The Physical Limits of AI Growth
The exponential growth in AI demand is encountering significant physical limitations. Christophe Fouquet, CEO of ASML, the dominant provider of extreme ultraviolet (EUV) lithography machines essential for advanced chip production, projected that the market will remain “supply limited” for the next two to five years. Despite accelerated chip manufacturing efforts, hyperscalers such as Google, Microsoft, and Amazon will likely face shortages of the advanced semiconductors they require.
Francis deSouza, COO of Google Cloud, underscored the scale of this demand. He noted that Google Cloud’s revenue surged by 63% to over $20 billion last quarter, with its committed revenue backlog nearly doubling in a single quarter to $460 billion. This rapid expansion points to a genuine and escalating demand for AI-driven infrastructure.
Qasar Younis, co-founder and CEO of Applied Intuition, identified a different bottleneck: real-world data acquisition. His company, which develops autonomy systems for vehicles and defense applications, finds that even advanced simulation cannot fully replace the necessity of collecting data from physical environments. Younis stated that fully training models on physical-world interactions synthetically remains a distant prospect.
The Looming Energy Challenge and Architectural Innovations
Beyond silicon constraints, energy consumption presents a major hurdle. DeSouza revealed that Google is actively exploring the viability of orbital data centers as a means to access more abundant energy resources, acknowledging the significant engineering challenges associated with heat dissipation in a vacuum.
He also emphasized the efficiency gains achieved through Google’s integrated AI stack, from custom Tensor Processing Units (TPUs) to sophisticated models. This co-engineering approach, deSouza argued, yields superior energy efficiency compared to off-the-shelf component integration. Fouquet concurred on the escalating costs associated with increased compute power, noting that the current industry investments, driven by strategic imperatives, come with substantial energy demands and associated expenses.
Eve Bodnia, founder of Logical Intelligence, presented a fundamentally different approach to AI architecture. Her company is developing energy-based models (EBMs) that diverge from the token-prediction paradigm of large language models (LLMs). EBMs aim to discern underlying data rules, mirroring biological reasoning processes rather than linguistic sequences. Bodnia asserted that her company’s models, with significantly fewer parameters than leading LLMs, operate orders of magnitude faster and can update knowledge without requiring complete retraining. She posited that EBMs are better suited for domains requiring an understanding of physical rules, such as robotics and autonomous systems, suggesting that the industry may need to look beyond scale alone.
Agents, Control, and Geopolitical Dimensions
Dmitry Shevelenko, Chief Business Officer of Perplexity, discussed the evolution of his company’s product from a search engine to a “digital worker.” Perplexity’s new offering is designed as an AI staff to augment knowledge workers. Addressing concerns about control, Shevelenko highlighted the importance of granular permissions for enterprise agents, distinguishing between read-only and read-write access. Perplexity’s agent requires user approval before taking actions, a safeguard he deems critical for maintaining security and trust, drawing parallels to the conservative security instincts of established financial institutions.
Younis introduced a geopolitical dimension, arguing that physical AI applications are intrinsically linked to national sovereignty. Unlike purely digital technologies, autonomous vehicles, drones, and other physical AI systems have tangible impacts within national borders, raising critical questions about safety, data governance, and foreign control. He observed that many nations are hesitant to permit physical AI systems controlled by foreign entities operating within their territories.
Fouquet elaborated on this point by examining China’s AI advancements. While acknowledging significant progress at the software layer, he noted that limitations in accessing advanced chip manufacturing technologies, specifically EUV lithography, create a fundamental disadvantage. He contended that the U.S. currently possesses a comprehensive ecosystem, encompassing data, computing access, semiconductor manufacturing, and talent, which China is seeking to replicate.
AI’s Impact on Future Generations
The panel concluded with reflections on AI’s potential impact on critical thinking skills. Panelists offered optimistic perspectives. DeSouza suggested that enhanced AI tools could empower humanity to tackle complex challenges in areas like disease research and climate change, fostering new levels of creativity. Shevelenko posited that while entry-level jobs may change, AI lowers the barrier for independent innovation, with personal curiosity serving as the primary constraint.
Younis distinguished between knowledge work and essential physical labor. He pointed to chronic labor shortages in sectors like agriculture, trucking, and mining, noting that physical AI is poised to fill these existing voids rather than displace willing workers. This perspective frames physical AI as a solution to demographic and labor challenges, particularly in industries that are unattractive to a modern workforce.
Business Style Takeaway: The AI industry is grappling with fundamental supply chain limitations and escalating energy demands, necessitating strategic innovation in both hardware and software architecture. Businesses must assess how these constraints and emerging AI paradigms, such as EBMs and advanced agent systems, will influence their operational strategies, competitive positioning, and long-term technological investments.
Learn more at : techcrunch.com