The AI economy is experiencing a unique set of challenges, as revealed by a panel of experts at the Milken Global Conference. The discussion highlighted the complex interplay between physical limitations, energy constraints, and the evolving nature of AI architecture. Here's a breakdown of the key insights and opinions shared by the panel:
The Bottlenecks and Supply-Demand Imbalance
The panel emphasized the real bottlenecks in the AI supply chain, which extend beyond the widely recognized chip shortages. Christophe Fouquet, CEO of ASML, noted the acceleration in chip manufacturing but expressed a strong belief that the market will be supply-limited for the next few years. This means that hyperscalers like Google, Microsoft, Amazon, and Meta may not get the full value of their chip investments.
Francis deSouza, COO of Google Cloud, underscored the magnitude of the issue, mentioning that Google Cloud's revenue crossed $20 billion in the last quarter, with a 63% growth rate. The backlog, which represents committed but not yet delivered revenue, nearly doubled in a single quarter, from $250 billion to $460 billion. This indicates a significant demand for AI chips, which is not being met by the current supply.
Qasar Younis, CEO of Applied Intuition, highlighted a different kind of bottleneck. His company focuses on building autonomy systems for various applications, and the constraint lies in the data. Applied Intuition's challenge is gathering real-world data, which cannot be fully replicated through synthetic simulations. This data is crucial for training models that operate in the physical world.
The Energy Conundrum
The energy problem is another critical aspect of the AI landscape. Google, according to deSouza, is exploring data centers in space as a potential solution to energy constraints. However, this approach comes with its own set of challenges. Space, being a vacuum, relies on radiation for heat dissipation, which is a slower and more complex process compared to conventional cooling systems. Despite the difficulties, Google is treating this as a legitimate path to address energy efficiency.
Fouquet echoed a similar sentiment, emphasizing that the industry's current investment is driven by strategic necessity. However, he pointed out that more compute power means higher energy consumption, and energy has a price. This highlights the need for efficient integration and optimization of the AI stack to balance performance and energy usage.
Redefining AI Architecture
Eve Bodnia, a quantum physicist turned entrepreneur, introduced a different perspective on AI architecture. Her company, Logical Intelligence, is built on energy-based models (EBMs), which differ from traditional large language models (LLMs). EBMs focus on understanding the underlying rules of data, mimicking the human brain's approach to reasoning. Bodnia's model, with 200 million parameters, runs thousands of times faster than leading LLMs and can adapt to changing data without full retraining.
She argues that EBMs are more suitable for domains like chip design, robotics, and other applications where physical rules need to be grasped. This perspective challenges the notion that scale alone is sufficient in AI, prompting a reevaluation of architectural choices.
Agents, Guardrails, and Trust
Dimitry Shevelenko, Chief Business Officer of Perplexity, discussed the evolution of their company's offerings. Perplexity has transitioned from a search product to a 'digital worker,' with their newest product, Perplexity Computer, acting as a staff that knowledge workers can direct. Shevelenko emphasized the importance of granularity in controlling these agents, allowing enterprise administrators to specify permissions and access levels.
The concept of trust is crucial in this context. Shevelenko, drawing from his experience at Lazard, highlighted the need for conservative instincts in protecting client trust, especially in industries with a long history, like finance. Granularity in agent permissions is essential for maintaining good security hygiene.
Geopolitics and Physical AI
Qasar Younis brought up a geopolitically charged topic, linking physical AI to national sovereignty. He argued that the real-world impact of physical AI, such as autonomous vehicles and defense drones, raises questions about safety, data collection, and control within a nation's borders. Countries are increasingly resistant to allowing foreign-controlled physical AI systems on their soil.
Fouquet provided a different perspective, focusing on the technological disparities between the United States and China. While China has made significant progress in AI, its limitations lie below the model layer, specifically in access to EUV lithography. This constraint hinders China's ability to manufacture advanced semiconductors, creating a compounding disadvantage despite their software advancements.
The Next Generation's Critical Thinking
The panel addressed the concern that AI advancements might impact the next generation's critical thinking abilities. However, the responses were generally optimistic. DeSouza pointed to the potential for more powerful tools to tackle complex global challenges, such as neurological diseases and climate change. He argued that AI could unleash humanity's creativity and problem-solving capabilities.
Shevelenko offered a pragmatic view, suggesting that while entry-level jobs may change, the ability to launch independent projects has become more accessible. Younis drew a clear distinction between knowledge work and physical labor, emphasizing that physical AI is filling existing labor gaps rather than displacing workers.
