The AI industry is facing a multitude of challenges, from physical bottlenecks to energy constraints and geopolitical considerations. Five experts from diverse sectors of the AI supply chain recently gathered at the Milken Global Conference to discuss these issues and their implications for the future of AI.
The Physical Bottlenecks
The AI boom is hitting hard physical limits, and the constraints are more widespread than many realize. Christophe Fouquet, CEO of ASML, a Dutch company that holds a monopoly on extreme ultraviolet lithography machines, predicts a supply-limited market for the next few years. This means that hyperscalers like Google, Microsoft, Amazon, and Meta won't get all the chips they're paying for.
Francis deSouza, COO of Google Cloud, highlights the magnitude of the issue. Google Cloud's revenue crossed $20 billion in the last quarter, growing 63%, while its backlog nearly doubled in a single quarter. The demand is real, but the supply chain is struggling to keep up.
Qasar Younis, co-founder and CEO of Applied Intuition, a $15 billion physical AI company, emphasizes the importance of real-world data. Applied Intuition builds autonomy systems for various applications, and the bottleneck isn't silicon but the data gathered from the physical world. Synthetic simulation can't fully replicate the complexities of the real world, and training models on physical data is a long-term process.
The Energy Crisis
Energy is the next major challenge. Google is exploring data centers in space as a response to energy constraints. While space offers abundant energy, it presents unique challenges. Data centers in space rely on radiation to shed heat, a slower and more complex process than conventional cooling systems. However, Google is treating this as a legitimate path forward.
DeSouza argues that Google's strategy of co-engineering its AI stack, from custom TPU chips to models and agents, provides significant energy efficiency. Running Gemini on TPUs is more energy-efficient than using off-the-shelf components, as chip designers are aware of the model's requirements.
Fouquet agrees, stating that the industry is investing heavily due to strategic necessity. However, more compute means more energy, and the price of energy is a critical consideration.
A Different Kind of Intelligence
Eve Bodnia, a quantum physicist and founder of Logical Intelligence, challenges the conventional AI architecture. Her company builds energy-based models (EBMs), which understand the rules underlying data rather than predicting the next token in a sequence. EBMs, according to Bodnia, mimic the human brain more closely.
Logical Intelligence's largest model has 200 million parameters, running thousands of times faster than leading LLMs. It's designed to update knowledge as data changes, eliminating the need for retraining. Bodnia argues that EBMs are more suitable for chip design, robotics, and other domains where physical rules are crucial.
Agents, Guardrails, and Trust
Dimitry Shevelenko, chief business officer of Perplexity, discusses the evolution of their search product into a 'digital worker.' Perplexity Computer is designed as a staff that knowledge workers direct, with granular control over permissions. This approach raises questions about control and security.
Shevelenko emphasizes the importance of granularity in managing agent permissions, especially in corporate systems. When Comet, Perplexity's computer-use agent, takes actions, it presents a plan and seeks approval, ensuring security and trust.
Sovereignty and Geopolitics
Qasar Younis highlights the geopolitical implications of physical AI. Autonomous vehicles, defense drones, and mining equipment have real-world consequences that governments can't ignore. Countries are increasingly concerned about the intelligence in physical forms controlled by other nations.
Fouquet agrees, noting that China's AI progress is constrained below the model layer due to a lack of access to EUV lithography. The United States has an advantage in data, computing access, chips, and talent, while China excels at the top of the stack.
The Generation Question
The panel addressed the concern that AI might impact the next generation's critical thinking. Francis deSouza remains optimistic, citing the potential for more powerful tools to address complex global issues. Dimitry Shevelenko argues that the accessibility of AI technology allows individuals to launch independent projects, fostering creativity.
Qasar Younis draws a clear distinction between knowledge work and physical labor, emphasizing that physical AI is filling existing labor gaps rather than displacing workers. The industry's challenges are significant, but the experts remain optimistic about the future of AI and its potential to address global problems.
