Earlier this week, five people who touch every layer of the AI supply chain sat down at the Milken Global Conference in Beverly Hills, where they talked with this editor about everything from chip shortages to orbital data centers to the possibility that the whole architecture that undergirds the tech is wrong. On stage with TechCrunch: Christophe Fouquet, CEO of ASML; Francis deSouza, COO of Google Cloud; Qasar Younis, co-founder and CEO of Applied Intuition; Dimitry Shevelenko, chief business officer of Perplexity; and Eve Bodnia, a quantum physicist and founder of Logical Intelligence. Here is what the five had to say.
The bottlenecks are real
The AI boom is running into hard physical limits, and the constraints begin further down the stack than many may realize. Fouquet was the first to say it, describing a “huge acceleration of chips manufacturing,” while expressing his “strong belief” that despite all that effort, “for the next two, three, maybe five years, the market will be supply limited.” This means the hyperscalers — Google, Microsoft, Amazon, Meta — will not get all the chips they are paying for, full stop.
Also read: Thinking Machines Lab unveils AI that listens while it talks, mimicking natural conversation
DeSouza highlighted how big — and how fast growing — an issue this is, reminding the audience that Google Cloud’s revenue crossed $20 billion last quarter, growing 63%, while its backlog nearly doubled in a single quarter, from $250 billion to $460 billion. “The demand is real,” he said with impressive calm.
For Younis, the constraint comes primarily from elsewhere. Applied Intuition builds autonomy systems for cars, trucks, drones, mining equipment and defense vehicles, and his bottleneck is not silicon — it is the data that one can only gather by sending machines into the real world. “You have to find it from the real world,” he said, and no amount of synthetic simulation fully closes that gap.
Also read: GM lays off 600 IT workers in deliberate shift toward AI-native talent
The energy problem is also real
If chips are the first bottleneck, energy is the one looming behind it. DeSouza confirmed that Google is exploring data centers in space as a serious response to energy constraints. “You get access to more abundant energy,” he noted. Of course, even in orbit, it is not simple. DeSouza observed space is a vacuum, so eliminates convection, leaving radiation as the only way to shed heat into the surrounding environment — a much slower and harder-to-engineer process than the air and liquid cooling systems that data centers rely on today. But the company is still treating it as a legitimate path.
The deeper argument DeSouza made was about efficiency through integration. Google’s strategy of co-engineering its full AI stack — from custom TPU chips through to models and agents — pays dividends in watts per flop that a company buying off-the-shelf components simply cannot replicate, he suggested. “Running Gemini on TPUs is much more energy efficient than any other configuration,” because chip designers know what is coming in the model before it ships. Fouquet echoed the point later: “Nothing can be priceless.”
A different kind of intelligence
While the rest of the industry debates scale, architecture, and inference efficiency within the large language model approach, Bodnia is building something very different. Her company, Logical Intelligence, is built on so-called energy-based models (EBMs), a class of AI that does not predict the next token in a sequence but instead attempts to understand the rules underlying data, in a way she argues is closer to how the human brain actually works.
“Language is a user interface between my brain and yours,” she said. “The reasoning itself is not attached to any language.” Her largest model runs to 200 million parameters — compared to the hundreds of billions in leading LLMs — and she claims it runs thousands of times faster. More importantly, it is designed to update its knowledge as data changes, rather than requiring retraining from scratch. For chip design, robotics and other domains where a system needs to grasp physical rules rather than linguistic patterns, she argues EBMs are the more natural fit.
Agents, guardrails, and trust
Shevelenko spent much of the conversation explaining how Perplexity has evolved from a search product into something it now calls a “digital worker.” Perplexity Computer, its newest offering, is designed not as a tool a knowledge worker uses, but as a staff that a knowledge worker directs. “Every day you wake up and you have a hundred staff on your team,” he said of the opportunity. “What are you going to do to make the most of it?”
His answer to questions about control was granularity. Enterprise administrators can specify not just which connectors and tools an agent can access, but whether those permissions are read-only or read-write — a distinction that matters enormously when agents are acting inside corporate systems. When Comet, Perplexity’s computer-use agent, takes actions on a user’s behalf, it presents a plan and asks for approval first. Some users find the friction annoying, Shevelenko said, but he considers it essential. “Granularity is the bedrock of good security hygiene,” he said.
Sovereignty, not just safety
Younis offered what may have been the panel’s most geopolitically charged observation: physical AI and national sovereignty are entangled in ways that purely digital AI never was. The internet initially spread as American technology and faced pushback only at the application layer — the Ubers and DoorDashes — when offline consequences became visible. Physical AI is different. Autonomous vehicles, defense drones, mining equipment, agricultural machines — these manifest in the real world in ways governments cannot ignore. “Almost consistently, every country is saying: we don’t want this intelligence in a physical form in our borders, controlled by another country.” Fewer nations, he told the crowd, can currently field a robotaxi than possess nuclear weapons.
Fouquet framed it a little differently. China’s AI progress is real — DeepSeek’s release earlier this year sent something close to a panic through parts of the industry — but that progress is constrained below the model layer. Without access to EUV lithography, Chinese chipmakers cannot manufacture the most advanced semiconductors. “Today, in the United States, you have the data, you have the computing access, you have the chips, you have the talent. China does a very good job on the top of the stack, but is lacking some elements below,” Fouquet said.
The generation question
Near the end of the panel, someone in the audience asked the obvious uncomfortable question: is all of this going to impact the next generation’s capacity for critical thinking? The answers were optimistic, though not naively so. DeSouza pointed to the scale of problems that more powerful tools might finally let humanity address — neurological diseases, greenhouse gas removal, and grid infrastructure that has been deferred for decades. “This should unleash us to the next level of creativity,” he said.
Shevelenko made a more pragmatic point: the entry-level job may be disappearing, but the ability to launch something independently has never been more accessible. “For anybody who has Perplexity Computer… the constraint is your own curiosity and agency.” Younis drew the sharpest distinction between knowledge work and physical labor. He pointed to the fact that the average American farmer is 58 years old and that labor shortages in mining, long-haul trucking, and agriculture are chronic and growing — not because wages are too low, but because people do not want those jobs. In those domains, physical AI is not displacing willing workers. It is filling a void that already exists and looks only to deepen from here.
Conclusion
The Milken panel made clear that the AI industry is facing a convergence of real-world constraints — chip supply, energy availability, data scarcity, and geopolitical friction — that no amount of software optimization alone can solve. At the same time, new architectural approaches and evolving agent capabilities suggest the path forward is not simply about scaling existing models. The coming years will test whether the industry can address these bottlenecks while maintaining the trust of governments, enterprises, and the public.
FAQs
Q1: What are the main bottlenecks holding back AI development in 2026?
The three primary bottlenecks are chip supply constraints (especially advanced EUV lithography), energy availability for powering massive data centers, and the difficulty of gathering real-world data for physical AI systems.
Q2: Why are companies like Google exploring space-based data centers?
Space offers access to abundant solar energy, but presents engineering challenges like heat dissipation in a vacuum. Google is treating orbital data centers as a serious option to bypass terrestrial energy constraints.
Q3: What are energy-based models (EBMs) and how do they differ from LLMs?
EBMs attempt to understand the underlying rules of data rather than predicting the next token in a sequence. They can be much smaller and faster than large language models, and are designed to update knowledge without full retraining, making them potentially more suitable for robotics and chip design.
Be the first to comment