When Will Models Run Locally vs. In Data Centers?

In this episode, ARK’s Brett Winton, Charles Roberts and Frank Downing sit down with Stephen Balaban, CEO and co-founder of Lambda Labs — a company building AI-specific cloud infrastructure. The conversation explores Lambda’s role in the AI value chain, the evolving economics of data centers, and why traditional hyperscalers might be too slow to meet the moment.

Stephen explains why he believes we're transitioning from deterministic, rule-based software to what he calls “neural software” — stochastic, neural network-driven systems that will eventually replace nearly all traditional software. He shares Lambda’s mission to enable this transformation by rapidly deploying GPU infrastructure and supporting the AI research and application build-out happening today.

The discussion spans infrastructure strategy, regulatory bottlenecks, AI safety, energy constraints, and long-term visions of neural operating systems. Stephen offers a bold perspective on the hardware demands and philosophical shifts required to usher in a world where software is generated, not written.

Key Points From This Episode:

• 00:01:21 How Lambda positions itself as a “neo-cloud” provider competing with AWS, Azure, and GCP for AI workloads.
• 00:02:46 Why ARK estimates $1.5 trillion in annual AI-related data center investment by 2030 and what it could mean for Lambda.
• 00:05:26 Why hyperscalers may be too slow to meet the unique demands of AI training compared to specialized players.
• 00:06:29 How AI infrastructure requires new rack designs, higher power density, and different utilization patterns.
• 00:09:20 Why AI may disrupt the entire computing stack—from Nvidia overtaking Intel to reshaping platform and cloud services.
• 00:14:50 Stephen explains Lambda’s “secret mission” to replace all traditional software with neural networks.
• 00:16:36 Why companies trust Lambda to deploy GPU infrastructure faster and more reliably than incumbents.
• 00:20:27 How the concept of a “neural operating system” reframes software as stochastic rather than deterministic.
• 00:23:04 How hallucinations in neural systems could be managed with checks and balances similar to financial approvals.
• 00:25:04 Why Stephen sees AI safety and alignment as the cybersecurity of the future.
• 00:39:00 How real-time AI tasks may run locally at the edge, while deeper reasoning gets pushed to the cloud.
• 00:44:11 Why running modern large language models still resembles the supercomputer era rather than the PC era.
• 00:46:06 How Stephen views the long-term convergence of AI with quantum computing and brain–computer interfaces.
• 00:50:20 Why scaling AI requires the “heroic effort” of Nvidia, TSMC, OpenAI, energy providers, and Lambda together.
• 00:53:43 Back-of-the-envelope math on CapEx per megawatt—from power plants and data centers to GPUs.
• 00:57:11 Why power infrastructure and deregulation could become the biggest stumbling blocks for AI growth.
• 01:02:02 How software creation is shifting from a labor-driven process to a capital-intensive one.
• 01:06:06 Why Stephen and Brett describe data centers as “AI factories” producing custom neural software.