Interview #84 Hagay Lupesko, SVP of AI Inference at Cerebras Systems

Join Hagay Lupesko, SVP of AI Inference at Cerebras Systems, for a deep dive into the rapidly evolving world of AI inference. Hagay breaks down why inference has overtaken training as the dominant AI workload, how Cerebras' wafer-scale chip architecture delivers 10-20x faster performance than NVIDIA GPUs, and why CUDA is no longer the moat many think it is. He also covers how DeepSeek wiping $600 billion off NVIDIA's market cap in a single day was both a foundational and deeply misunderstood moment for the industry, the growing energy crisis in AI infrastructure, and what it will take to support the explosive rise of AI agents in the enterprise.

Join Shiva Pillay, SVP and General Manager of Veeam Americas, as he discusses why 80-90% of AI projects fail due to fundamental data issues, revealing that enterprises are attempting to feed AI models with fragmented, poorly governed, and inconsistently labeled data that was never designed for AI consumption in the first place. Pillay explains the dangerous disconnect between C-suite executives confident in their AI-ready initiatives and SMEs still struggling in pilot mode, warning that without proper guardrails, AI errors propagate exponentially faster than human errors with potentially catastrophic consequences from faulty manufacturing parts to compliance violations in regulated industries. Drawing from Veeam's acquisition of Securiti AI, he introduces the concept of data command graph as critical infrastructure for AI resilience, arguing that data protection, recoverability, and visibility must be embedded directly into the AI stack rather than treated as underlying infrastructure, while emphasizing that as agentic AI systems gain autonomy, the existential question shifts from what can AI do to what is it allowed to do.

Out of 7 million podcasts on Spotify, The Artificial Intelligence Podcast achieved top 2% Most Shared, top 9% Most Talked About, and top 7% listener retention in 2025. In this year-end review episode, Dr. Tony Hoang reflects on interviewing 80+ enterprise leaders from Microsoft, Intel, NVIDIA, and Fortune 500 companies, while previewing 2026's bigger AI guests, deeper enterprise AI case studies, and exciting new distribution opportunities. Our community of AI executives and tech leaders doesn't just listen- you implement AI strategies, share insights with teams, and drive digital transformation in your organizations. Want to sponsor the fastest-growing enterprise AI podcast? Visit tonyphoang.com to explore strategic partnerships for 2026.

Join Christopher Brown, Managing Partner of The Nova Rose Group, as he examines the collision course between AI innovation and governance, arguing that America's fragmented 50-state regulatory patchwork creates unprecedented challenges for an industry evolving too rapidly for traditional democratic experimentation. Brown critiques the Trump administration's AI litigation task force threatening to sue states over AI laws, warning that while maintaining competitive dominance against China is valid, the U.S. risks becoming what it opposes by sacrificing privacy protections and anti-discrimination safeguards in the rush to innovate. He frames AI regulation as an infrastructure problem requiring a comprehensive "national highway system" approach rather than piecemeal solutions, contending that companies pouring trillions into AI investments while ignoring workforce impacts, healthcare risks, and algorithmic bias are creating a financial and social time bomb that could undermine American leadership rather than secure it.

Join Vrajesh Bhavsar, CEO of Operant AI, as he reveals the emerging threat landscape facing production AI systems, from sophisticated prompt injection attacks to zero-click exploits that can exfiltrate sensitive data without any user interaction. Bhavsar explains how traditional security tools like code scanning, network firewalls, and cloud security posture management become largely ineffective against AI agents operating with non-deterministic behavior and authorized access to critical systems, requiring runtime security solutions that function as AI-layer firewalls. He discusses the Shadow Escape attack class discovered by Operant targeting Model Context Protocol (MCP) servers, where the rapid proliferation of 20,000 mostly open-source MCP servers creates dangerous shared memory spaces across disparate API endpoints, enabling data poisoning and exfiltration at scale while traditional security teams remain blind to these agent-to-agent interactions happening within the "cloud within the cloud."

Join Raj Shukla, CTO of SymphonyAI, as he discusses the critical distinction between AI demos and production-ready systems, revealing that enterprises consistently underestimate the "last mile" challenges of authentication, authorization, and data scalability that break POCs when moving to production. Shukla explains how SymphonyAI's vertical AI approach pre-trains models on industry-specific ontologies and knowledge graphs for retail, financial services, industrial manufacturing, and enterprise IT, enabling faster ROI by providing pre-built agents and domain-specific context rather than expecting generic LLMs to solve everything. He addresses the hidden costs that shock CFOs—not LLM inference which has dropped 1000x, but the expensive work of making data and APIs AI-ready through proper governance layers and MCP server implementations, while warning that enterprises overestimate the autonomy achievable in the short term and underestimate the infrastructure work required for real process automation at scale.

Join Balaji Raghavan, Head of Engineering at Postman, as he discusses the critical gap between AI adoption and API readiness, revealing that while 80% of developers use AI, only 24% design APIs with AI agents as the intended consumer. Drawing from Postman's 40 million developer user base, Raghavan explains how human-designed APIs create ambiguity problems for AI systems, requiring additional tooling layers that often introduce security vulnerabilities through proxy credentials and unauthorized access risks. He addresses the uncomfortable reality that the industry is still in early stages of making AI reliably call APIs at scale, with hallucinations and context limitations preventing effective orchestration across hundreds of endpoints, while warning that judicious leaders must distinguish between deterministic flows and cases where expensive AI-based approaches are truly necessary to manage infrastructure costs and prevent cascade failures.

Join Stelios Diamantidis, Chief Product Officer at Cognichip, as he explores how artificial intelligence is revolutionizing semiconductor development by enabling more holistic design processes that can reduce development time by half and costs by 75%. He discusses how AI tackles complex challenges across the entire chip design workflow—from early product definition through manufacturing—including verification, debugging, and hardware-software co-design optimization. Diamantidis envisions a near future where AI agents serve as true co-designers, helping engineers navigate the complex trade-offs between performance, power efficiency, and chip area while enabling rapid creation of bespoke accelerators tailored to specific AI workloads.

Join Paul Canetti, CEO of Skej, as he discusses the unique challenges of building AI products that operate without traditional user interfaces, instead functioning as virtual humans with email addresses, phone numbers, and Slack handles that interact through natural language. Drawing from his experience in UX design at Apple during the iPhone era, Canetti explains how building non-deterministic AI systems fundamentally differs from traditional software, requiring multiple quality assurance layers to prevent hallucinations and ensure AI assistants know when to remain silent in group conversations. He explores the shift toward anthropomorphized AI assistants with distinct personalities, arguing that as forms become obsolete and natural language interfaces become mainstream, the future lies in liberating people to do uniquely human work while AI handles generic tasks that anyone could accomplish but everyone suffers through.

Join Zachary Hanif, VP of Data and AI at Twilio, as he discusses the fundamental differences between building AI systems in regulated financial services versus communication platforms, drawing from his experience at Capital One to implement rigorous model governance frameworks that reduce maintenance costs while accelerating development timelines. Hanif addresses the critical balance between explainable AI and high-performing black box models, emphasizing that organizations must identify where their use cases fall on the explainability spectrum rather than applying blanket requirements. He explores privacy-by-design principles for real-time AI systems, the challenge of moving from proof-of-concept to production (with 80% of AI pilots failing), and provides a practical framework for successful AI implementation that includes clear objective criteria, close collaboration between technical teams and domain experts, and properly tempered expectations for experimental development timelines.