Former Intel CEO on What Went Wrong, What's Next + Lovable CEO on the Real Promise of Vibe Coding

Summary

This episode splits between semiconductor geopolitics and the next phase of AI software creation. Pat Gelsinger argues Intel’s problems were structural: the company shifted from technologists to business leaders, underinvested in manufacturing, and failed to keep pace with TSMC’s foundry model and Nvidia’s CUDA-driven software ecosystem. He also warns that Taiwan energy constraints make a blockade a severe global economic risk, and says AI’s near-term limiter is energy capacity rather than model hype. On the software side, Anton Osika says Lovable is seeing explosive usage from both technical and non-technical users, routing work across multiple frontier and open-weight models while training on its own failure cases. He frames the company’s roadmap as moving from helping people build software to helping them run businesses with an AI co-founder that can use platform data to suggest strategy and operations.

Notes

• Gelsinger says Intel's decline began when it was increasingly run by business leaders instead of technologists.
• Intel had returned $100B to shareholders before Gelsinger returned, instead of reinvesting.
• He says Intel had not built a new factory in a decade.
• He argues buying EUV machines was a technologist-level decision Intel should have made.
• Apple's move to in-house silicon is presented as a staged competence-building exercise.
• Nvidia's rise is described as a shift from graphics hardware to a software stack around CUDA and HPC.
• TSMC's foundry model is framed as a structural industry change enabled by standardization.
• He says TSMC was already producing 5x Intel's wafers in 2001, now roughly 7:1.
• He warns Taiwan has less than three weeks of energy reserves under blockade conditions.
• A fab shut down by brownout may take 90 days to restart, making disruption especially severe.
• He says AI infrastructure is constrained by energy capacity, not just model demand.
• Global energy capacity is growing 4-5%, while the U.S. had a decade at 1%.
• He aims to make AI 10,000x better by cutting token cost and energy use by five orders of magnitude.
• He expects meaningful quantum results this decade, with encryption impacts around 2032-2033.
• He cites trapped ions, photonics, and spin approaches as promising quantum modalities.
• Osika says Lovable sees one million new projects built every week.
• Lovable claims more than 50 million apps built on the platform in 20 months.
• About 20% of Lovable users are technical and 80% are non-technical.
• Lovable handles secure payments, security scans, and background monitoring as part of the platform.
• Osika says Lovable can sit on top of Google, Microsoft, Slack, Salesforce, and HubSpot while preserving data-flow visibility.
• A customer example replaced more than 10 bespoke tools and saved over $1M a year.
• Lovable is routed across multiple frontier and open-weight models by task.
• Lovable's own models get better through agent-harness usage, especially on tool calling and integration creation mistakes.
• Osika says the Stockholm research team is focused on post-training and being scaled up.
• Lovable uses reinforcement learning on model mistakes and customer-impact prioritization.
• The platform’s usage stream is a training asset because of the million-projects-per-week volume.
• Lovable monitors margins but will not choose a cheaper model if it is worse for customers.
• Customers hit subscription caps and can top up across tiers.
• Osika estimates about 60% of lowest-tier customers top up.
• Jason argues overages are acceptable when software is worth far more than the billed amount.
• Osika says AI usually helps teams move much faster, making the spend worthwhile.
• Osika supports rapid experimentation and comparing multiple solutions rather than settling on one local optimum.
• He says humans and agents still need to decide what to build, even as generation improves.