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

2026-07-15 - 50 min - source - Read full transcript

Jason Calacanis (host)Pat GelsingerAnton Osika

Key insights

Intel's decline traces to a leadership shift from technologists to finance-driven executives.

Gelsinger says Intel's founding generation (Grove, Noyce, Moore) were deeply technical, and the executive staff he joined was mostly PhDs. When business leaders who promote other business leaders took over, hardcore technical investment decisions started getting made through a spreadsheet instead of engineering judgment, which he identifies as the root failure.

intel-decline-and-recovery

Intel returned about $100 billion to shareholders in dividends and buybacks instead of investing in capacity.

In the years before Gelsinger's return as CEO, Intel had not built a new fab in a decade and passed on manufacturing chips for the iPhone. Gelsinger argues that capital went to shareholders instead of EUV machines and new factories precisely because the economics looked bad on a spreadsheet, even though a technologist would have made the investment.

intel-decline-and-recovery

TSMC beat Intel by inventing the open pure-play foundry model Intel refused to adopt.

Intel ran as an IDM, keeping its process and fabs proprietary and never opening them to third-party designs. TSMC said it would manufacture for anyone, cut the proprietary EDA overhead in half, and rode Apple as an anchor customer to scale. By 2001, when Gelsinger returned to Intel, TSMC was already producing five times Intel's wafer volume; today it's closer to seven times.

semiconductor-supply-chain

Apple began hedging away from Intel years before it announced its own silicon.

Gelsinger recounts Steve Jobs quietly building internal chip competency through small acquisitions (P.A. Semi) starting years before the public pivot, and Jobs telling him Apple had already ported its OS to x86 across four prior releases in preparation. Jobs stopped trusting Intel to stay far enough ahead of the industry and chose to optimize system design against Apple's own silicon instead.

intel-decline-and-recovery

Taiwan has less than three weeks of energy reserves, making it a single point of catastrophic failure for global chip supply.

Citing a recent Wall Street Journal report, Gelsinger says a blockade wouldn't need a shot fired: after three weeks without oil or LNG the island browns out, and a fab that loses power doesn't restart for 90 days. He calls the economic impact of a Taiwan brownout potentially larger than the Great Depression, and notes China has run blockade exercises in the Taiwan Strait seven times over the past four years.

taiwan-geopolitical-risk

Energy capacity is the natural governor that keeps the AI buildout from becoming an unconstrained bubble.

Gelsinger argues nobody can buy GPUs and build data centers faster than the grid can supply power, and U.S. energy capacity growth was stuck near 1 percent for a decade before recently accelerating. That ceiling, in his view, bounds how far speculative buildout can run ahead of real demand.

ai-buildout-economics

Gelsinger is betting on a multi-decade AI buildout driven by Jevons paradox, not a short-lived spike.

His explicit goal as an investor is to help make AI 10,000x cheaper, dropping cost-per-token and cost-per-energy-unit by five orders of magnitude so usage explodes rather than plateaus. He frames this as a couple-decade buildout, not a couple-year one, while still expecting periodic corrections along the way as over-hyped segments (which he compares to a coming 'SaaS apocalypse') get repriced.

ai-buildout-economics

Gelsinger predicts meaningful, commercially relevant quantum computing results before 2030.

He points to multiple qubit modalities (trapped ions, photonic, spin) now independently demonstrating error correction and reasonable results, meaning modality is no longer the bottleneck, only engineering scale. He expects the 'Trinity of Computing' - classical, AI, and quantum working together - to unlock currently uncomputable chemistry and biology problems this decade, with cryptography-breaking implications following in the early 2030s.

ai-buildout-economics

Lovable has scaled to roughly $500-600 million in revenue after 20 months, driven by non-developers building real production software.

Osika reports a million new projects created weekly, more than 700 million monthly visits across apps built on the platform, and over 50 million apps live, with the fastest growth now coming from enterprise customers. About 80 percent of users are non-technical, and the platform has moved from mockup generation a year ago to production-grade software with built-in security scanning, payments, and hosting.

vibe-coding-maturity

Vibe-coded internal tools are replacing five- and six-figure enterprise software spend, often built without formal IT approval.

Calacanis describes an employee who built a full company intranet in Lovable in four to eight hours without asking permission, at a cost he estimates under $2,000 versus the roughly $500,000 it would have cost to build conventionally years earlier. Osika cites a customer at a large healthcare company who replaced more than ten internal bespoke tools this way, saving over a million dollars a year.

vibe-coding-maturity

Lovable routes work across multiple frontier and its own post-trained open-weight models, prioritizing customer outcomes over margin.

Osika says Lovable never chooses a cheaper model if it measurably underperforms for the customer, contrasting this with competitors he says are 'token dumping' - reselling tokens at a loss to appear cheap. Lovable's own models are refined weekly using reinforcement learning targeted at the specific mistakes frontier models make on Lovable's workloads, informed by its large weekly token volume.

vibe-coding-maturity

Rapid, independent parallel experimentation (co-opetition) is replacing single-track software development inside teams.

Osika describes borrowing the CERN practice of letting isolated teams solve the same problem independently before comparing results, applying it inside companies now that engineering time is no longer the bottleneck. He recommends letting multiple people build separate Lovable projects for the same problem and then cross-pollinate the best features rather than forcing one unified build from the start.

vibe-coding-maturity

Books referenced

Steve Jobs - Walter Isaacson - Gelsinger cites it while describing how ruthless and demanding Steve Jobs was as an Intel customer before Apple moved to its own silicon.

Companies

Intel - Gelsinger's employer for 34 years and twice its CEO; central subject of the decline-and-recovery discussion.
Nvidia - Discussed as the company that turned graphics chips into general-purpose computing platforms via sustained CUDA software investment.
TSMC - Won the semiconductor industry by pioneering the pure-play foundry model that Intel refused to adopt; was producing 5x Intel's wafer volume by 2001.
Apple - Moved from IDM buyback strategy to building its own silicon starting around 2008-2009 after losing confidence in Intel's roadmap.
Cerebras - Named as one of the inference-chip startups building the picks-and-shovels layer of the AI buildout.
Groq - Named alongside Cerebras as an inference-silicon company benefiting from the AI buildout.
d-Matrix - Named as another inference-chip company in the AI hardware ecosystem Gelsinger is investing behind.
Lovable - Anton Osika's AI app-building platform; the main subject of the second half of the episode.
AWS - Cited as the incumbent cloud/hosting layer Lovable is now competing with and building on top of via its new hosting product line.
Anthropic - Osika discusses using Anthropic's latest coding model as one of several frontier models Lovable routes tasks to.

Techniques and frameworks

IDM vs. foundry model - Gelsinger contrasts Intel's integrated device manufacturer approach (design and fab everything in-house) with TSMC's open pure-play foundry model that served any customer's designs.
Jevons paradox - Gelsinger's framework for why falling cost-per-token should expand AI usage and revenue rather than shrink the market, provided costs drop by orders of magnitude.
Co-opetition - Osika's CERN-derived practice of letting separate teams solve the same problem independently before comparing and merging results, to avoid getting stuck in a local optimum.

Summary

This episode splits into two back-to-back interviews recorded, judging by the closing exchange, on location near Paris. In the first half, Jason Calacanis interviews former two-time Intel CEO Pat Gelsinger on what went wrong at Intel and what he sees coming next in semiconductors and computing. Gelsinger's central diagnosis is a leadership shift: Intel's founding generation of PhDs and technologists gave way to finance-driven executives who evaluated hard technical bets through a spreadsheet rather than engineering judgment, leading the company to return roughly $100 billion to shareholders instead of building fab capacity, adopting EUV early, or opening its manufacturing to outside customers the way TSMC did. That refusal to become a foundry, combined with Apple's quiet multi-year hedge toward its own silicon (foreshadowed by Steve Jobs privately porting Apple's OS to x86 years in advance) and Nvidia's patient, unglamorous build-out of the CUDA software stack, let competitors overtake Intel's core businesses one at a time.

Gelsinger then turns to geopolitics and macro risk. He frames Taiwan's chip dominance as sitting on top of an alarming vulnerability: the island reportedly has less than three weeks of energy reserves, meaning a blockade could brown out its fabs (which take 90 days to restart once powered down) without a single shot fired - an event he says would have economic impact larger than the Great Depression. He notes China has run blockade exercises in the strait seven times in four years. On AI, he calls himself an optimist who expects a multi-decade buildout rather than a short bubble, arguing that finite energy capacity growth naturally caps how far speculative infrastructure spend can outrun real demand, and that his personal investing thesis is built around Jevons paradox: driving cost-per-token down by five orders of magnitude to unlock a much larger token economy. He closes with a prediction that quantum computing will produce commercially meaningful results before 2030, since multiple qubit modalities have independently cracked error correction and the remaining challenge is pure engineering scale.

The second half shifts to Anton Osika, founder of Lovable, in conversation with Calacanis about the state of "vibe coding." Osika reports Lovable has grown to roughly $500-600 million in revenue after 20 months, with a million new projects created weekly, over 700 million monthly visits across apps built on the platform, and more than 50 million apps live - growth now led by enterprise adoption even though 80 percent of users are non-technical. He traces a shift from a year ago, when AI tools produced impressive-looking mockups that didn't hold up, to today's production-grade software with built-in security scanning, payments infrastructure, and a new hosting product line that competes directly with AWS.

Calacanis illustrates the shift with a concrete example: an employee on his team built a full company intranet in Lovable in four to eight hours, without asking permission, for a fraction of the roughly $500,000 it would have cost to build conventionally - and then kept extending it unprompted, including an economic-impact calculator for a startup accelerator program. Osika confirms this pattern at larger scale, citing an enterprise customer that replaced more than ten internal bespoke tools this way, saving over a million dollars a year. On the model layer, Osika describes routing every task across multiple frontier models plus Lovable's own post-trained models (developed by a growing research team in Stockholm), always optimizing for customer outcome rather than cost - a contrast, he says, with competitors "token dumping" underpriced usage to fake margin. He also describes borrowing a "co-opetition" practice from his time at CERN: letting separate teams solve the same problem independently, without sharing progress, then merging the best results afterward, which he argues now works better than a single unified build now that engineering speed is no longer the bottleneck.

Notable Quotes

"Steve was an incredible leader. He was also a ruthless leader." - Pat Gelsinger

"The island of Taiwan has less than three weeks of energy reserves... you don't need a shot to be fired. You just need to say, no energy for three weeks." - Pat Gelsinger

"There has not been a time in human history where it's been better to be a technologist than the one we're in right now." - Pat Gelsinger

"It's this $500,000 piece of software... built in four hours by an employee... for less than $2,000 in a year." - Jason Calacanis

"We've never had the decision to say, let's use a cheaper model here if it's measurably worse for our customers." - Anton Osika