aipod / all-in / 2026-07-15-former-intel-ceo-on-what-went…
ALL-IN · Jul 15, 2026

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

🎙 Jason Calacanis · Pat Gelsinger · Anton Osika
29 claims5 predictions22 facts6 notable quotes raw JSON ↗

In a two-part interview taped in Paris, Jason Calacanis first sits with former Intel CEO Pat Gelsinger for a post-mortem on Intel's decline: Gelsinger argues the company lost its way once business and finance people rather than technologists ran it, returned about $100B to shareholders instead of building fabs and buying EUV machines, and missed the shifts to Apple Silicon, NVIDIA's CUDA-powered GPUs, and TSMC's foundry model. Gelsinger then lays out a macro thesis — Taiwan's fab concentration and sub-three-week energy reserves are a civilization-scale supply-chain risk; energy capacity is the natural governor that keeps the AI buildout from becoming a runaway bubble; token value as a measure of intelligence is 'somewhat infinite'; and quantum computing will deliver meaningful results before 2030 (with 'Q-day' encryption breaks by roughly 2033), converging with classical and AI compute in a 'trinity of computing.' In the second half, Lovable founder Anton Osika describes a platform seeing a million new products a week, 700M+ monthly app visits, and $500M in revenue after 20 months, serving an roughly 80% non-technical customer base by imposing opinionated architecture, security, and payments so builders 'don't go off a cliff.' Osika argues vibe coding has moved from mockups to deployable, secure production software, that bespoke apps will increasingly replace generic SaaS while interoperating with it, and that Lovable improves its own open-weight models via post-training and reinforcement learning on the mistakes frontier models make across its enormous token distribution. Both close on the human bottleneck: even with more capable models like Anthropic's Fable, deciding what to build — not writing code — is the constraint.

SEGMENTS

1:41
Intel's decline: technologists vs. business leaders

Gelsinger argues Intel went off the rails once business and finance people rather than technologists ran it, and that returning ~$100B to shareholders via dividends and buybacks starved it of capital for new fabs and EUV machines.

5:33
Apple Silicon and Steve Jobs' long game

Gelsinger recounts how Steve Jobs quietly built Apple's internal silicon competence over multiple OS releases to move off Intel and co-optimize system and chip design.

7:56
NVIDIA, CUDA, and GPUs becoming general-purpose compute

Gelsinger describes how Intel scoffed at Jensen Huang's GPUs until CUDA and HPC turned them into general-purpose compute, and how Intel's competing Larrabee project was killed a week after his first departure.

11:32
TSMC and the rise of the foundry model

Gelsinger explains how TSMC's foundry vision beat Intel's integrated IDM model to become the semiconductor industry standard, out-producing Intel 5-7x on wafers, and why Intel's comeback requires becoming a foundry.

14:11
Onshoring chips, the CHIPS Act, and Taiwan blockade risk

Gelsinger credits the CHIPS Act with raising US leading-edge share from ~12% to ~18%, then warns that Taiwan's sub-three-week energy reserves make a Chinese blockade a brownout risk with economic impact greater than the Great Depression.

17:28
Is AI a bubble? Energy as the governor and token economics

Gelsinger argues energy capacity is a hard upper bound that keeps the AI buildout from overheating, and that the value of a token as a measure of intelligence is 'somewhat infinite,' targeting a 10,000x reduction in cost and energy per token.

21:21
AI valuations, healthy corrections, and the trinity of computing

Gelsinger contrasts today's real-revenue AI companies with dot-com speculation, welcomes periodic corrections, and points to the convergence of classical, AI, and quantum computing — predicting meaningful quantum results before 2030.

25:30
Lovable: mission, growth, and the customer for vibe coding

Anton Osika frames Lovable's mission as empowering humans, citing a million new products a week, 700M+ monthly app visits, 50M+ apps, 20 months in market, and a customer base that is roughly 80% non-technical.

29:17
From mockups to production software: the intranet story

Jason recounts an employee building a full Founder University intranet on Lovable in about four hours for under $2,000 versus a ~$500K legacy cost, while Osika describes Lovable handling architecture, payments, and security for non-technical builders.

36:29
Bespoke software replacing SaaS incumbents

Osika's Nursa story (replacing 10+ internal tools, saving $1M+/yr) and his view that bespoke apps will grow for specific requirements while Lovable keeps interoperating with Salesforce, HubSpot, Slack, Google, and Microsoft.

39:40
Lovable's model strategy, token economics, co-opetition, and Fable

Osika describes multi-model routing toward open weights, a Stockholm post-training/RL team mining frontier-model mistakes, caps with overages (~60% of low-tier users top up), the CERN 'co-opetition' model, and his read on Anthropic's Fable.

NOTABLE QUOTES

“when you're making these hardcore technical decisions that affect billions of dollars, you don't do that through a spreadsheet.”
Pat Gelsinger · 3:01 — Crystallizes Gelsinger's entire Intel post-mortem — technology businesses need technologists, not finance leaders, making the calls.
“The economic impact of a brownout of Taiwan is greater than the Great Depression in the world.”
Pat Gelsinger · 15:13 — The starkest framing of the Taiwan semiconductor supply-chain risk — a blockade needing no shots fired.
“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 · 20:38 — The optimistic capstone of Gelsinger's AI-buildout thesis, set against his bubble caution.
“This decade.”
Pat Gelsinger · 22:48 — His blunt, falsifiable answer to when quantum computing will finally do something meaningful.
“many engineers, they don't look at the code, they don't write code anymore, and that means that you don't need to be an engineer to create software.”
Anton Osika · 30:09 — The thesis of the vibe-coding half — software creation decoupled from writing code.
“I'm still paying somewhere between 0.1% and 1% of what I would have paid three years ago. Who cares?”
Jason Calacanis · 44:13 — The economic-leverage argument for why token overages are worth paying.

PREDICTIONS ON THE RECORD

Dated and falsifiable — full record at /api/episode/<id>.

Pat Gelsinger · a couple of decades · 18:41
The AI buildout will be a couple-of-decades-long expansion, not a couple of years.
“I am an optimist that we're in a couple of decade build out.”
Pat Gelsinger · unspecified · 18:41
The cost and energy per token will drop by roughly five orders of magnitude, making AI about 10,000x better and vastly more accessible.
“We want to drop by five orders of magnitude, the cost per token, the energy per token”
Pat Gelsinger · before 2030 · 23:47
Quantum computing will deliver meaningful results across multiple industries before 2030.
“my prediction is meaningful results before 2030”
Pat Gelsinger · 2028-2033 · 23:15
Quantum computers will break encryption ('Q-day') roughly in the 2028-2033 window.
“it's probably going to be 2028, 2032, 2033, when we solve things like encryption, where you'll have the fundamental Q day kind of implications.”
Anton Osika · within a year · 35:28
Within about a year, Lovable will offer an 'AI co-founder' that works while you sleep and returns strategic directions and optimizations each morning.
“give them access to a co-founder that works for you even when you're sleeping. And it comes back to you in the morning and says like, here are some strategic directions you could go”

WHO CLAIMED WHAT

Every position taken in the episode, attributed to its holder, with conviction (0–1) and the verbatim receipt.

Pat Gelsinger — 15 claims

opinionconviction 0.852:08
Intel's fundamental failure was being run by business and finance people rather than technologists, since business leaders promote other business leaders and hollow out technical decision-making.
“one of the things that went off the rail was when it started to be run by business people.”
opinionconviction 0.853:01
Hardcore technical decisions that affect billions of dollars cannot be made through a spreadsheet, so a technology business fundamentally needs technologists running it.
“when you're making these hardcore technical decisions that affect billions of dollars, you don't do that through a spreadsheet.”
opinionconviction 0.83:40
Returning ~$100B to shareholders in the years before he came back was a mistake; that capital should have stayed on Intel's balance sheet to build fabs and buy EUV machines.
“What I wouldn't have done for another $100 billion on the balance sheet.”
assertionconviction 0.85:33
Steve Jobs deliberately built Apple's internal semiconductor competence over years to reduce reliance on Intel and co-optimize system and silicon design.
“He had been preparing the core technologies inside of Apple for something that might happen in the future.”
opinionconviction 0.68:37
Killing Intel's Larrabee project — an attempt to make x86 do what NVIDIA's GPUs did — was a pivotal error; had it survived, the industry would look very different.
“the project was killed a week after I left. And the world would have been so much different.”
assertionconviction 0.8511:32
TSMC's foundry model became the model of the entire semiconductor industry, displacing Intel's integrated design-and-manufacturing approach except for Intel itself and memory.
“that model of foundry became the model of the semiconductor industry, with two exceptions, Intel and memory.”
opinionconviction 0.811:32
A core thesis of Intel's comeback strategy is that Intel must become a foundry serving third parties, not just an integrated manufacturer for its own chips.
“We must become a foundry as well.”
assertionconviction 0.815:13
A Chinese blockade of Taiwan could brown out the island and its fabs without a shot being fired, with economic impact greater than the Great Depression.
“The economic impact of a brownout of Taiwan is greater than the Great Depression in the world.”
opinionconviction 0.817:10
The US needs far more resilient semiconductor supply chains because China's intentions toward Taiwan have been clear over a sustained period.
“Their intentions have been clear over a sustained period of time. We need more resilient supply chains for it.”
opinionconviction 0.818:28
Energy capacity is a hard upper bound on the AI buildout — nobody builds data centers or buys GPUs without energy — which guarantees the AI bubble can't get too far ahead of itself.
“there is a silver lining here that guarantees we don't get too far ahead of ourselves in terms of bubble, and that is energy capacity.”
opinionconviction 0.7518:41
The value of a token as a measure of intelligence is 'somewhat infinite,' so the economic potential AI unleashes is vast, especially given developed-world labor shortages.
“the potential value that we unleash in a token economic world is somewhat infinite, particularly with labor shortages”
opinionconviction 0.8520:38
There has never been a better time in human history to be a technologist or to be alive, because AI will let us solve chemistry, language, materials, cancer, and poverty.
“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.”
opinionconviction 0.7521:45
Unlike dot-com speculation, today's AI companies have real revenues and margins; periodic corrections that keep multiples reasonable are healthy because the trajectory won't be a smooth curve.
“periodic corrections that keep the multiple, you know, earnings multiples and so on, and reasonable things is good because this will not be a smooth curve.”
opinionconviction 0.721:45
The most exciting frontier is the 'trinity of computing' — classical, AI, and quantum computing converging.
“what I call the trinity of computing, classical computing, AI computing, and quantum computing. And when those three come together, okay, that's when things get really exciting.”
assertionconviction 0.7523:15
Quantum error correction has now been proven across multiple modalities, so the remaining challenge is engineering scale, not the underlying physics.
“We know how to build qubits. We know how to error correct qubits. We now have algorithmics against quantum, and now it's just about the engineering scale.”

Anton Osika — 10 claims

assertionconviction 0.830:09
Many engineers no longer read or write code, which means you no longer need to be an engineer to create software.
“many engineers, they don't look at the code, they don't write code anymore, and that means that you don't need to be an engineer to create software.”
opinionconviction 0.830:09
Lovable's core value is imposing opinionated architecture and best practices — payments, security, discoverability — so non-technical builders don't go off a cliff.
“it creates a structure for the architecture of the software that you build, and it makes sure that you don't go off a cliff”
opinionconviction 0.830:09
Being opinionated from day one and focusing on the non-technical '99%' is what made Lovable successful.
“us being opinionated from day one and being focused on making this for the 99%.”
opinionconviction 0.737:46
Where existing tools don't fit specific requirements, companies will increasingly build bespoke solutions, even as Lovable continues to interoperate with the incumbent SaaS tools underneath.
“In those cases, I think yes, you will have more bespoke solutions.”
assertionconviction 0.840:22
Lovable's strategy is to always do whatever is best for the customer, which means routing every request to the most suitable model across multiple commercial and open-weight vendors.
“we're using multiple models. And so if you ask Lovable now, it's actually routed to the model that's most suitable to whatever you want to do.”
opinionconviction 0.7543:23
Lovable never picks a cheaper model when it is measurably worse for customers, because moving fast with more intelligence is usually worth the spend.
“We've never had the decision to say, let's use a cheaper model here if it's measurably worse for our customers.”
assertionconviction 0.7542:06
Lovable improves its own models by mining the specific mistakes frontier models make across its enormous token distribution and applying reinforcement learning to those failures.
“we're looking at the mistakes that any of the models do right now, and then we prioritize them by what drives most impact for our customers, and then we make the models, we create datasets, we did something called reinforcement learning specifically for the problems where the frontier models are making mistakes for us right now.”
opinionconviction 0.7545:54
Now that engineering is no longer the bottleneck, running many isolated teams at the same problem ('co-opetition') is a good way to avoid local minima and find the best solution.
“now since the engineering is less of the bottleneck, it's more the question of what is the right thing to build. I think it's a great thing to have if you have the sufficiently many humans right to do to try to attempt the solving the same problem in different ways.”
opinionconviction 0.747:34
Anthropic's Fable can produce very sophisticated, good-looking output on the first attempt, especially 3D games.
“it can in the first attempt, create very sophisticated things that look really good.”
opinionconviction 0.848:01
Even with more capable models like Fable, the binding constraint is the human work of deciding what to build and plugging in the right data — that is not changing as fast as raw model capability.
“on figuring out what to build, figuring out what are the right strategic directions or experiments you should run to improve the outcomes for your business, that's not changing as fast.”

Jason Calacanis — 4 claims

opinionconviction 0.64:37
Apple squandered the opportunity of its enormous capital hoard by only doing small, marginal acquisitions instead of big transformative bets.
“They buy little, tiny acquisitions on the margins.”
assertionconviction 0.834:53
Vibe coding has leapt in a year from producing mockups and wireframes to building deployable, secure production software directly.
“A year ago, we were at mockups. Now we're at functionality and secure and super viable for deployment.”
opinionconviction 0.636:56
Foundational SaaS like Slack could be replaced by bespoke vibe-coded software integrated more deeply into a company's operations.
“maybe I should vibe code my own Slack.”
opinionconviction 0.844:13
Paying large token overages is rational because you are still paying only 0.1%-1% of what the equivalent software would have cost a few years ago.
“I'm still paying somewhere between 0.1% and 1% of what I would have paid three years ago. Who cares?”

QUERY THIS EPISODE

agent access
# full structured record
curl 'https://aipod-two.vercel.app/api/episode/all-in-2026-07-15-former-intel-ceo-on-what-went-wrong-what-s-next-lovable-ceo-'

# or over MCP
aipod_episode { "id": "all-in-2026-07-15-former-intel-ceo-on-what-went-wrong-what-s-next-lovable-ceo-" }