Former Intel CEO on What Went Wrong, What's Next + Lovable CEO on the Real Promise of Vibe Coding
In a Nutshell
Pat Gelsinger argues Intel lost its edge by letting finance and business leaders prioritize buybacks over factory and process investment, enabling TSMC's foundry model and Nvidia's CUDA-driven GPU dominance to take over. Lovable's CEO shows that AI-driven "vibe coding" now lets both engineers and non-technical users ship production-grade, secure software in hours instead of months, with the platform hitting $500M ARR in 20 months by handling architecture, security, payments, and operations automatically.
These notes were generated by AI and may contain inaccuracies.
Pat Gelsinger spent 34 years at Intel, joining at age 18 and progressing through puberty at the company. He worked under deeply technical leaders including Andy Grove, Gordon Moore, and Bob Barrett, who served as mentors. When Gelsinger first joined the executive staff, approximately 15 of the 20 people in the room held PhDs. Intel was once one of America's greatest companies with the iconic "Intel Inside" campaign, but was later displaced by Nvidia, TSMC, and to some extent Apple.
A critical turning point occurred when Intel began being run by business people and finance professionals rather than technologists. Gelsinger became CEO in 2001 as the first technical leader in approximately 15 years. He notes that when business leaders are in charge, they tend to promote other business leaders. Great technology companies today are typically deeply technical and founder-led, with examples including Satya Nadella and Sundar Pichai who, while not founders, are deeply technical individuals.
In the five to six years before Gelsinger's return, Intel distributed $100 billion to shareholders through dividends and stock buybacks. The company had not built a new factory in a decade and had not purchased EUV machines. These decisions were viewed as economically rational through a spreadsheet lens but missed the technological imperative that a technologist would recognize. Gelsinger contrasts this with Apple, which also pursued significant buybacks and dividends, becoming the largest holder of capital of any company, while making only small acquisitions such as Beats for demographic access.
Steve Jobs was described as an incredible but ruthless leader. When Apple moved to Intel's Centrino chip, the company placed extraordinary demands on Intel for smaller chips and lower power consumption. When Jobs lost confidence that Intel could continue meeting those demands, he initiated an internal silicon project. Apple acquired small companies including P Semi and began building internal competency through small chip projects that gradually scaled.
Jobs had been preparing for this transition for years. During conversations about porting the operating system to Intel's x86 architecture, he revealed he had already ported the last four releases to x86, shocking Gelsinger. Jobs' philosophy was that if Intel could not remain sufficiently ahead of the industry, Apple could optimize system design with silicon design internally rather than relying on a Windows-optimized chip.
Jensen Huang initially built high-performance throughput machines that Intel dismissed as graphics cards for gamers during Intel's CPU dominance. Nvidia's development of the CUDA software stack and SIMT technology gradually improved over successive releases. A defining moment occurred when Japanese HPC researchers began using graphics cards for high-performance computing workloads, transforming GPUs from graphics devices into general-purpose computing platforms.
Nvidia benefited from serendipity as cryptocurrency and AI workloads proved ideal for their architecture, though they had not specifically predicted these applications. The "street finds its own use for technology," with the hacker community identifying optimal applications. Gelsinger had initiated a similar project at Intel called Larrabee to apply x86 to parallel computing, but the project was killed a week after his first departure from the company.
TSMC's founding vision was to become the factory for the semiconductor industry, offering manufacturing to any designer regardless of whose chip it was. This contrasted with Intel's IDM (Integrated Device Manufacturer) model, where the company never made its processes and factories available to third parties in a standardized way with PDKs and design tools.
Early in his career, Gelsinger contributed to the foundations of EDA including the first place and route, first standard cells, and first high-level description language. Intel's proprietary approach meant they maintained internal EDA tools. TSMC's model of accepting any design proved transformative. By the time Gelsinger returned to Intel in 2001, TSMC was producing five times the wafers that Intel produced. The foundry model became the dominant industry structure, with only Intel and memory companies remaining vertically integrated.
When Gelsinger returned to Intel in 2001, the US produced approximately 12% of leading-edge wafers. That figure has improved to approximately 18% through the CHIPS Act. Intel is developing as a real foundry. TSMC's factories are operating at scale in the US, alongside Samsung.
Taiwan has less than three weeks of energy reserves. A blockade causing a three-week brownout would shut down fabs that require 90 days to restart. The economic impact of a Taiwan brownout would exceed the Great Depression. China has blockaded the Taiwan Straits seven times in the last four years, with intentions appearing clear through sustained provocative exercises.
Energy capacity expansion serves as a natural constraint on AI infrastructure growth. Global energy capacity expands at 4-5% annually, while the US expanded at only 1% for a decade. No one will build data centers without available energy, creating an upper bound on how far ahead of demand the industry can build.
The value of a token, as a measure of intelligence, is potentially infinite in its ability to improve supply chains, finance, logistics, and other domains. Gelsinger is optimistic about a couple-of-decades buildout rather than a couple of years. The objective is to make AI 10,000x better by reducing the cost per token and energy per token by five orders of magnitude.
Current AI companies have real revenues and real margins, distinguishing them from dot-com era speculation. Examples include Eleven Labs at $600 million in revenue and Lovable at $500-600 million. Periodic corrections that keep earnings multiples reasonable are viewed positively, as the trajectory will not be a smooth curve. Disruptions will occur as industries are impacted by new capabilities.
The convergence of classical computing, AI computing, and quantum computing represents the next phase of excitement. Quantum computing is predicted to deliver meaningful results this decade, before 2030. Applications will include problems that cannot be computed today, such as chemistry and biology, as well as logistics optimization like the traveling salesman problem.
Encryption implications are expected around 2032-2033 with Q-day. Modalities including trapped ions, photonic, and spin approaches are all demonstrating results with proven error correction. The race is now about engineering scale. Gelsinger's portfolio company, PsiQuantum, is one contender among multiple viable approaches.
Lovable is working on two gaps: building the product and building a business around that product. The platform sees a million new projects built every single week. On the business side, the company is investing in making it easier to run a business, helping users get discovered, and supporting the full business operations whether running a small business or a large one. The platform has achieved more than 700 million visits to applications every month. There are more than 50 million apps built on the platform to date.
Lovable has been in market for 20 months. The platform is seeing both first-time founders and enterprise leaders using it. Enterprise is where Lovable is growing fastest. The platform has opinionated pieces on how to create software, how to operate that software, and how different applications within a company connect to each other over time.
Ten years ago, no-code and low-code platforms were emerging. Every 20th company in an incubator would have an MBA or non-developer who had built something using these platforms. The software did not look good, did not work perfectly, and was slow, but the promise was there. LLMs and new intelligence have made it possible to create actually good software.
Approximately 20% of Lovable users have a technical background or are engineers. Engineers appreciate that the platform is opinionated and incorporates best practices into software architecture. The platform makes it seamless to set up payments securely with one prompt and run security scans after every change, with background monitoring of projects.
Four out of five users are non-technical. Non-technical users often use the platform first to figure out what is the right thing to build, which has been a strength of Lovable. People are running businesses making more than a million dollars of revenue on the platform. There is cross-pollination where someone discovers Lovable at a large company, runs a side hustle that makes hundreds of thousands of dollars, and then becomes a founder.
A year ago, vibe coding was viewed as a way to make mockups, think about product, create wireframes, or build workable prototypes. Now users can go directly to building the actual product in a day or two. After achieving product-market fit, additional steps are required including setting up payments, implementing security, and ensuring data is not lost or leaked. These capabilities have changed dramatically over the last 12 months.
Many engineers no longer look at or write code themselves. The platform creates structure for the architecture of the software and prevents users from going off a cliff. Things like setting up payments, emails, getting discovered by AI chat engines and Google search are handled. Users do not need to know the details of how these work. The platform handles data security and secure connections to other tools. Being opinionated from day one and focused on making the platform for the 99% has been key to success.
An internal team at This Week in Startups used Lovable to build an internet for Founder University, a pre-accelerator program. Ten years ago, this would have cost $500,000. The team would not have approved the project due to budget constraints. In 4 to 8 hours, the team built the entire internet plus additional features that were not requested. The person running Founder University built it on her own without initial permission, putting it on a corporate card.
The software is now driving the program. The program exists in Saudi Arabia and Japan because of its economic impact. An economic impact tracker was then built for the 50 companies in the program, calculating how many people work at each company, taxes paid, rent, and average salaries. The platform costs $50 a month on a business plan and starts at $25. The economic impact of what was built would have cost $500,000 two years ago. It was built in 4 hours by an employee plus the cost of the software for less than $2,000 total.
Security was a concern for the Founder University internet. The team went through security review with Lovable and confirmed it is secure. Lovable is now asking penetration testers to compare all tools and verify the invisible security and trust work. Free users also receive security scanning running in the background.
A year ago the platform was at mock-ups. Now it delivers functionality that is secure and viable for deployment. Lovable has launched a new product line for the hosting part, including AI hosting and normal hosting. This product line is growing faster than the building product. It lets users run all their software and works with companies like AWS under the hood.
Lovable is being used as an AI co-founder and partner that users talk to about everything in their business. When apps and tools are on the platform, talking to Lovable provides access to all company data. The company is working with customers in pre-release to give access to a co-founder that works even when the user is sleeping and returns in the morning with strategic directions, optimizations for growing the business faster, and serving customers better.
The evolution is toward operation and intelligence, driving toward outcomes for the business. The goal is to build the software but stay to build and operate the business. Every time Lovable makes a mistake, it goes into a genetic system with engineers improving it. This compounding intelligence is applicable to customers running their business on the platform.
Internal tools could become 100% bespoke. A user at Nursa built out a new product line called Nurse Study for educating more nurses, including admin tools, scheduling, license management, and certification management. This was taken to market and also used internally in the back office, replacing more than 10 tools. They are saving more than a million dollars per year.
Bespoke solutions will increase where specific requirements are not met by existing tools. Lovable will continue to interoperate with existing tools including the Google suite, Microsoft suite, Slack, Salesforce, and HubSpot. The platform guides users through connection steps and provides an overview of data flows to ensure access is not granted to the wrong people or wrong data. Users can continue using familiar tools under the hood with a bespoke interface on top.
Lovable reached $500 million in annual recurring revenue in May. The company has been declared dead every 6 months since starting but continues growing by approximately $100 million in annual revenue every 6 months.
The strategy is to do whatever is best for customers. Lovable uses multiple models and routes requests to the model most suitable for the task. This includes commercial frontier models from multiple vendors and increasingly open weight models. When routed to their own model, it becomes more intelligent for the agent harness, especially on mistakes around which tool to call, which integration to create, and how to guide users toward business success.
The company has multiple partnerships and is investing to be close with partners. The goal is to get the fastest performance at the lowest cost. A research team in Stockholm works on post-training, applying best practices and scaling the team significantly. This is viewed as important for the European ecosystem.
Lovable looks at mistakes models make, prioritizes them by customer impact, creates datasets, and uses reinforcement learning for problems where frontier models are making mistakes. The platform has an enormous token distribution from a million new products being built every week. This provides signals for improving both the agent harness and the internal skills that help the agent remember facts from software engineers on how to build good software. Both are modified every single week.
The company monitors margins but does what is best for customers, which often means more intelligence. They have never chosen a cheaper model if it is measurably worse for customers.
There are multiple subscription tiers with caps and overages. Approximately 60% of customers from the lowest subscription tier hit caps and top up. Customers are willing to pay overages because of the value received. The comparison is that paying $600 or even $6,000 a year for a $500,000 piece of software is between 0.1% and 1% of what would have been paid three years ago.
Everything is about moving fast. More AI usually lets users move much faster, so the spend is worth it.
Multiple people in an organization may try to solve the same software problem and compete with each other. One example involved building an internet for Japan and a separate one for the US. Previously this would have resulted in one track of software with Franken software trying to accommodate all needs. Now two different people built two different Lovable projects, providing two swings at the problem.
Lovable supports rapid experimentation. At CERN, co-opetition was practiced with two isolated teams working on the same particle accelerator at different places, not sharing results until publication. This prevents getting stuck in a local minimum. Free markets work because of competition. Academia does this as well. Since engineering is less of a bottleneck and the question is more about what is the right thing to build, having sufficiently many humans attempt solving the same problem in different ways is valuable.
When doing this on Lovable, a new project can be brought up and the agent asked to check out another project, take the three things that are liked, bring them over, and potentially run an A/B split test to see if metrics improve for customers.
Lovable uses Fable as one of the models. Fable can create very sophisticated things that look really good on the first attempt. When evolving, the human still needs to think and often should plan together with the agent about what is the right thing to do. The bottleneck remains the human knowing how to use the tool, plug in the right data, and take the right decisions for moving the product and business forward.
On some tasks, more intelligence is great, such as creating beautiful 3D games. On figuring out what to build, strategic directions, or experiments to run to improve business outcomes, progress is not changing as fast.
Lovable is worth trying. Users should put it on a corporate card and start building. The economic outcome for business is extraordinary.
Keep All-In Podcast in your library
Save the videos and channels worth coming back to, and find them again in one place.
