Gelsinger Warns Taiwan Blockade Triggers Global Great Depression
All-In Podcast
A three-week Chinese blockade of Taiwan would deplete the island's energy reserves and shut down semiconductor fabrication plants for 90 days, triggering a global economic collapse worse than the Great Depression.
## Background
Published on July 15, 2026, on YouTube by the All-In Podcast, this episode features host Jason Calacanis interviewing former Intel CEO Pat Gelsinger and Lovable founder Anton Osika. The page description outlines an in-depth conversation covering what went wrong culturally at Intel, the geopolitical stakes of Taiwan's semiconductor dominance, and how Lovable's "vibe coding" platform is radically slashing software development costs.
## TL;DR
The episode explores the critical shift in technology leadership, tracing how Intel's transition from technical founders to finance-driven executives led to missed opportunities in EUV lithography and mobile silicon, while Nvidia and TSMC surged ahead. Pat Gelsinger warns of the catastrophic global economic impact of a potential Taiwan blockade and outlines a future defined by energy-constrained AI and quantum computing by 2030. Finally, Anton Osika demonstrates how Lovable has scaled to $500 million in annual revenue by enabling non-technical users to build, host, and run highly secure, bespoke enterprise-grade applications at a fraction of traditional engineering costs.
## Atmosphere & style
The conversation carries a reflective yet highly charged intellectual energy, balancing sober geopolitical and corporate warnings with infectious technological optimism. Gelsinger speaks with the seasoned, candid authority of a veteran engineer who has lived through Silicon Valley's golden eras, bringing a mix of regret for Intel's past missteps and urgent passion for hardware sovereignty. When Osika joins, the vibe shifts into an exciting, fast-paced exploration of the future of software, characterized by mutual enthusiasm between the host and guest as they marvel at the democratization of coding.
## Key takeaways
- [00:00](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=0s) Intel's historical decline stemmed from a cultural shift where technical leadership was replaced by business and finance executives who prioritized short-term financial returns over long-term capital investments.
- [03:21](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=201s) Before Gelsinger's return, Intel returned $100 billion to shareholders through dividends and buybacks instead of investing in new fabrication plants or critical extreme ultraviolet (EUV) lithography technology.
- [05:12](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=312s) Apple's transition to custom silicon was a highly strategic, long-term effort planned by Steve Jobs years in advance, starting with the quiet acquisition of companies like P.A. Semi to build core chip capabilities.
- [07:56](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=476s) Nvidia’s dominance in AI and GPU computing was built on persistent software innovation through the CUDA ecosystem, whereas Intel made a major strategic error by canceling its parallel Larrabee throughput project shortly after Gelsinger's initial departure.
- [11:00](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=660s) TSMC's pure-play foundry model succeeded by standardizing interfaces and serving as a shared industry factory, rapidly outpacing Intel's internal-only manufacturing scale to achieve a 7-to-1 wafer volume advantage.
- [14:11](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=851s) A potential Chinese blockade of Taiwan poses a severe risk to the global economy; cutting off Taiwan's limited energy reserves for just three weeks would cause a power loss, shutting down semiconductor fabs and triggering a fallout larger than the Great Depression.
- [17:27](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1047s) While global energy capacity is growing, the stagnation of the US electrical grid acts as a physical ceiling and upper bound on the rapid expansion of physical AI infrastructure.
- [19:08](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1148s) Gelsinger aims to make AI 10,000 times more efficient to trigger Jevons Paradox, where dramatically lowering the cost and energy per token will exponentially drive up global consumption of AI.
- [20:53](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1253s) Unlike the speculative dot-com era, today's leading AI startups are generating substantial and real revenues, exemplified by companies like ElevenLabs and Lovable.
- [23:11](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1391s) Practical, industry-wide quantum supremacy and utility across hard problems like encryption are projected to be achieved within this decade, specifically by 2030 to 2033.
- [25:00](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1500s) Lovable has seen explosive adoption for its 'vibe coding' platform, supporting 50 million apps built, 1 million new weekly projects, and an 80% non-technical user base within 20 months of launch.
- [31:12](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1872s) AI-assisted development on platforms like Lovable has dramatically altered software economics, enabling non-technical users to build sophisticated internal tools in hours for under $2,000, which previously would have cost $500,000.
- [36:56](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=2216s) Bespoke, custom-built applications created through AI are starting to replace expensive back-office SaaS tools and internal legacy enterprise systems, saving companies millions of dollars.
- [39:41](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=2381s) To protect itself against rapid shifts in frontier models, Lovable employs a dynamic routing architecture that balances tasks across different commercial and proprietary open-weight models.
- [41:10](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=2470s) Lovable officially reached a massive financial milestone of $500 million in annual revenue in May 2026, supported by a specialized research team in Stockholm focusing on post-training methodologies.
- [43:01](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=2581s) Despite usage caps, roughly 60% of users on Lovable's lowest subscription tier pay overages for token top-ups because the ROI of building highly valuable software at low cost remains extremely high.
- [47:08](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=2828s) Modern AI capabilities have shifted the bottleneck of software engineering away from raw code generation to the strategic human-to-agent planning process.
## Detailed walkthrough
### [00:00](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=0s) What Went Wrong at Intel: Technologists vs. Bean Counters
The discussion begins with a retrospective on Intel's historic trajectory. The host notes that Intel—having spent "only 34 years" under the guest's initial tenure—was once arguably "one of the greatest American companies ever," but eventually went off the rails and got "demolished" by Nvidia, TSMC, and to a degree, Apple. The host nostalgically recalls the iconic "Intel Inside" campaign and the Pentium processor era, questioning what went wrong during the period when the guest stepped away before returning as CEO.
Pat Gelsinger reflects on joining Intel at age 18, joking that he "went through puberty at Intel" under legendary, deeply technical mentors like Andy Grove, Robert Noyce, Gordon Moore, and Craig Barrett. He recalls his first time joining the executive staff, where 15 out of 20 people in the room held PhDs. Gelsinger diagnoses the primary cultural shift that derailed Intel: the transition from being run by deeply technical leaders to being run by "business people," "bean counters," and "finance people." When Gelsinger became CEO in 2021, he was the first technical leader to hold the position in essentially 15 years. He argues that business leaders naturally promote other business leaders, whereas today's greatest tech companies—even those not led by founders, like Microsoft under Satya Nadella or Alphabet under Sundar Pichai—are led by deeply technical individuals.
Gelsinger asserts that making high-stakes, multi-billion-dollar technical decisions through a spreadsheet is a "lousy" approach unless the underlying technology trends justify the investment. He points out that in the five to six years before his return, Intel returned $100 billion to shareholders through dividends and stock buybacks, yet failed to build a single new factory for a decade or invest in extreme ultraviolet (EUV) lithography machines. A technologist would have made those capital investments because they understood the long-term roadmaps, even if the short-term spreadsheet economics looked unfavorable.
* **Mentors Mentioned:** Andy Grove, Gordon Moore, Robert Noyce, Craig Barrett.
* **Capital Allocation Failure:** Intel distributed $100 billion in dividends and buybacks in the five to six years prior to Gelsinger's return, rather than investing in new fabrication facilities or EUV technology.
---
[05:12](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=312s) The Apple Silicon Divergence
The host draws a comparison to Apple, noting that while Apple also engaged in massive buybacks and dividends, they made strategic, highly integrated moves rather than wasting time on minor acquisitions. The host asks about Steve Jobs’ decision around 2008–2009 to design Apple's own silicon, inquiring whether it was a covert project or if Intel was aware of it.
Gelsinger characterizes Steve Jobs as an "incredible" but "ruthless" and highly demanding leader. When Apple transitioned Mac computers to Intel chips (utilizing the Centrino platform), they pushed Intel constantly to make the silicon smaller and lower-power. Once Jobs became unconvinced that Intel could maintain that aggressive pace, he quietly began building internal capability. Apple acquired P.A. Semi and other small companies to build core competencies, starting with small chips before scaling up.
Gelsinger shares a telling anecdote about discussing Apple's transition from PowerPC chips to x86 with Jobs. Gelsinger proudly offered Intel's compiling and operating system expertise to help Apple port their OS, to which Jobs replied, "I've been working on that for the last four releases." Jobs had been preparing Apple’s core software for years in anticipation of a hardware shift. This long-term preparation allowed Apple to eventually design their own custom silicon, optimized directly for iOS and macOS rather than relying on Intel’s general-purpose silicon optimized for Windows.
---
[07:56](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=476s) Nvidia, Larabee, and the Rise of GPU Computing
The conversation shifts to Nvidia and its CEO, Jensen Huang. The host asks whether Nvidia’s massive success—becoming the backbone for cryptocurrency mining and AI workloads—was due to luck, skill, or a combination of both.
Gelsinger explains that Nvidia's trajectory was a masterclass in continuous innovation. Initially, Intel scoffed at Nvidia’s throughput graphic machines, viewing them as niche hardware for gamers compared to Intel's dominant CPUs. However, Nvidia steadily built out a robust software ecosystem, introducing CUDA and SIMT (Single Instruction, Multiple Threads) technology, improving it release by release. A defining moment occurred when Japanese High-Performance Computing (HPC) researchers realized they could leverage these dense graphic cards for complex scientific calculations. This shifted Nvidia's focus from pure graphics to general-purpose GPU computing.
Meanwhile, AI was emerging from its "fifth nuclear winter." While many in the industry doubted AI would ever achieve a mainstream breakthrough, the hacker and academic communities kept developing on top of Nvidia's compounding CUDA software stack. Gelsinger reveals that he had championed a parallel project at Intel called Larrabee, which aimed to use x86 architecture for many-core throughput computing. Highly illustratively of Intel's cultural missteps, the Larrabee project was killed just one week after Gelsinger first departed the company. "The world would have been so much different," Gelsinger reflects, had Intel sustained that long-term research and development.
---
[11:00](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=660s) The Foundry Model: How TSMC Changed the World
The discussion moves to TSMC (Taiwan Semiconductor Manufacturing Company) and how they overtook Intel in manufacturing. Gelsinger explains that TSMC pioneered a pure-play foundry vision, aiming to be the shared "factory for the industry" to offset the extreme $20 billion to $30 billion cost of modern semiconductor fabrication.
In contrast, Intel operated strictly as an Integrated Device Manufacturer (IDM), keeping its process technologies and factories proprietary for its own CPUs. Intel lacked standardized Process Design Kits (PDKs) and electronic design automation (EDA) tools for third parties, relying instead on highly proprietary, in-house software. TSMC, however, standardizing the interface, essentially said, "I don't care whose chip it is... I'll be your manufacturing partner."
With Apple as an anchor customer driving TSMC's manufacturing scale, the foundry model rapidly outpaced Intel's internal-only model. By the time Gelsinger returned to Intel in 2021, TSMC was producing five times (5x) the wafer volume of Intel—a gap that has since widened to roughly seven to one (7x). This dramatic shift convinced Gelsinger that Intel had to pivot to become a world-class foundry as well.
---
[14:11](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=851s) Geopolitics and the Vulnerability of Taiwan
The host addresses the geopolitical realities surrounding Taiwan and the semiconductor supply chain, highlighting the US CHIPS Act and asking whether the US can onshore manufacturing quickly enough to prevent a global catastrophe if China were to blockade Taiwan.
Gelsinger outlines three critical dynamics:
1. **The CHIPS Act is working:** When the legislative push began around 2021, the US manufactured only about 12% of leading-edge semiconductors; by 2026, that figure has risen to roughly 18%.
2. **Foundry progress is real:** Intel is making meaningful progress toward operating as a commercial foundry, alongside TSMC and Samsung building out US-based capacity.
3. **The "ugly" reality of Taiwan's vulnerability:** Gelsinger references a striking Wall Street Journal report highlighting that the island of Taiwan maintains less than three weeks of energy reserves.
Gelsinger notes that a Chinese blockade would not even require a single shot to be fired; simply cutting off oil and liquefied natural gas (LNG) imports for three weeks would cause the island to brown out. In semiconductor manufacturing, turning off a fab due to power loss means it cannot safely return to operation for at least 90 days. Gelsinger warns that the economic fallout from a Taiwan blackout would be "greater than the Great Depression." This extreme vulnerability underscores the urgent global need to build resilient, geographically diversified supply chains.
[16:54](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1014s) Resilient Supply Chains and the Looming Threat over Taiwan
The conversation shifts to geopolitical realities and their massive impacts on global supply chains, specifically regarding China and Taiwan. Pat Gelsinger underscores that a Taiwan blockade is not merely a theoretical scenario for the future. Over the last four years, China has blockaded the Taiwan Strait seven times through highly provocative military exercises.
> "This isn't a theory... their intentions have been clear over a sustained period of time."
Gelsinger notes that whether a full-scale conflict or disruption occurs in 2027, 2030, or 2035, the trajectory demands that the United States build far more resilient supply chains. While progress is being made, Gelsinger urges that the transition to local manufacturing must go faster and become more meaningful to protect the global economy.
---
[17:27](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1047s) The AI Buildout, Energy Caps, and Jevons Paradox
Addressing the massive infrastructure buildout supporting artificial intelligence, the host asks Gelsinger if the industry is currently experiencing a bubble, comparing it to historical tech booms like the PC, server, and internet revolutions. The concern is whether companies are overspending on GPUs and data centers without generating the necessary return on investment (ROI) to justify their valuations. Gelsinger responds with a unique perspective, noting that there is a built-in physical ceiling preventing the bubble from growing too far out of control: global energy capacity.
While global energy capacity is growing at about 4% to 5% annually, the U.S. electrical grid suffered from a "hideous" decade-and-a-half period of stagnation, growing at just 1% per year. Because nobody will buy GPUs if they cannot plug them in, energy availability acts as an ultimate upper bound on AI hype. Beyond this constraint, Gelsinger is a long-term optimist, predicting a multi-decade AI buildout rather than a brief spike. He has set a personal and industry goal to make AI $10,000\times$ more efficient. By dropping the cost and energy per token by five orders of magnitude, the industry will trigger Jevons Paradox—where making a resource cheaper and more efficient dramatically explodes its overall consumption.
- **Energy Grid Growth**: Global energy capacity is growing at 4% to 5%, whereas the U.S. sat at roughly 1% growth for a decade.
- **AI Efficiency Goal**: A target of making AI $10,000\times$ better, reducing token costs by 5 orders of magnitude (a factor of $10^5$).
- **Technological Optimism**: Gelsinger highlights specialized silicon companies like Cerebras, Groq, and dMatrix that are rapidly driving down the costs of running inference.
---
[20:53](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1253s) High Valuations, Market Corrections, and the Trinity of Computing
The discussion pivots to the staggering valuations of modern AI startups. Unlike the speculative dot-com bubble of 1999–2000, Gelsinger points out that today's leading AI companies are generating "real revenues" and "real margins." For instance, companies like ElevenLabs and Lovable are already booking hundreds of millions of dollars in revenue. While some market corrections are inevitable and even healthy to keep earnings multiples grounded, Gelsinger views these disruptions as natural bumps on a twenty-year upward trajectory.
Looking further ahead, Gelsinger shares his excitement for what he calls the "trinity of computing"—the convergence of classical computing, AI computing, and quantum computing. When asked when quantum computing will finally move past its perpetual "five years away" status to deliver practical utility, Gelsinger confidently declares it will happen "this decade," specifically by 2030.
- **Startup Revenues**: Companies like ElevenLabs and Lovable are cited as having real revenues in the $500M to $600M range, signaling true market demand compared to the dot-com era.
- **Quantum Supremacy Timeline**: Anticipated meaningful results and quantum supremacy across multiple industries by 2030, with hard problems like encryption and Q-day implications being solved around 2032 or 2033.
- **Quantum Modalities**: Success is being proven across multiple quantum approaches, including trapped ions, photonic systems, and silicon spin qubits.
---
[25:00](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1500s) Lovable's Explosive Growth and the Vibe Coding Revolution
Anton Osika, the founder of Lovable, joins the conversation to discuss how his platform is redefining software development. Since launching 20 months ago, Lovable has seen extraordinary adoption, scaling to over 50 million applications built on the platform and generating more than 700 million visits to these apps every month. Osika explains that Lovable's mission is focused on "empowering humans" by bridging two major gaps: building a product and building a business around it.
Unlike the "no-code/low-code" wave of ten years ago, which often resulted in slow, poorly designed software that required extensive workarounds, the integration of LLMs has enabled the creation of high-quality, production-ready software. Lovable has achieved rapid traction in the enterprise market because it is highly opinionated, automatically structuring code architecture, enforcing security scans, and setting up payments seamlessly.
- **Platform Stats**: Over 50 million apps built, 1 million new projects created weekly, and over 700 million monthly visits across those apps within 20 months of launch.
- **User Demographics**: 80% (four out of five) of Lovable's users are non-technical, while the remaining 20% are engineers who appreciate the platform's architectural best practices.
- **Economic Power**: Users are now running fully-fledged businesses generating over $1 million in revenue directly on Lovable.
---
[31:12](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1872s) High-Leverage Vibe Coding: A Case Study in "Boss-to-AI" Prompting
The host shares a personal story of how his own team leveraged Lovable to build a fully functional internal portal for Founder University—their startup pre-accelerator. In the past, developing a custom intranet of this caliber would have cost upwards of $500,000 and taken months of developer resources. Instead, a non-technical program manager built the entire system in just four to eight hours on her own initiative, putting the subscription on her corporate card.
The software has since become the core engine driving their global accelerator programs in countries like Saudi Arabia and Japan. When the host wanted to visualize the economic impact of the 50 participating startups, the manager was able to quickly prompt Lovable to build a complex calculator. By factoring in variables like local tax rates, rental costs, employee headcount, and average salaries, the platform produced a sophisticated tool that would have previously been financially unjustifiable to build.
> "She asked Lovable to do it. I gave her some, you know, prompting—human prompting, boss-to-AI... and it built something that I would have never been able to afford to build."
[33:32](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=2012s) Lovable's Evolution and the True Cost of Software
The economics of modern "vibe coding" are radically altering the financial landscape of software development. Jason highlights that a business plan for Lovable starts around $25 to $50 a month, yet the economic output is staggering. A custom corporate intranet that Jason's team built in just four hours would have easily commanded a $500,000 price tag and months of engineering effort just two years ago. Factoring in employee hours and software fees, the entire tool was deployed for under $2,000.
As these tools shift from simple mock-ups to highly functional, deployable applications, security remains a top priority. When Jason raised concerns about potential data leaks, Anton reassured him that Lovable's team rigorously addresses security and trust behind the scenes. Lovable actively encourages penetration testing to compare its security standards against other tools, providing continuous background security scanning even for free-tier users.
[35:00](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=2100s) Lovable as an AI Co-founder and Hosting Platform
Lovable's vision for the next year is to bridge the gap between building a product and actually running a business. To achieve this, they have introduced a dedicated hosting product line that manages both AI workloads and traditional application hosting. While they collaborate with infrastructure giants like AWS under the hood, Lovable is increasingly viewed by its customers as an "AI co-founder."
By hosting applications directly on the platform, Lovable gains a holistic view of a company's data. In pre-release trials, the platform acts as an automated partner that works around the clock. Business owners can wake up to tailored strategic directions, growth optimizations, and customer service recommendations generated overnight. Rather than just building software, the goal is to help founders operate and scale their entire business, leveraging an agentic system where engineering feedback loops continuously correct mistakes and compound platform intelligence.
[36:56](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=2216s) Replacing Enterprise Software with Bespoke Applications
The rise of AI-assisted development raises a profound question: Will bespoke, custom-built software eventually replace massive enterprise platforms like Slack, Salesforce, HubSpot, Google Suite, and Microsoft Office? Jason notes that while a $10,000 annual Slack bill is well worth the cost, the temptation to "vibe code" a highly integrated, custom version of such tools is becoming increasingly attractive.
Anton shares a compelling story about a user named Ned at Nursa, a large healthcare platform. Eager to build out a new product line called "Nurse Study" to educate and license nurses, Ned used Lovable to build all the necessary administrative, scheduling, and certification management tools. Beyond client-facing products, Ned brought Lovable into Nursa's back office, replacing over ten bespoke internal tools and saving the company over $1 million per year. While specialized requirements will drive more companies to build bespoke software, Anton emphasizes that Lovable is designed to interoperate seamlessly with existing giants like Google, Microsoft, and Slack, guiding users step-by-step through secure data-flow integrations.
[39:41](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=2381s) Model Routing, Open-Weight Models, and Stockholm's Research Team
Lovable’s underlying architecture is designed to put the customer first by utilizing a mixture of multiple AI models. Rather than relying on a single vendor, Lovable dynamically routes tasks to whichever model is most suited for the job, utilizing both commercial frontier models and open-weight models. When a query is routed to Lovable's in-house models, it trains their agentic harness to be more intelligent, specifically learning from past mistakes regarding tool calling and integration steps.
This customer-first approach is fueling explosive financial growth. Despite critics claiming Lovable is "dead" every six months when new frontier models launch, the company has scaled aggressively, officially hitting $500 million in annual revenue in May 2026. This rapid growth is supported by a robust research team based in Stockholm, Sweden. The team focuses heavily on post-training methodologies, aiming to bolster the broader European tech ecosystem with advanced AI capabilities.
- **Revenue Milestone:** Lovable surpassed $500 million in annual revenue in May 2026.
- **Dynamic Routing:** Lovable routes user requests across commercial frontier models and proprietary open-weight models depending on task suitability.
- **Stockholm R&D:** The Swedish research team is actively scaling up to focus on advanced post-training techniques.
[43:01](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=2581s) Token Economics, Subscription Caps, and "Co-opetition"
With a million new products being built on the platform, Lovable burns through an enormous volume of tokens. Anton notes that they refuse to use cheaper, inferior models if it degrades the user experience, prioritising intelligence over margin-cutting. While they offer multiple subscription tiers with usage caps, customers are highly willing to purchase top-ups. In fact, roughly 60% of customers on the lowest subscription tier regularly hit their caps and pay overages because the value of generating a multi-hundred-thousand-dollar piece of software for a fraction of the cost makes the extra token spend trivial.
This speed of execution is also changing how organizations approach software development. Jason shares an anecdote where two separate team members built different versions of an intranet—one for Japan and one for the US—using Lovable. In traditional software engineering, this duplication would be avoided to prevent a bloated "Franken-software" mess. However, Anton is a major proponent of this rapid, parallel experimentation. Drawing on his experience working at the CERN particle physics laboratory, Anton explains the concept of "co-opetition," where isolated teams work on different parts of the same particle accelerator and only share results at the end. This structure prevents teams from getting stuck in "local minima." Because engineering is no longer the bottleneck, companies can easily run parallel projects on Lovable, run split tests, and merge the best features of both into the final product.
- **Overage Adoption:** Approximately 60% of users on Lovable's lowest subscription tier choose to pay for token top-ups to keep building.
- **Co-opetition:** Parallel development on Lovable allows businesses to test multiple software variations simultaneously without the friction of traditional engineering bottlenecks.
[47:08](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=2828s) The Impact of Claude and the Bottleneck of Human Strategy
The conversation turns to the capabilities of the latest AI models, such as Anthropic’s Claude (transcribed as Fable), which developers have famously used to build complex 3D games and functional clones of games like Fortnite on the first attempt. Anton confirms that Lovable integrates Claude as part of its model suite, noting that while the model is a massive step function in creating sophisticated, highly polished visual outputs on the first try, the core bottleneck of development remains unchanged.
Ultimately, the bottleneck is not the generation of code, but the human-to-agent planning process. A human developer must still work closely with the AI to determine the right strategic directions, figure out what to build, and run the right experiments to drive real business outcomes. Jason closes the session with a strong endorsement, praising Lovable's incredible business outcomes and celebrating Anton's ability to "survive and thrive" through every major shift in the AI landscape.
## Notable quotes
> [16:54](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1014s) This isn't a theory... their intentions have been clear over a sustained period of time. — Pat Gelsinger
> [31:12](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1872s) She asked Lovable to do it. I gave her some, you know, prompting—human prompting, boss-to-AI... and it built something that I would have never been able to afford to build. — Jason
## Mentioned
Intel, Nvidia, TSMC, Apple, Pat Gelsinger, Andy Grove, Robert Noyce, Gordon Moore, Craig Barrett, Satya Nadella, Microsoft, Sundar Pichai, Alphabet, Steve Jobs, P.A. Semi, Jensen Huang, Larrabee, Samsung, ElevenLabs, Lovable, Anton Osika, Founder University, Slack, Salesforce, HubSpot, Google Suite, Microsoft Office, Nursa, CERN, Claude, Anthropic
基辛格警告封锁台湾将引发全球大萧条
中国对台湾为期三周的封锁将耗尽该岛的能源储备，并导致半导体晶圆厂停产90天，从而引发比大萧条更为严重的全球经济崩溃。
本期内容探讨了科技领导权的关键转移，追溯了Intel如何因管理层从技术型创始人转向财务导向型高管，导致错失EUV光刻与移动芯片的发展机遇，而Nvidia和TSMC则借势迅猛崛起。Pat Gelsinger对潜在的台湾封锁可能给全球经济造成的灾难性冲击提出警告，并展望了到2030年由受能源限制的AI与量子计算所构成的未来图景。最后，Anton Osika展示了Lovable如何通过让非技术用户以远低于传统工程成本的方式构建、托管和运行高度安全的定制企业级应用，从而实现年营收规模达到5亿美元。
英特尔前 CEO 帕特·基辛格（Pat Gelsinger）指出，英特尔衰退的根源在于技术领导力被追求短期财报效益的财务高管取代；在其回归前数年间，公司向股东派息并回购了 1000 亿美元股票，却长达十年未建新晶圆厂也未投资 EUV 光刻技术。 (0:00)
基辛格回顾了英特尔的多项战略失误，包括在离职后一周便取消了关键的并行吞吐计算项目 Larrabee，而英伟达则凭借 CUDA 生态实现持续创新；同时台积电凭借纯代工模式建立起相对英特尔约 7 比 1 的晶圆产能优势。 (7:56)
基辛格强调台湾若遭遇封锁断电仅需三周便会导致半导体晶圆厂瘫痪，造成的全球经济损失将超过大萧条；此外，美国电网等能源基础设施的瓶颈正成为制约 AI 算力无限制扩张的硬性天花板。 (14:11)
Lovable 创始人 Anton Osika 分享平台上线 20 个月已构建超 5000 万款应用并突破 5 亿美元年收入，其 80% 用户为非技术人员，通过 AI 动态模型路由彻底颠覆传统软件开发成本结构并催生出多团队并行实验的全新研发模式。 (25:00)
Reading summary
Pat Gelsinger attributes Intel's historic decline to financial executives prioritizing short-term capital distributions over manufacturing and technical roadmaps, while Lovable CEO Anton Osika demonstrates how AI-driven software development is radically collapsing engineering costs and shifting the primary bottleneck to human strategic planning.
Executive Leadership and Capital Misallocation at Intel (0:00)
Gelsinger argues that Intel's competitive decline stemmed from a leadership transition away from deeply technical founders and PhDs toward financial executives focused on spreadsheet economics. In the five to six years before Gelsinger's 2021 return as CEO, Intel distributed $100 billion to shareholders through buybacks and dividends, but failed to build a new fabrication facility for a decade or invest in extreme ultraviolet (EUV) lithography.
Intel also abandoned critical forward-looking projects, terminating the Larrabee many-core throughput computing architecture one week after Gelsinger's initial departure, whereas Nvidia sustained research and development around CUDA and general-purpose GPU computing. Furthermore, Intel's proprietary Integrated Device Manufacturer (IDM) framework fell behind TSMC's open foundry model; supported by anchor customer Apple, TSMC's manufacturing volume expanded to roughly seven times Intel's wafer output.
Taiwan Supply Chains and Geopolitical Fragility (14:11)
Gelsinger highlights the extreme concentration and fragility of leading-edge semiconductor manufacturing in Taiwan, pointing out that China has blockaded the Taiwan Strait seven times in four years during military exercises. Referencing a Wall Street Journal report, Gelsinger notes that Taiwan possesses less than three weeks of energy reserves, meaning an import blockade on oil and liquefied natural gas could force power brownouts without direct military strikes.
Because an unexpected fab shutdown requires at least 90 days to safely resume operations, Gelsinger projects that a blackout in Taiwan would trigger an economic collapse greater than the Great Depression. While Gelsinger reports that US CHIPS Act initiatives have increased domestic leading-edge chip production from 12% in 2021 to 18% by 2026, he maintains that geographic diversification and commercial foundry transitions must move faster to secure supply chains.
Energy Bottlenecks, Jevons Paradox, and Quantum Computing (17:27)
Addressing concerns over an artificial intelligence capital expenditure bubble, Gelsinger identifies power grid infrastructure as the binding physical ceiling on hardware deployment. While global power capacity expands at 4% to 5% annually, the US electrical grid grew at roughly 1% per year across a 15-year stagnation period, restricting the pace at which data centers can activate GPUs.
Rather than anticipating a market crash, Gelsinger forecasts a multi-decade buildout aimed at making AI 10,000 times more efficient. Lowering token costs by five orders of magnitude will, in Gelsinger's view, trigger Jevons Paradox, vastly expanding net resource consumption. Looking to long-term computational horizons, Gelsinger predicts practical quantum supremacy across multiple industries by 2030, with cryptographic challenges and Q-day resolutions following around 2032 or 2033.
Generative Software Economics and Strategic Planning Bottlenecks (25:00)
Lovable CEO Anton Osika outlines how modern large language models have transformed software production, scaling Lovable to over 50 million created applications and $500 million in annual revenue by May 2026. With 80% of users lacking technical backgrounds, internal enterprise applications that previously cost upwards of $500,000 and months of development are now deployed by non-engineers in hours for under $2,000.
Lovable leverages dynamic model routing between commercial models like Claude and proprietary open-weight systems, while encouraging rapid parallel experimentation ('co-opetition') across teams. Osika and host Jason Calacanis emphasize that because software synthesis is increasingly automated, code generation is no longer the limiting constraint; development speed is instead gated by human-to-agent planning, prompt architecture, and business strategy.
While semiconductor manufacturing and AI application development are advancing rapidly, both domains face hard non-technical limits: hardware independence is constrained by the speed of capital-intensive onshoring against geopolitical timelines, while automated software production shifts organizational friction from coding execution to energy availability and high-level strategic reasoning.
## Background
该视频由知名科技与投资播客《All-In Podcast》于 2026 年 7 月 27 日在 YouTube 平台发布。节目简介显示，主持人 Jason 先后对话了前英特尔 CEO 帕特·基辛格（Pat Gelsinger）探讨英特尔的战略失策与台海局势对供应链的冲击，以及 Lovable 创始人 Anton Osika 分享“氛围编码”（Vibe Coding）如何重塑软件开发生态与降低构建成本。
## TL;DR
前半部分中，前英特尔 CEO 帕特·基辛格深度复盘了英特尔因财务官僚掌舵、千亿回购和削减资本开支而错失移动芯片与先进制程的惨痛教训，并严厉警告台海能源脆弱性可能引发超越 1929 年大萧条的全球供应链灾难，同时看好 AI 算力与量子计算在 2030 年前的实质突破。后半部分中，Lovable 创始人 Anton Osika 展示了平台每周诞生百万款新应用的惊人生态，详解其如何通过多模型路由与主见化架构为 80% 的非技术用户兜底生产级软件开发。主持人 Jason 结合团队亲历案例，力证这种 AI 原生开发范式已将传统数十万美元的企业系统交付成本打入谷底，彻底打破了传统 SaaS 的既有壁垒。
## Atmosphere & style
整场对话在厚重深沉的产业反思与极度兴奋的前沿创新之间形成了强烈的戏剧性反差。基辛格的访谈兼具老派工程领袖的痛心疾首与地缘博弈的冷峻警醒，其对千亿资金流失和台湾三周能源底线的推演令人凛然；而随后与 Anton 的连线则迅速将气氛拉入硅谷极客的高频狂欢，二人言语间充满对人机协同颠覆软件生产力的绝对自信与幽默打趣，整体观感既是一堂充满血泪教训的商业史大课，又是一场充满技术乐观主义的敏捷开发实战秀。
## Bulletin
- [14:11](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=851s) 前英特尔 CEO 帕特·基辛格指出，台湾关键能源储备不足三周，一旦遭封锁停电，先进制程晶圆厂重启需长达 90 天，其对全球经济的破坏力将超越 1929 年大萧条。
- [17:27](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1047s) 基辛格认为全球年均仅 4% 至 5% 的能源增速是防止 AI 盲目泡沫的“物理安全阀”，并断言量子计算在 2030 年前必将取得实质突破并实现量子霸权应用。
- [36:56](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=2216s) AI 应用构建平台 Lovable 年化收入（ARR）已在 2026 年 5 月突破 5 亿美元，累计生成超过 5000 万个独立应用，月访问量突破 7 亿次。
- [44:55](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=2695s) “Vibe Coding”（氛围编码）正使企业内部定制软件边际成本骤降 99%，团队通过类似 CERN 的多方案“竞合（Co-opetition）”并行试错模式替代传统臃肿软件与通用 SaaS。
## Key takeaways
- [02:30](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=150s) 英特尔被财务官僚管理者掌舵后，将千亿美元资金用于股票回购与分红，导致资本开支窒息、十年未建新厂，并先后错失初代 iPhone 芯片代工与 EUV 光刻机红利。
- [05:16](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=316s) 史蒂夫·乔布斯早在 2008 年就敏锐预见英特尔无法满足移动能耗需求，提前秘密完成 Mac 系统的跨平台移植并收购半导体团队，奠定了 Apple Silicon 自主生态。
- [09:58](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=598s) 英特尔内部曾启动类似英伟达 GPGPU 路线的 Larrabee 项目，但因基辛格首次离职而仓促终止；英伟达则凭借对 CUDA 和底层软件栈的长线投入构建起 AI 计算霸权。
- [11:33](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=693s) 台积电通过中立开放的代工生态打破了英特尔高度封闭专有的 IDM 模式，出货规模拉大至英特尔的 5 到 7 倍，倒逼英特尔全面启动代工转型。
- [14:11](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=851s) 台湾全岛液化天然气与石油储备不足三周，任何海上封锁将直接瘫痪需 90 天清洗重启周期的晶圆厂，推动美国必须加速本土半导体自主可控进程。
- [17:27](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1047s) 电网基建缓慢的扩张周期（美国年增速仅约 1%）给 AI 数据中心算力设下了物理硬上限，长远看需通过推动 AI 效率提升 10,000 倍以降低单 Token 推理成本与能耗。
- [21:21](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1281s) 未来的计算范式将由经典计算、AI 计算与量子计算构成的“三位一体”主导，适度的市场估值回调是预防泡沫无序膨胀的必要防线。
- [22:41](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1361s) 量子计算的核心理论纠错与算法已成型，2030 年前将在分子模拟、化学合成及大规模路径优化中攻克经典计算无法企及的极限难题。
- [25:00](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1500s) Lovable 平台每周生成新项目超 100 万个，其核心客群由 80% 非技术业务人员与 20% 专业工程师组成，正从纯应用生成拓展为全天候运转的“AI 联合创始人”。
- [29:17](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1757s) “Vibe Coding”使软件开发跨越了概念原型阶段直接进入生产环境，由平台在底层自动接管支付集成、安全防护、邮件系统与 AI 搜索引擎可见性（SEO）。
- [31:12](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1872s) 企业构建内部系统的成本出现结构性暴跌，过去耗费 50 万美元外包预算的内网管理平台，非技术员工使用 Lovable 仅需 4 小时、以不足 2,000 美元成本即可高质交付。
- [36:56](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=2216s) 自研定制化软件正在逐步蚕食传统通用 SaaS 的市场份额，企业通过针对特定业务自建前端，不仅能大幅节省外部授权开支，还能避免数据孤岛。
- [40:28](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=2428s) Lovable 采用商业前沿大模型与自研开源微调模型的动态分流路由，依托斯德哥尔摩研发中心将真实报错与开发者经验通过强化学习沉淀为系统技能集。
- [43:41](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=2621s) AI 编程工具正告别低价倾销算力的恶性竞争，由于投入产出比极高，低门槛订阅套餐中有近六成用户因真实生产力需求主动追加充值购买算力。
- [44:55](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=2695s) 借鉴 CERN 粒子物理攻坚的“竞合”思维，团队允许多个成员针对同一需求各自用 AI 并行构建方案，通过 A/B 真实测试择优融合，根除臃肿妥协的“弗兰肯斯坦软件”。
- [47:08](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=2828s) 前沿基础模型具备惊艳的单次生成表现，但软件持续演进的瓶颈依然在于人类对战略方向、核心业务规则和真实痛点的深度协同把控。
## Detailed walkthrough
### [00:00](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=0s) 英特尔的辉煌与陨落：技术巨头如何偏离航向
英特尔（Intel）曾是美国商业史上最伟大的科技公司之一，凭借风靡全球的 "Intel Inside" 品牌战略和经典的奔腾（Pentium）处理器，让无数消费者仅凭这句标志性的宣传语和开机音效就选购电脑。然而，这家芯片霸主随后却遭遇了惨烈的滑铁卢，被英伟达（Nvidia）、台积电（TSMC）以及在一定程度上的苹果（Apple）彻底超越。前英特尔 CEO 帕特·基辛格（Pat Gelsinger）在英特尔度过了整整 34 年的漫长岁月，他见证了英特尔的巅峰，也亲历了公司陷入困境、错失战略机遇的惨痛历程。
基辛格回顾道，自己 18 岁便加入英特尔，开玩笑说自己几乎是“在英特尔度过青春期的”。早期培养他的导师是安迪·格鲁夫（Andy Grove）、罗伯特·诺伊斯（Robert Noyce）、戈登·摩尔（Gordon Moore）和克雷格·贝瑞特（Craig Barrett）等半导体领域的传奇人物。彼时的英特尔管理层充满浓厚的技术氛围——基辛格回忆自己首次列席高管会议时，全场 20 人里至少有 15 人拥有博士学位。然而，英特尔后来偏离正轨的根本转折点，正是当技术人才让位于“算账的”财务与商科背景管理者（bean counters）。
```mermaid
flowchart TD
accTitle: 英特尔领导层演变与企业失策路径
accDescr: 展示英特尔从早期技术专家领导转向财务经理人执掌后，如何导致战略性研发投入不足的链条。
subgraph Phase1["技术立企时代"]
A["传奇工程师领导<br/>(格鲁夫 / 摩尔 / 诺伊斯)"] --> B["高管超75%为博士<br/>深厚技术洞察"]
end
subgraph Phase2["财务官僚时代"]
B --> C["非技术背景高管掌舵<br/>依财务报表做决策"]
C --> D["优先回购与分红<br/>(1000亿美元流向股东)"]
end
subgraph Phase3["核心能力落后"]
D --> E["削减资本开支<br/>十年未建新工厂"]
D --> F["错失EUV光刻机<br/>放弃iPhone芯片"]
end
```
### [02:30](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=150s) 财务官僚的短视：千亿美元回购与资本开支的窒息
当非技术人员接管科技公司后，管理层往往会提拔更多同质的业务线管理者，导致工程基因被逐渐边缘化。基辛格强调，如今最顶级的科技企业要么是由创始人执掌，要么是由兼具深厚技术素养的领袖掌舵——即使像微软的萨提亚·纳德拉（Satya Nadella）或谷歌的桑达尔·皮查伊（Sundar Pichai）并非创始人，他们本身也是精通技术的专家。在半导体这类动辄需要数十亿美元豪赌的核心领域，仅靠 Excel 电子表格来评估回报是致命的，因为许多技术投资在立项之初从财务模型来看都是“糟糕的投资”，必须由真正懂得技术发展趋势的人来判断其战略价值。
在基辛格 2021 年重返英特尔担任 CEO 之前的大约五六年间，前任管理层将多达 1000 亿美元的资金投入到股票回购和分红中，以此取悦华尔街。基辛格感慨，如果能把这 1000 亿美元留在账上，英特尔能完成多少不可思议的战略突破：
- 在基辛格重新执掌英特尔之前，公司竟然已有整整十年未曾新建一座芯片工厂。
- 面对极紫外光刻技术（EUV）的革命性节点，当时的决策层因初始经济效益不佳而迟迟未能全力采购 ASML 的 EUV 机器。
- 缺乏技术远见的领导层曾断然拒绝为初代 iPhone 代工制造芯片，彻底错过了移动半导体时代的最大红利。
### [05:16](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=316s) 苹果自研芯片的伏线：乔布斯早有准备的降维打击
访谈中提到了苹果公司的资本运用，苹果同样进行了巨额回购和分红，但凭借极高的利润率坐拥庞大的现金流。而在供应链掌控上，史蒂夫·乔布斯（Steve Jobs）早在 2008 至 2009 年间便开始谋划自研芯片。基辛格披露了当年苹果从 IBM PowerPC 转向英特尔 x86 芯片及 Centrino 平台背后的惊人内幕。当时苹果是一个极其严苛的客户，对芯片的体积、功耗提出了极度苛刻的要求。
当乔布斯预感到英特尔无法长期满足苹果对芯片低功耗和高性能的极限迭代时，便开始通过收购 P.A. Semi 等小型半导体设计公司，在内部悄悄构建自研芯片的核心能力。基辛格回忆道，在当年苹果准备将 Mac 操作系统从 PowerPC 迁移到 x86 时，英特尔的高管们对自家的编译器和系统软件能力非常自信，热情地向乔布斯表示可以协助苹果将系统迁移到 x86。乔布斯却冷冷地回应：“我已经在过去的四个操作系统版本里全部完成了向 x86 的移植，我看我们自己能搞定。”这种深谋远虑的工程掌控力让基辛格深感震撼。苹果不想再受制于为 Windows 优化的英特尔芯片，而是追求系统级软硬件协同，最终走出了一条完全自主化的 Apple Silicon 之路。
### [07:56](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=476s) 英伟达的豪赌：从游戏显卡到 HPC 与 AI 的通用算力革命
与此同时，黄仁勋（Jensen Huang）带领下的英伟达走出了一条看似充满机缘巧合、实则是极高技术远见的道路。在英特尔 CPU 处于绝对统治地位的黄金时代，英特尔内部其实对英伟达的显卡嗤之以鼻，认为那不过是“一些狂热游戏玩家才会买的图形卡”，根本无法与强大的通用 CPU 相提并论。
然而，英伟达的策略与乔布斯暗度陈仓的打法颇有异曲同工之妙。黄仁勋带领团队在底层软件栈上下死功夫，重金研发 CUDA 和单指令多线程（SIMT）等并行计算技术，并在每一个产品版本中持续迭代。转折点发生在注重密集计算的日本高性能计算（HPC）科研圈，一批学者敏锐地发现英伟达显卡具有极高的算力密度，开始尝试将其应用于超算任务。这次突破让黄仁勋和英伟达意识到，显卡不再仅仅用于绘制游戏画面，而是能够解决全球最复杂算力瓶颈的高吞吐量通用计算平台。尽管当时人工智能已经历了五次所谓的“AI 寒冬”，整个行业普遍悲观地认为 AI 永远无法取得真正的技术飞跃，但英伟达围绕并行计算构建的开发者生态早已在暗中蓄力，静候质变时刻的到来。
### [09:58](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=598s) Larrabee 的夭折与英伟达 CUDA 的崛起
基辛格回顾了英特尔在并行计算领域错失的战略先机。当年英特尔内部曾启动过名为 Larrabee 的项目，试图基于 x86 架构实现通用 GPU（GPGPU）计算，走一条与英伟达类似的技术路线。然而，在基辛格第一次离开英特尔时，仅仅过了一周，这个极具前瞻性的项目就被直接砍掉。基辛格感慨道：“如果当时坚持下去，今天的世界格局会截然不同。”
主持人 Jason 借用科幻作家威廉·吉布森的名言“街道会为技术找到它自己的用途”来点评持续创新、承担风险与底层研发的复利效应。英伟达当年投入巨资研发 CUDA 时，并没有预料到后来的比特币挖矿潮会将其计算架构推向极致，也无法完全预知当下的 AI 大爆发。正是因为英伟达提供了最底层的优秀算力工具，黑客社区和极客开发者们自发在上面找到了破局之道，并不断迭代进化。
- **英特尔 Larrabee 项目**：试图用 x86 架构处理图形和高并行计算，但在基辛格首次离职后仅一周即被终止。
- **英伟达 CUDA 的壁垒**：凭借一代又一代软硬件结合的持续投入，最终在加密计算与 AI 浪潮中成为绝对基石。
---
### [11:33](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=693s) 台积电代工模式的逆袭与英特尔 IDM 帝国的失守
随着苹果放弃英特尔转向自研 Apple Silicon、英伟达统治算力，半导体版图上另一家台湾巨头——台积电（TSMC）更是改变了整个产业格局。基辛格深入剖析了台积电崛起的商业逻辑：台积电最初带着纯粹的晶圆代工（Foundry）愿景起步，立志成为全行业的制造工厂。建立先进制程晶圆厂极其昂贵，单座晶圆厂投资就高达 200 亿到 300 亿美元，需要庞大的工程团队和长期持续的资金投入。
与台积电形成鲜明对比的是当时的英特尔。英特尔一直固守 IDM（集成设计与制造）模式，从未考虑将自己的制造工艺和工厂开放给第三方使用。英特尔内部的逻辑是“我们自研 CPU 的出货量足够大，多余产能做做主板芯片组和周边元器件就好”，因此整套制造流程高度封闭，从未进行标准化，也不对外提供标准制程设计套件（PDK）或兼容通用的 EDA 设计工具。基辛格回忆，早年自己在英特尔牵头搭建了 EDA 工具的基础体系，包括最早的布局布线（Place and Route）、标准单元库及高层次硬件描述语言，但这套体系完全属于英特尔私有专有技术。
台积电则彻底打破了这种垂直垄断，宣称“我不在乎是谁设计的芯片，只要你有图纸，我就是你的制造伙伴”。最初英特尔对这种细碎的代工业务根本不屑一顾，认为市场份额微不足道。然而台积电凭借长期扎实的积累，加上大客户苹果对极限制程的严苛驱动，实现了质的飞跃。当基辛格 2021 年重返英特尔担任 CEO 时，惊觉台积电的晶圆出货量已经是英特尔的 5 倍之多，如今这一差距更是拉大到 7 倍左右。全球半导体行业除了内存（Memory）领域和英特尔之外，几乎全盘倒向了“开放代工 + 通用 EDA 工具”的模式。这也促使基辛格回归后坚定推行重塑英特尔的战略核心：英特尔必须转型，建立属于自己的代工业务。
```mermaid
flowchart TD
accTitle: 半导体制造模式演进对比
accDescr: 对比传统封闭的 IDM 模式与现代开放的代工生态模式
subgraph IDM ["英特尔传统 IDM 模式（封闭垂直整合）"]
A1["自研芯片设计 (x86 CPU)"] --> B1["专有私有 EDA 工具 / 内部 PDK"]
B1 --> C1["自有内部晶圆厂"]
C1 --> D1["仅供自家产品与芯片组使用"]
end
subgraph Foundry ["台积电 Foundry 模式（开放分工生态）"]
A2["全球无晶圆厂客户 (苹果 / 英伟达 / AMD 等)"] --> B2["行业标准化 EDA 工具 & 开放式 PDK"]
B2 --> C2["超大规模共享代工厂"]
C2 --> D2["海量多元芯片交付全球生态"]
end
```
---
### [14:11](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=851s) 芯片法案、台湾能源死穴与地缘危机推演
Jason 将话题引向地缘政治现实：外界高度担忧台海局势可能对美国经济造成毁灭性打击，甚至有传闻探讨若局势恶化是否会炸毁晶圆厂、撤离工程师。美国能否通过《芯片法案》（CHIPS Act）在本土复刻这些尖端制造能力？
基辛格直言这一议题极其严峻。他指出《芯片法案》确实正在见效，2021 年他重返英特尔时，美国先进制程制造在全球占比仅有约 12%，如今已回升至 18% 左右；英特尔正在稳步推进代工业务，台积电在美工厂也已规模化运转，三星同样在建厂。但与 50% 的安全目标相比，美国依然任重道远。
随后，基辛格谈到了一个令全行业“不寒而栗”的现实隐患：根据《华尔街日报》此前的报道，台湾全岛的能源储备不足三周。基辛格警告，封锁甚至无需真正擦枪走火，只要切断三周的液化天然气（LNG）与石油补给，全岛就会面临大范围断电。半导体先进制程晶圆厂一旦彻底断电停机，从完全断电到重新清洗恢复生产至少需要 90 天。台湾电网崩溃对全球经济造成的破坏力，将“直接超越 1929 年的大萧条”。基辛格强调，过去四年中国大陆已在台湾海峡进行了多次封锁演练，“这不是假设理论，而是正在发生的现实威胁”。无论是 2027 年、2030 年还是 2035 年，全球都必须不惜一切代价加速建立更具韧性的本土供应链。
- **先进制程本土化份额**：美国本土先进制程产能占比已从基辛格回归时的约 12% 提升至目前的 18% 左右。
- **台湾能源安全储备极限**：全岛关键能源储备不足 3 周，极易受海上封锁影响。
- **晶圆厂停机代价**：先进晶圆厂一旦非正常断电停产，重新恢复正常运转需要长达 90 天。
---
### [17:27](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1047s) AI 基础设施是否存在泡沫？能源瓶颈与代币经济的万亿愿景
面对当前轰轰烈烈的 AI 数据中心扩建、万亿级资本开支与算力争夺战，Jason 提出质疑：这是否会重蹈过往科技泡沫的覆辙？如果硬件建设过度超前，而应用端无法产生足够的收益，上市公司终将面临投资者的无情抛弃。
基辛格对此给出了一个反直觉的冷静判断：**能源供给的天然刚性，恰恰是防止 AI 陷入盲目泡沫的“安全阀”**。全球能源产能每年的增速只有 4% 到 5%，而美国在过去十余年间的电网增速几乎停滞在 1% 左右。哪怕资本市场再狂热、囤积再多 GPU，没有充足的电力支撑，谁也无法凭空点亮一座数据中心。正是电网基础设施的建设周期，给 AI 算力的过度扩张设下了物理硬上限，让行业不至于过度透支。
另一方面，基辛格对 AI 长期价值持有坚定的乐观态度。他认为每个代币（Token）的增量价值本质上是“智能的度量衡”，在自动化物流、供应链优化、金融风控等领域的应用价值几乎是无限的，尤其是在发达国家面临劳动力结构性短缺的大背景下。基辛格坚信，眼下的 AI 基础设施建设并非两三年的短周期热潮，而是一场长达数十年的超级周期。他给自己设定的长期核心目标之一，就是“将 AI 的效率提升 10,000 倍”——通过将单 Token 的推理成本与能耗降低多达五个数量级，使智能真正普惠全社会。
### [20:10](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1210s) 杰文斯悖论、AI 推理芯片与技术黄金时代
Jason Calacanis 与 Pat Gelsinger 深入探讨了当前人工智能领域的经济学规律。Jason 指出，过去一年里“杰文斯悖论”（Jevons Paradox）体现得淋漓尽致：由于模型 Token 成本大幅暴跌、开发工具日趋强大，用户和开发者开始全天候高频调用，直到收到巨额账单时才意识到必须追求明确的投资回报率（ROI）。与此同时，硬件层面正涌现出一批极具创新力的公司，例如 Cerebras、Groq 以及 d-Matrix，它们专注于构建专用推理芯片，正在根本性重构推理侧的算力经济学。
Pat Gelsinger 对此充满乐观，认为如果行业能够实现 Token 经济效益数个数量级的跃升、保证高可用性并显著降低能源消耗，人类社会将迎来极其辉煌的未来二十年。他由衷感慨，人类历史上从未有过比当下更适合成为技术人员的时代：“我们将解决化学难题、攻克自然语言、发明全新材料、重塑交互方式、战胜癌症，并将无数人从贫困中解救出来。没有比现在活着更美好的时代了，而作为技术人员，我们正坐在时代的驾驶座上。”
- **推理芯片破局者**：Groq、Cerebras 与 d-Matrix 等芯片设计商正针对大模型推理阶段优化架构，推动算力成本与能耗持续指数级下降。
- **技术红利领域**：化学与生物材料设计、疾病根治（如癌症靶向治疗）、多模态交互革命。
---
### [21:21](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1281s) AI 估值周期与计算的“三位一体”
谈及当前 AI 创业公司的高估值时，Jason 联想到当年的互联网泡沫（Dot-com Bubble），但他也敏锐地指出现在的格局截然不同：如今的头部公司拥有真实的营收支撑，例如 ElevenLabs 营收已达数亿美元规模，Lovable 的估值与业务体量也迅速突破五六亿美元，与当年的纯概念投机不可同日而语。Pat 赞同这一判断，指出当今的商业模式展现出了真实的收入与利润率，但他同时强调，任何时候只要估值倍数过高，周期性回调都是健康且必要的。
在 Pat 看来，未来二十年的繁荣绝非一条平滑上升的曲线，中间必然会遭遇各种震荡与行业洗牌。正如软件领域曾经历过“SaaS 大浩劫”（SaaS Apocalypse），未来各行业在面对被 AI 释放的颠覆性能力时，同样会迎来各自的重构周期。每当市场出现回调，“我们应该心存感激”，因为这防止了泡沫盲目膨胀失控。更关键的是，目前的浪潮还只是序幕，真正的科技聚变将发生在 Pat 所称的“计算三位一体”（Trinity of Computing）交汇之时：
```mermaid
flowchart TD
accTitle: 计算的三位一体架构
accDescr: 展示经典计算、AI计算与量子计算融合协同的未来计算架构
subgraph ComputingTrinity["计算的三位一体 (Trinity of Computing)"]
A["经典计算<br/>(Classical Computing)"]
B["AI 智能计算<br/>(AI Computing)"]
C["量子计算<br/>(Quantum Computing)"]
end
A <--> D["终极计算能力跃迁"]
B <--> D
C <--> D
D --> E["彻底重塑材料、密码学与分子科学"]
```
---
### [22:41](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1361s) 量子霸权落地倒计时：2030 年前见证实质突破
面对 Jason 半开玩笑式的质疑——“过去25年里，量子计算似乎总被说成‘还有5年就会到来’，它究竟何时才能产生实际价值？”——Pat Gelsinger 斩钉截铁地回答：“就在这个十年，2030年之前它必将发挥实质性作用。”他解释道，届时量子计算将攻克当今经典架构完全无法计算的极限难题，特别是在生物学、复杂化学分子模拟，以及如旅行商问题（Traveling Salesman Problem）等超大规模物流路线优化场景中，直接输出全局最优解。
至于密码学破解以及关乎国家安全与数据保密的“量子日”（Q-Day），Pat 预计可能会落在 2032 到 2033 年左右。但就本年代而言，各行业将实打实地迎来“量子霸权”（Quantum Supremacy）的应用成果。底层的工程突破已经奠定：行业已经掌握了物理量子比特的构建方式，攻克了量子纠错（Error Correction）的核心理论，并开发出了匹配的量子算法，眼下剩下的纯粹是工程化规模扩张问题。
- **多路线齐头并进**：当前离子阱（Trapped Ions）、光量子（Photonic）、自旋量子（Spin Qubits）等四到六种不同的物理模态均展现出良好进展，技术路线已不再是阻碍。
- **代表性布局**：作为其投资组合的一部分，Pat 高度看好光量子路线的领军企业 PsiQuantum。
- **倒计时紧迫感**：Jason 提醒观众，距离 2030 年其实只剩下大约 40 个月的时间，产业变革近在眼前。
---
### [24:34](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1474s) 赞助插播：Plaud 智能录音硬件
在节目的商业赞助环节，Jason 推介了专为重度沟通者打造的可穿戴 AI 录音硬件 Plaud。商业世界中最关键的决策与火花极少发生在写字楼的工位上，更多是在走廊水池旁、晚餐席间或与创始人的突发通话中产生。Plaud NotePin 能够直接吸附或别在衣物上，实现全程免提录音，并由后台的 Plaud 智能引擎直接提炼成条理分明的会议笔记和具体待办事项，让商务人士在完全专注于现场交谈的同时，轻松保留关键信息。
---
### [25:00](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1500s) 对话 Lovable 创始人 Anton Osika：从产品构建到商业闭环
随后，Jason 连线了他极为推崇的创业者——Lovable 创始人 Anton Osika。Jason 称赞 Anton 打造了一款让用户彻底上瘾的现象级产品（连 Jason 自己的团队都在狂热使用），并盛赞他对“赋能全人类构建优秀软件”这一北极星目标的极致专注。Anton 阐述了 Lovable 创立的终极愿景——“赋能人类”（Empowering humans），并指出横亘在普通人与商业成功之间的两道巨大鸿沟：第一道鸿沟是“如何从零做出一个产品”，第二道则是“如何围绕这个产品建立起真正的商业闭环”。
Lovable 正在双管齐下解决这两大难题。在第一道鸿沟上，Lovable 已经取得了惊人的突破，目前平台每周新增生成的软件项目突破 100 万个；而在第二道鸿沟上，公司正大规模投入基础设施建设，帮助用户让构建的软件被更多人发现、完成商业化转化并稳定运行。整个生态的扩张速度令人咋舌，已演变成一个庞大的软件分发网络：
- **生态规模破纪录**：平台上线仅 20 个月左右，累计生成的独立应用程序已超过 5000 万个。
- **高频真实访问**：依托 Lovable 构建部署的应用，每月外部访问总量已超过 7 亿次。
- **企业端爆发式增长**：除了初创独立黑客，企业级客户已成为 Lovable 增长最快的业务板块，企业团队借助其高度规范的技术范式，大幅提升了跨部门协作与系统互联的交付速度。
---
### [26:58](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1618s) 从低代码玩具到生产级系统：Lovable 的用户画像与架构哲学
Jason 回顾了 10 年前孵化器中流行的“所见即所得”（WYSIWYG）及低代码/无代码（No-Code / Low-Code）浪潮。当时非技术背景的 MBA 创业者做出的无代码产品往往粗糙丑陋、运行迟缓且极其脆弱。直到如今大语言模型与前沿 AI 技术的介入，软件构建的上限才被彻底拔高。针对 Jason 关于“客户到底是程序员还是普通人”的提问，Anton 剖析了 Lovable 独特的双重用户结构：
```mermaid
flowchart LR
accTitle: Lovable 平台双客群构成与赋能模型
accDescr: 展示 Lovable 平台中专业开发者与非技术人员的不同使用路径及商业化闭环
subgraph NonTech["80% 非技术人群 (业务人员/独立创业者)"]
NT1["自然语言提示描述需求"] --> NT2["极速验证产品想法 (PMF)"]
NT2 --> NT3["发展副业，迈向百万美元营收"]
end
subgraph Tech["20% 专业工程师 (技术架构师/研发团队)"]
T1["认可平台规范化架构最佳实践"] --> T2["一键式集成高安全性支付"]
T2 --> T3["后台自动化安全扫描与监控"]
end
NT2 -.协作桥梁.-> T1
```
Anton 透露，平台用户中约有 20% 是专业软件工程师或技术极客。这群专业人士之所以青睐 Lovable，是因为它拥有非常清晰严谨的“主见化”（Opinionated）架构，将行业公认的软件工程最佳实践固化在框架中——从一键安全集成 Stripe 等支付网关，到每次代码变更后自动触发后台安全扫描与持续健康监控。而剩余 80% 的非技术用户，则利用 Lovable 极低的心智门槛快速探索产品雏形、验证市场契合度（PMF）。更有趣的是两个群体间的交叉碰撞：许多大型企业员工在公司内部接触 Lovable 后，在业余时间开启副业项目，其中不少副业甚至做到了数百万美元营收，促使他们最终全职创业。
---
### [29:17](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1757s) “Vibe Coding”重塑开发：跨越原型直接进入生产
Jason 敏锐地指出，大众对“Vibe Coding”（氛围编码）的认知正在发生本质蜕变。如果放在一年前，外界普遍认为这类工具仅仅是做产品线框图（Wireframe）、高保真原型或视觉概念的工具；然而到了今天，传统的画原型环节已经被彻底淘汰——任何人都能在短短一两天内直接敲定一个全功能、可上线的成熟产品。
更重要的是，从一个令人自豪的初步产品，到真正经得起市场考验的商业实体，中间有着无数繁重的工程暗礁。绝大多数人并未意识到 Lovable 在底层所承担的脏活累活：如何确保多端用户支付流程的稳定与合规？如何抵御注入与恶意爬虫攻击？如何确保高并发下的数据库持久化与数据绝不丢失？Lovable 正将这些原本需要庞大专业后端团队数月才能搞定的系统工程，封装为普通人触手可及的基础能力。
### [30:04](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1804s) 面向 99% 的非技术人群：用平台架构兜底软件开发
在过去 12 个月里，整个软件开发范式发生了剧烈的蜕变。如今许多工程师已经不再手动写代码，甚至都不用去查看底层代码，这标志着构建软件的门槛已被彻底打破——任何人即使没有工程背景，也能独立创造软件。Lovable 的核心价值在于，它不仅仅是一个代码生成器，而是为非技术用户构建了一套严谨的软件架构。它为用户在底层铺设好安全护栏，确保开发过程“不会跌落悬崖”；诸如支付网关设置、邮件系统集成，以及针对 Google 搜索和各家 AI 聊天引擎的 SEO 与可发现性优化，全都在后台被自动妥善处理。
对于普通用户来说，他们完全不必去探究这些繁琐的技术底层细节。平台自动承担了数据安全合规以及与其他第三方生产力工具的安全互通职责。Lovable 从第一天起就坚持“极具主见（opinionated）”的架构路线，坚定地把重心锁定在“占人口 99% 的非技术受众”这一广阔市场，正是这一战略选择奠定了其爆发式的商业成功。
```mermaid
flowchart TD
accTitle: Lovable 架构保护与赋能非技术人员流程图
accDescr: 展示非技术用户通过 Lovable 平台构建应用，平台在后台自动处理架构、安全、支付与发现性等能力。
subgraph User["创作者层 (99% 用户)"]
U["非技术用户 / 运营人员"] -->|"自然语言提示词 / 意图"| L["Lovable 智能引擎"]
end
subgraph Core["Lovable 平台兜底架构"]
L --> S["规范化架构 & 错误防护"]
S --> SEC["后台安全扫描 & 数据加密"]
S --> INT["第三方生态安全互联 (API)"]
end
subgraph Ops["端到端服务与集成"]
SEC --> PAY["支付系统集成"]
SEC --> EML["邮件通信服务"]
INT --> SEO["SEO 与 AI 引擎可发现性"]
end
```
---
### [31:12](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1872s) Jason 团队的实战案例：4 小时替掉 50 万美元的内网开发
Jason 分享了自己团队在实际业务中探索各类 AI 工具的真实故事。最初，他给团队开放了所有能想到的工具：有人尝试过 Lovable 但没能做完完整的内网（Intranet）；有人去尝试 Cursor 或 Claude Code，虽然开启了所谓的“Vibe Coding”（氛围编码），却始终无法把产品完整落地交付；也有人去折腾 CodeWorks 或 Perplexity Computer。直到主持 Founder University（创业者大学，Jason 旗下的早期孵化器项目）的女性负责人为了给项目搭建一套内网管理系统，在完全没有预先申请预算和许可的情况下，直接刷公司信用卡购买了 Lovable 订阅，只花了 4 到 8 个小时就独立做出了整套成熟好用的内网系统，甚至还自主扩充了多项 Jason 从未要求过的超预期功能。
这套软件立刻成为了支撑该孵化器日常运转的核心引擎。随后，由于该孵化器在沙特阿拉伯和日本等地落地，需要向当地政府与合作伙伴量化其经济拉动效应，Jason 顺势提出需求：“你能否为参与项目的 50 家创业公司做一个经济影响力测算模型？”负责人再次通过人类业务语言层层提示（prompting），Lovable 迅速建成了这套数据模块，把每家公司的雇员人数、纳税额度、租房开销以及平均薪酬水平全盘考虑了进去。Jason 感慨，如果倒退 10 年或者仅仅 2 年前，做一个同等水准的内部系统至少需要耗费 50 万美元的工程预算，公司绝不可能批准；而如今，仅靠一位时薪五六十美元的普通员工在 4 小时内完成，配合 Lovable 极具性价比的订阅费，全套交付成本直接骤降到了 2,000 美元以内。
- **价格与价值对比**：Lovable 的入门订阅价格每月仅需 25 美元，商业版（Business Plan）约为每月 50 美元，Jason 调侃这个定价“低得不可思议”。
- **渗透测试与企业安全**：面对 Jason 早期对数据泄露与网络安全的顾虑，Anton 表示 Lovable 在不可见的底层投入了巨额资金，所有免费和付费用户都享受着后台持续运行的安全扫描，甚至能够从容应对安全专家的渗透测试对比。
---
### [34:53](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=2093s) 从代码工具到全天候“AI 联创”，托管业务迎来爆发
回顾技术迭代的跨度，仅仅一年前平台还只能输出死板的前端原型图（mock-ups），而现在不仅能交付全功能的可用系统，还能直接实现企业级安全部署。当谈及未来一年的规划时，Anton 透露，Lovable 已经开辟了一条增速甚至超越了单纯“代码构建器”的新产品线——原生应用托管（Hosting）与 AI 算力服务。这项服务使用户在平台内生成的应用能够一键运行并面向全球提供服务，底层则协同 AWS 等传统云巨头展开深度基础设施合作。
更为深远的是，Lovable 正在演变为用户的“AI 联合创始人（AI Co-founder）”。由于企业的所有应用、工具与实时业务数据均在 Lovable 的平台上流转，大模型得以全盘穿透公司的运营指标。当前 Lovable 正在小范围内测一项功能：这位全天候运作的“AI 联创”会在创始人入睡期间自主运行，并在次日清晨递交一份详细的战略建议报告，内容涵盖业务扩张路径、流程降本增效建议以及客户体验优化策略。
- **留存逻辑的蜕变**：“用户因快速构建软件而来，但因经营运转企业而留下（You come to build the software, but you stay to build the business）”。
- **复利智能（Compounding Intelligence）**：平台长期建立了一套自我修正的 Agent 机制，Lovable 每一次犯错都会被反馈给内部工程师组成的系统进行迭代修复，这种能力提升同样会直接反哺给所有正在其上运营生意的企业客户。
---
### [36:56](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=2216s) 自研软件 vs 传统 SaaS：Nursa 案例与 5 亿美元营收里程碑
Jason 提出了一个关键的产业疑问：未来定制化自制软件（Bespoke Software）是否会 100% 取代企业内部使用的传统基础 SaaS 工具？比如公司内部每年花上万美元购买 Slack，未来是否只需 Vibe Coding 一套深度契合自身业务的即时通信工具？包括 Salesforce、HubSpot 以及微软、Google 办公套件等传统企业软件霸权是否会瓦解？
Anton 引用了一个典型客户案例来回答：美国大型机构 Nursa 的负责人 Ned 最初利用 Lovable 仅仅是为了孵化新业务“Nurse Study”（面向护士群体的继续教育与认证考试培训项目），并顺利构建了全套行政调度后台与执照资质管理平台；但随后，Ned 将其进一步引入内部后台流程，一口气用 Lovable 生成的定制化应用替代了企业原先采购的 10 多款外部 SaaS 工具，每年为公司硬生生省下了 100 多万美元的软件开支。不过，Anton 认为未来的格局不会是非黑即白的彻底取代，而是更具兼容性：
| 维度 | 传统通用 SaaS (Salesforce, HubSpot 等) | Lovable 定制化内部应用 |
|:--- |:--- |:--- |
| **应用场景** | 提供行业成熟的标准通用数据模型与底层生态 | 针对企业极其独特的特定业务需求与特殊流程 |
| **集成关系** | 继续充当底层数据仓库与核心服务载体 | 在传统 SaaS 之上构建一层完全贴合团队习惯的个性化交互前端 |
| **协同连接** | 依赖复杂的原生 API 或第三方 Zapier 类连接工具 | Lovable 主动提供结构化向导，清晰掌控权限与数据流向，杜绝泄密 |
在访谈尾声，针对市场上“每隔 6 个月就有人预言 Lovable 会被前沿大模型杀死”的质疑，Jason 调侃 Lovable “每被宣判死亡一次，营收就暴涨一亿美元”。Anton 当场披露了最新财务数据：Lovable 在 2026 年 5 月正式突破了 5 亿美元的年化收入（ARR），展现出了惊人的逆势增长。面对前沿基础大模型（Frontier Models）与开源模型的竞争，Lovable 坚持保持中立与用户利益至上，将底层最强模型直接吸收并封装进自身的端到端交付生态之中。
### [40:28](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=2428s) 多模型动态路由与开源权重后训练
Anton Osika 深入探讨了 Lovable 背后的底层模型架构与技术路径。他指出，Lovable 并不依赖单一的人工智能模型，而是采用多模型动态路由系统（Multi-model Routing）。用户在 Lovable 中输入构建需求时，系统会自动评估任务类型，并分发给最合适该场景的模型：既包括主流头部大模型厂商的商业前沿模型（Commercial Frontier Models），也越来越多地接入团队自主微调的开源权重模型（Open-weight Models）。随着任务不断路由到内部开源模型，这些模型在 Lovable 专属的 Agent Harness（智能体运行框架）环境中进化得更加智能，尤其是在修复模型常见错误、精准决定调用哪类工具（Tool Calling）、构建系统集成以及业务引导方面表现优异。
面对 Jason Calacanis 追问 Lovable 是否准备完全倒向开源（All-in Open Source），Anton 澄清说，Lovable 采取的是双轨并进策略：一方面与顶级前沿模型实验室保持紧密的商业合作并持续大额投入；另一方面，为了给客户带来“极致速度与最低成本”，Lovable 在斯德哥尔摩建立了一支极具实力的欧洲本土科研团队，全力专注模型后训练（Post-training）技术。Anton 强调，在欧洲本土沉淀并扩展这种核心 AI 能力，对于欧洲本土科技生态而言具有非常深远的战略意义。
```mermaid
flowchart TD
accTitle: Lovable 智能体路由与后训练架构
accDescr: 展示用户请求如何在商业前沿模型与自研开源模型之间分流，并通过真实错误反馈驱动后训练进化。
subgraph UserLayer ["用户交互层"]
U["用户构建需求与提示词"]
end
subgraph RoutingEngine ["动态路由调度层"]
R{"Lovable Agent Harness<br>智能路由分发"}
end
subgraph Models ["模型执行矩阵"]
CFM["商业前沿模型<br>(Anthropic / OpenAI 等)"]
OWM["自研开源权重模型<br>(斯德哥尔摩团队研发)"]
end
subgraph FeedbackLoop ["数据飞轮与后训练"]
T["百万级产品生成数据流<br>(Token 流分布)"]
Err["错误归因与优先级排序"]
RL["强化学习 (RL) / 技能微调"]
end
U --> R
R -->|复杂创新/逻辑任务| CFM
R -->|高频适配/低成本任务| OWM
CFM --> T
OWM --> T
T --> Err
Err --> RL
RL -.更新 Agent Harness 技能库与模型权重.-> OWM
```
- **开源与闭源协同**：既利用商业顶级前沿大模型的能力天花板，又通过内部开源模型做垂直场景优化以压低边际成本。
- **斯德哥尔摩研发中心**：Lovable 在瑞典斯德哥尔摩组建了强劲的后训练与强化学习研究团队。
---
### [41:53](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=2513s) 数据飞轮、强化学习与拒绝“倾销 Token”
在谈到是否借助外部数据标注公司训练模型时，Anton 说明了 Lovable 的专有数据闭环：团队核心精力放在捕捉前沿模型在实际编程中暴露出的错误上，并按照对用户业务的影响力进行优先级排序。利用 Lovable 每周诞生上百万个新产品的真实使用场景，团队沉淀出规模庞大的 Token 分布与交互信号。针对模型频繁踩坑的痛点，斯德哥尔摩团队通过强化学习（Reinforcement Learning）和构建专属数据集开展专项训练；同时，团队将优秀工程师构建软件的经验提炼为内置“技能（Skills）”，每周对 Agent Harness 及技能库进行持续迭代。
Jason 提及当前 AI 软件市场存在的恶性竞争与“Token 倾销（Token Dumping）”现象——某些平台以 50 美元亏本倒卖价值 100 美元的算力，沦为注定亏损的 Token 倒爷。Anton 回应道，Lovable 一贯非常严格地监控利润率与业务基本面。更重要的是，Lovable 绝不为了省钱去牺牲客户体验，从不在能用更聪明模型解决问题时故意切到劣质廉价模型。
- **每周上百万款新产品落地**：巨大的 Token 吞吐量和交互数据直接转化为产品与模型的微调燃料。
- **内置工程师技能集**：系统沉淀了资深开发者的经验法则，教会 Agent 在恰当的时机调取架构规范与工程记忆。
- **不打低质价格战**：坚持以智能水平和业务成效为先，拒绝用体验缩水换取虚假的低价。
---
### [43:41](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=2621s) 额度超额与 ROI：百倍收益下的价值重构
针对定价与使用上限，Anton 透露 Lovable 并非无上限免费供应，而是设置了多档订阅套餐（Subscription Tiers）和用量上限（Caps）。当用户消耗完套餐内额度后，可以通过额外付费（Top-up）购买算力包。在最低档套餐中，约有 60% 的活跃客户因为深度使用而主动触顶并追加充值。
Jason 认为这一数据反映了真实软件生产力的质变：用户愿意为超额用量付费，根本原因在于巨大的投资回报率（ROI）。以往企业开发一款专有软件可能需要动辄 50 万美元的预算与数月研发周期；如今即便用户拼命使用、一年在 Lovable 上跑满 6,000 美元的 Token，其开销也仅仅是过去传统外包或雇佣团队成本的 0.1% 到 1%。在这个追求快速迭代的时代，“多用 AI 能让人跑得更快”，用户完全不在意几千美元的算力账单，因为省下的时间与人力成本呈几何级数放大。
- **60% 触顶追加充值**：入门级订阅层中有近六成用户因高频使用而触发超额付费。
- **0.1%~1% 的极端成本优势**：6,000 美元/年的 AI 算力投入即可替代过去 50 万美元规模的定制软件开发。
---
### [44:55](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=2695s) 竞合模式打破“弗兰肯斯坦软件”，快速试错成常态
Jason 抛出了一个他在自身团队中观察到的新现象：同一个内部工具，团队成员由于上手门槛极低，日本团队和美国团队各自在 Lovable 上独立构建了一套完全不同的系统，形成了“内部赛马”甚至竞争格局。在以往，这种重复造轮子的行为会被严格禁止，企业会强行将所有部门需求捏合进同一个项目，最终拼凑出臃肿笨拙的“弗兰肯斯坦式怪物软件（Franken-software）”。
Anton 对这种“赛马式竞合”表达了强烈支持，并分享了自己早年在欧洲核子研究中心（CERN）从事粒子物理研究的一段经历：
> “在 CERN，他们引入了‘竞合（Co-opetition）’的概念——两支彼此独立的科研团队围绕同一个粒子加速器的不同部位展开攻关，在正式发表成果前完全互不沟通。这样既能对比两种技术路径的优劣，又不会过早陷入局部最优解（Local Minimum）的泥潭。”
Anton 分析道，随着 AI 抹平了软件工程的编码瓶颈，核心问题已经从“如何把它写出来”转变为“到底什么是正确的产品形态”。如果组织内有多位优秀人才，让大家各自挥棒尝试，随后再将各方案中最出色的部分整合进来，甚至直接上线开展 A/B 分流测试（Split Test），才是最大化商业成效的高效方法。
```mermaid
flowchart LR
accTitle: 传统集中式开发与 Lovable 竞合开发对比
accDescr: 对比传统单一堆砌需求的开发模式与现代多方案并发竞争试错的敏捷模式。
subgraph Traditional ["传统开发模式 (局部最优)"]
TReq["多团队繁杂需求"] --> TMonolith["单一系统强行捏合<br>(Franken-software 泥潭)"]
TMonolith --> TSlow["周期漫长且妥协折中"]
end
subgraph LovableCoopetition ["Lovable 竞合敏捷模式 (全局最优)"]
LTeamA["团队 A (Lovable 方案 1)"]
LTeamB["团队 B (Lovable 方案 2)"]
LTeamA --> Comp{"上线 A/B 分流测试<br>与指标验证"}
LTeamB --> Comp
Comp --> Best["融合双方优秀特性<br>产生最佳生产力工具"]
end
```
- **告别弗兰肯斯坦软件**：无需强迫多方在一套代码库里相互妥协，避免产生庞杂难用的劣质架构。
- **CERN 物理实验室哲学**：利用内部竞争打破思维定势，规避路径依赖与局部最优陷阱。
- **跨项目模块吸纳**：直接要求 Lovable 查看另一项目的闪光点并迁移整合，通过真实业务指标验证胜者。
---
### [47:08](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=2828s) 前沿模型边界与商业终局
对话转向业界对前沿多模态及推理模型的追捧，例如使用 Anthropic 尖端模型（如 Claude 新版/代号模型）直接复刻 3D 游戏（如《堡垒之夜》Fortnite）。Anton 确认 Lovable 同样接入并运用了这类顶级模型，并客观评价了其优势与局限：新一代模型确实具备惊人的“单次生成能力（Step Function）”，第一次提示词交互就能直接输出非常精致、高复杂度的 3D 界面或原型；但在软件长线演进与持续迭代中，真正的瓶颈依然在于人类开发者。人类必须同 Agent 并肩规划战略方向、打磨业务逻辑、接入正确的数据，并决定究竟应该解决什么商业问题，而这些核心判断力是 AI 无法在一夜之间取代的。
在节目的尾声，Jason 对 Lovable 给企业带来的颠覆性降本增效成果赞不绝口。他呼吁所有企业与创业者别再犹豫，“直接刷公司信用卡，立刻动手用 Lovable 构建”。面对外界每隔半年就断言“基础大模型升级会彻底杀死 Lovable”的质疑，Lovable 却能在每次模型迭代中不断浴火重生、实现营收数亿美元的跃升，靠的正是深耕客户真实需求和工程落地。最后，身处巴黎参加 RAISE 峰会的两人幽默调侃凡尔赛宫的宏伟，戏称未来人类终将借助 Lovable 与 Optimus 人形机器人共同构建如此壮丽的奇迹工程。
- **模型初试惊艳 vs. 长期迭代瓶颈**：新模型极度擅长首轮视觉与复杂任务生成，但持续产品演进依然高度依赖人机协同规划。
- **每六个月“重生”一次**：打破大模型厂商封杀中间层软件的悲观预言，凭借贴近客户真实业务流持续高速增长。
## Notable quotes
> [00:00](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=0s) 在英特尔度过青春期的 — Pat Gelsinger
> [05:16](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=316s) 我已经在过去的四个操作系统版本里全部完成了向 x86 的移植，我看我们自己能搞定。 — Steve Jobs
> [09:58](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=598s) 如果当时坚持下去，今天的世界格局会截然不同。 — Pat Gelsinger
> [09:58](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=598s) 街道会为技术找到它自己的用途 — William Gibson
> [11:33](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=693s) 我不在乎是谁设计的芯片，只要你有图纸，我就是你的制造伙伴 — TSMC
> [14:11](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=851s) 直接超越 1929 年的大萧条 — Pat Gelsinger
> [14:11](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=851s) 这不是假设理论，而是正在发生的现实威胁 — Pat Gelsinger
> [17:27](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1047s) 将 AI 的效率提升 10,000 倍 — Pat Gelsinger
> [20:10](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1210s) 我们将解决化学难题、攻克自然语言、发明全新材料、重塑交互方式、战胜癌症，并将无数人从贫困中解救出来。没有比现在活着更美好的时代了，而作为技术人员，我们正坐在时代的驾驶座上。 — Pat Gelsinger
> [22:41](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=1361s) 就在这个十年，2030年之前它必将发挥实质性作用。 — Pat Gelsinger
> [34:53](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=2093s) 用户因快速构建软件而来，但因经营运转企业而留下 — Anton Osika
> [44:55](https://www.youtube.com/watch?v=-ILKiOU5iAQ&t=2695s) 在 CERN，他们引入了‘竞合（Co-opetition）’的概念——两支彼此独立的科研团队围绕同一个粒子加速器的不同部位展开攻关，在正式发表成果前完全互不沟通。这样既能对比两种技术路径的优劣，又不会过早陷入局部最优解（Local Minimum）的泥潭。 — Anton Osika
## Mentioned
Pat Gelsinger, Jason Calacanis, Anton Osika, Andy Grove, Robert Noyce, Gordon Moore, Craig Barrett, Satya Nadella, Sundar Pichai, Steve Jobs, Jensen Huang, William Gibson, Ned, Intel, Nvidia, TSMC, Apple, Microsoft, Google, ASML, P.A. Semi, Groq, Cerebras, d-Matrix, ElevenLabs, Lovable, PsiQuantum, Plaud, Airwallex, Cursor, Claude Code, CodeWorks, Perplexity Computer, Founder University, AWS, Nursa, Salesforce, HubSpot, Slack, Anthropic, OpenAI, CERN, Intel Inside, Pentium, Centrino, Apple Silicon, PowerPC, x86, CUDA, Larrabee, Plaud NotePin, Optimus, Fortnite
## Video description
(0:00) Former Intel CEO Pat Gelsinger joins Jason!
(1:41) What Went Wrong at Intel
(15:19) Why a Taiwan Blockade Would Cripple the US Economy
(25:00) Lovable's Anton Osika: One Million New Apps a Week
(33:38) How Lovable is Bringing Down Builder Costs
Thanks to our partners for making this possible!
Airwallex is a leading global payments and financial platform for modern businesses, offering trusted solutions to manage everything from business accounts, payments, treasury, and spend management to embedded finance. https://airwallex.com/allin
Plaud - If your work depends on conversations — meetings, deal flow, interviews, customer calls — Plaud helps you capture and organize everything with highly accurate AI-generated notes that are not just simple summaries, but also highlight pain points, key decisions, next steps, and customizable summary templates. Check out Plaud at Plaud.ai/allin and use code ALLIN for up to 20% off! Which is also available on Amazon: https://amzn.to/43URLff (Code: ALLIN20X)
Follow Pat:
https://x.com/PGelsinger
Follow Anton:
https://x.com/antonosika
Follow the besties:
https://x.com/chamath
https://x.com/Jason
https://x.com/DavidSacks
https://x.com/friedberg
Follow on X:
https://x.com/theallinpod
Follow on Instagram:
https://www.instagram.com/theallinpod
Follow on TikTok:
https://www.tiktok.com/@theallinpod
Follow on LinkedIn:
https://www.linkedin.com/company/allinpod
Intro Music Credit:
https://rb.gy/tppkzl
https://x.com/yung_spielburg
Intro Video Credit:
https://x.com/TheZachEffect
And a special thank you to @RAISESummit for hosting us in Paris
#allin #tech #news
[Source](https://www.youtube.com/watch?v=-ILKiOU5iAQ)
Source quotes
[0:00] Pat Gelsinger: 在英特尔度过青春期的
[5:16] Steve Jobs: 我已经在过去的四个操作系统版本里全部完成了向 x86 的移植，我看我们自己能搞定。
[9:58] Pat Gelsinger: 如果当时坚持下去，今天的世界格局会截然不同。
[9:58] William Gibson: 街道会为技术找到它自己的用途
[11:33] TSMC: 我不在乎是谁设计的芯片，只要你有图纸，我就是你的制造伙伴
[14:11] Pat Gelsinger: 直接超越 1929 年的大萧条
[14:11] Pat Gelsinger: 这不是假设理论，而是正在发生的现实威胁
[17:27] Pat Gelsinger: 将 AI 的效率提升 10,000 倍
[20:10] Pat Gelsinger: 我们将解决化学难题、攻克自然语言、发明全新材料、重塑交互方式、战胜癌症，并将无数人从贫困中解救出来。没有比现在活着更美好的时代了，而作为技术人员，我们正坐在时代的驾驶座上。
[22:41] Pat Gelsinger: 就在这个十年，2030年之前它必将发挥实质性作用。
[34:53] Anton Osika: 用户因快速构建软件而来，但因经营运转企业而留下
[44:55] Anton Osika: 在 CERN，他们引入了‘竞合（Co-opetition）’的概念——两支彼此独立的科研团队围绕同一个粒子加速器的不同部位展开攻关，在正式发表成果前完全互不沟通。这样既能对比两种技术路径的优劣，又不会过早陷入局部最优解（Local Minimum）的泥潭。
Source quotes
[16:54] Pat Gelsinger: This isn't a theory... their intentions have been clear over a sustained period of time.
[31:12] Jason: She asked Lovable to do it. I gave her some, you know, prompting—human prompting, boss-to-AI... and it built something that I would have never been able to afford to build.
Original transcript
[0:00] spent a long time at Intel.
[0:02] >> Yeah.
[0:02] >> And uh
[0:03] >> only 34 years.
[0:04] >> 34 years.
[0:05] >> Yeah.
[0:06] >> Probably one of the greatest American
[0:08] companies uh ever. And then absolutely
[0:12] went off the rails and got absolutely
[0:16] demolished by Nvidia, TSMC,
[0:21] uh and I guess Apple to a certain
[0:23] extent. So you had this incredible Intel
[0:26] inside moment. We bought our computers
[0:29] based on, you know, hey, the Pentium and
[0:32] that sound
[0:33] >> Intel inside, baby. Intel insideum.
[0:37] And so, let's talk about how things went
[0:39] wrong. What went right and then
[0:42] >> how did it and and you were there for a
[0:45] long time. You took a break and then you
[0:47] came back. But there seemed to be have
[0:50] been some critical mistakes that we can
[0:52] learn from. So, let's just embrace it
[0:54] and go right into it. tremendous success
[0:56] as an American company coming back now I
[0:58] think uh reasonably but when you when we
[1:02] look back on it and we do our
[1:04] post-mortem what were the mistakes a and
[1:07] what would we change in terms of the
[1:09] direction of that company
[1:12] >> if you were building a global financial
[1:14] system from first principles today you
[1:16] wouldn't build it on 50-year-old legacy
[1:18] rails you'd build airwalls one AI native
[1:21] platform for global accounts cards and
[1:23] payments it's designed to make the
[1:24] entire entire world feel like a local
[1:26] market. Others are bolting AI onto
[1:28] broken infrastructure, but Airwall was
[1:30] built for the intelligent era from day
[1:32] one. Stop paying the legacy tax and
[1:34] start building the future at
[1:36] airwall.com/allin.
[1:38] Airwall built for the future.
[1:41] >> Having spent so much of my life there,
[1:43] you know, I view it. I joined when I was
[1:45] 18. I went through puberty at Intel,
[1:48] right? I joke, right? I was just like,
[1:50] you know, I am so early. uh Grove,
[1:53] Noise, Bar uh Barrett, right? Uh and uh
[1:57] uh you know, they they were the people I
[1:59] grew up at, right? You know, so on. They
[2:01] were my mentors. They were the people I
[2:03] adored uh for it and they were deeply
[2:06] technical.
[2:07] >> Andy Grove,
[2:08] >> Andy Grove, Gordon Moore, Bob, you know,
[2:11] co-inventor, you know, these were deeply
[2:13] technical leaders. I remember when I
[2:15] joined the executive staff for the first
[2:17] time, you know, there was, you know,
[2:19] probably 15 of the 20 people that were
[2:21] in the room were PhDs, right? You know,
[2:25] it was just that technical. And, you
[2:27] know, I view one of the things that went
[2:28] off the rail was when it started to be
[2:30] run by business people
[2:32] >> as opposed to technical,
[2:33] >> the bean counters, the finance people.
[2:34] >> Yeah. And you know when I became uh CEO
[2:38] uh in 2001 that was the first technical
[2:41] leader in essentially 15 years
[2:44] >> right you know associated with it you
[2:46] know and if you have a business leader
[2:47] who does he promote business leaders and
[2:49] you know right you know so I think one
[2:52] of the fundamental things is and you
[2:54] know as you look at the great technology
[2:56] companies uh today you know they're
[2:58] deeply technical
[2:59] >> and founderled typically
[3:01] >> you know and even if they're not you
[3:02] know Satcha is not a founder no right?
[3:04] You know, Sundar is not a founder as
[3:07] well, but they're deeply technical
[3:09] individuals. And when you're making
[3:10] these hardcore technical, you know,
[3:13] decisions that affect billions of
[3:15] dollars, you don't do that through a
[3:17] spreadsheet. That's a lousy investment,
[3:20] right? Unless the technology trends make
[3:23] it the right investment. And I think
[3:25] that's one of the fundamental things.
[3:26] And obviously you know in the five years
[3:29] uh five six years before I came back you
[3:31] know Intel gave $100 billion to
[3:35] shareholders.
[3:36] >> Oh the dividends
[3:38] >> and stock buybacks
[3:40] >> a hundred. What I wouldn't have done for
[3:42] another hundred billion dollar on the uh
[3:47] what would you have done? You probably
[3:48] would have made chips for iPhone which
[3:51] Intel passed on. Yeah.
[3:53] >> Yeah. you know, but you know, it hadn't
[3:55] built a new factory in a decade when I
[3:57] got there. It's like, you know, how can
[3:59] you not be building? How could you not
[4:01] buy EUV machines? You know, there's just
[4:03] all of these things, you know, that you
[4:05] would only do as a technologist because
[4:08] the economics behind them by themselves
[4:10] were not good.
[4:11] >> So, you know, it's getting back to the
[4:13] core of technology to me that was, you
[4:16] know, the fundamental thing. You know,
[4:18] you make good decisions, you make bad
[4:19] decisions as leaders. Every business uh
[4:22] does that uh as they go along but uh you
[4:25] know fundamentally this is a technology
[4:27] business and you need technologists
[4:29] running technology uh that then hires
[4:33] technologists that are sitting at the
[4:34] staff that then hire the best
[4:36] technologist you know you know
[4:38] >> and take big swings at you know
[4:40] categories that could matter in the
[4:43] future like skating to where the puck's
[4:44] going. If you look at Apple, they did
[4:45] the same thing for the past 15 years,
[4:48] buying back the stock, tremendous amount
[4:50] of dividends. They're the largest holder
[4:52] of capital of any company, I believe, to
[4:54] this date. And what if what companies do
[4:58] they buy? They buy little tiny
[4:59] acquisitions on the margins. I think the
[5:01] largest ones was was Beats cuz they
[5:04] wanted to get inroads into, you know,
[5:06] certain uh demographic segments like in
[5:08] the Android space that they couldn't get
[5:10] into. But my god, what a colossal waste
[5:12] of time. Like you said, they could have
[5:13] done so many amazing things. Tell me
[5:16] about Steve Jobs
[5:18] in 2008, 2009 deciding, I think we're
[5:22] going to make our own silicon and that
[5:23] impact because was that a covert product
[5:26] project or did you guys know he was
[5:28] doing that? Did he inform you?
[5:30] >> Well, that seemed to be another one of
[5:32] those forks in the road. Yeah,
[5:33] >> you know, Steve was an incredible
[5:35] leader. You know, he was also a ruthless
[5:37] leader, right? You know, very difficult.
[5:39] you know, read Walder Isaxson's book on
[5:42] him uh as well. I had many many
[5:44] conversations with Steve over the years
[5:46] uh you know, for it. Um but you know,
[5:48] when they moved to Intel and the
[5:50] Centrino chip, it was a big deal.
[5:53] >> Yeah.
[5:53] >> Right. And they were putting
[5:55] extraordinary demands on Intel. You
[5:58] know, make the chip smaller, drive lower
[6:00] power. They're demanding uh customer.
[6:04] And when he was no longer convinced that
[6:06] we could continue to do that, you know,
[6:09] he started the project,
[6:10] >> right? You know, and if you remember
[6:12] what was it, you know, you know, uh, P
[6:15] semi, you know, they acquired some small
[6:17] companies, started to build some
[6:18] competency, but, you know, they did a
[6:20] few little chips internally. It wasn't a
[6:22] big deal and then the little chips got a
[6:24] little bit bigger, you know, and Steve
[6:25] was a master of this, you know, just
[6:28] starting, you know, these small efforts
[6:30] to build core competence inside the
[6:32] company. Uh I remember when we had the
[6:34] first conversation with Steve about uh
[6:37] porting the uh operating system to the
[6:42] Intel chip from the power chip that they
[6:44] were running on before they moved to
[6:45] Intel. And we were quite proud of the
[6:49] silicon software competencies that we
[6:52] had and compilers and operating systems,
[6:55] you know. So Steve, we'll help you port
[6:57] the operating system to the x86. And I
[7:00] remember that Steve said, "I've been
[7:02] working on that for the last four
[7:04] releases."
[7:06] >> He had been preparing the core
[7:08] technologies inside of Apple for
[7:11] something that might happen uh in the
[7:14] future, you know, and he was already,
[7:15] you know, to me, I just remember I was
[7:16] just shocked. I, you know, I've ported
[7:18] the last four releases to the x86. I
[7:20] think we got this.
[7:21] >> Yeah.
[7:22] >> Right. You know, it was that kind of
[7:23] thing. And that's how they got into the
[7:24] semiconductor, you know, doing their own
[7:26] semiconductor. Hm. I'm not sure I can
[7:28] rely on Intel to be that much ahead of
[7:31] the industry and I can start optimizing
[7:34] the system design with the silicon
[7:36] design as opposed to relying on one
[7:38] that's been somewhat optimized for a
[7:40] Windows environment versus an iOS
[7:42] environment, you know, in their uh
[7:44] operating system. And you know, it was
[7:46] just, you know, uh, you know, it was
[7:49] never that kind of thing. They sort of
[7:51] said, you know, right, you failed as a
[7:53] supplier. No, I can supply myself
[7:55] better.
[7:56] >> Yeah. and Jensen uh decides, hey, he's
[8:00] going to go all in making these video
[8:01] cards and talk about just incredible
[8:06] uh serendipity that these happen to be
[8:09] also very applicable for cryptocurrency
[8:12] and running these AI jobs.
[8:15] >> Yeah.
[8:16] >> Was that luck or skill or combination of
[8:19] both there? Well, you know, when you
[8:21] think about that progression, you know,
[8:23] Jensen, he was just building high
[8:25] performance computers, you know,
[8:27] throughput machines. You know, when we
[8:29] were at the height of our strength on
[8:32] CPUs, uh, at Intel, we sort of scoffed
[8:35] at his machines. Yeah.
[8:36] >> Right. You like, oh, that's a graphic
[8:38] machine. You, you know, there's some
[8:40] gamers who want to use that kind of
[8:42] stuff, right? You know, it was always
[8:43] the big CPU and those little GPUs. But
[8:46] when they started to build a real
[8:48] software stack, Yes. with it, right? You
[8:50] know, sort of, okay, this CUDA thing and
[8:52] SIMT as a technology, you know, uh, you
[8:55] know, uh, multi-threading and so on. And
[8:58] it just sort of kept getting a little
[8:59] bit better and a little bit better and
[9:01] it was a little bit jobslike in that
[9:03] way. You know, we're just making it
[9:04] better every release and it's becoming
[9:07] more robust and all of a sudden, you
[9:10] know, the crazy, you know, uh, Japanese
[9:13] HPC guys said, "Hey, we could take those
[9:15] graphics cards and maybe start using
[9:16] them in HPC." H,
[9:18] >> right? you know, and that was sort of
[9:20] defining moment where it wasn't just
[9:22] about doing graphics anymore. This was a
[9:25] more computationally dense platform to
[9:28] start attacking some of the world's most
[9:30] interesting workloads. And I think
[9:32] Jensen would agree that was a defining
[9:34] moment and them sort of saying, "Oh,
[9:36] these aren't just graphics cards
[9:37] anymore. You know, these are
[9:39] generalpurpose computing devices that
[9:42] can start applying to these other uh
[9:44] workloads." And you know AI was you know
[9:47] had gone through what its fifth nuclear
[9:49] winter by that point. We're just like
[9:51] man you know you know this is never
[9:53] going to matter right we're never going
[9:54] to you know get the breakthroughs but
[9:56] the community around it was continuing
[9:58] to develop uh you know for it and the
[10:01] CUDA software kept getting better uh
[10:04] generation by generation and uh you know
[10:07] I had a project at Intel Larabe right
[10:09] where we were trying to take the x86 and
[10:12] essentially do the same thing right you
[10:14] know for it and you know in my first
[10:16] departure from Intel the project was
[10:18] killed a week after I left
[10:20] >> huh
[10:21] >> and the world would have been so much
[10:23] different right I
[10:24] >> I mean it really I think it's
[10:27] illustrative of illustrative of what
[10:30] continuous innovation taking some risks
[10:32] and doing that fundamental research and
[10:36] the compounding power of technology
[10:37] because I think it was William Gibson
[10:39] who said the street finds its own use
[10:40] for technology like
[10:43] Nvidia did not predict that this Bitcoin
[10:46] project would take over and that this
[10:48] would be the best way to do those
[10:49] computations, nor did they anticipate, I
[10:52] think, you know, that AI would take off,
[10:54] but because it was the best solution,
[10:56] the hacker community could kind of
[10:58] figure that out. Well, as we wrap on the
[11:00] Intel portion of your uh career, um,
[11:04] okay, Apple Silicon, that's one. Uh, and
[11:07] then you have Nvidia, and then you have
[11:09] this Taiwanese company, uh, that starts
[11:13] making, you know, really great at
[11:15] fabricating the these chips. Um and
[11:19] Intel missed that as well. Yeah. And and
[11:21] maybe you could talk a little bit about
[11:23] TSMC and their surging and we can even
[11:26] get into a little bit of the the
[11:27] politics of it now and then we'll get
[11:29] into some of these AI chips and venture
[11:31] investing. You know the thing with TSMC
[11:34] was they started with a vision of
[11:36] foundry
[11:37] >> right you know they were going to become
[11:39] the factory for the industry and again
[11:41] these factories are so expensive 20
[11:44] billion 30 billion and uh the
[11:47] engineering and the continuous
[11:49] investment required to do it and you
[11:51] know it was a stunning you know vision
[11:54] uh at that point in time Intel was IDM
[11:57] as we called it the integrated design
[11:59] and manufacturing you We never worked to
[12:03] make our process and our factories
[12:06] available for third parties
[12:08] >> right you know it was always this thing
[12:09] hey it's you know we do enough CPUs
[12:12] oursel you know we reuse it for chipsets
[12:15] and some of the other things that we're
[12:16] doing but it was never standardized in a
[12:19] way that it could be made available for
[12:21] a broad ecosystem you know using PDKs
[12:24] and all the design tools you know we did
[12:27] a lot of our own EDA tools ourself you
[12:29] know one of the projects that I started
[12:31] early in my career was the foundations
[12:33] of EDA, right? Uh as well, the first
[12:36] place and route, you know, the first
[12:38] standard cells, the first highle
[12:40] description language, you know, it was
[12:42] so proprietary and TSMC basically cut
[12:45] that in half and says, I don't care
[12:47] whose chip it is. I don't care what
[12:48] you're designing, I'll be your
[12:50] manufacturing partner. And at the time,
[12:52] that was such a trivial piece of the
[12:55] business, Intel didn't even care,
[12:57] >> right? You know, so on. And then over
[13:00] steady progress over a long period of
[13:02] time and Apple as a customer driving
[13:06] them to be could be become really
[13:08] meaningful. You know obviously the world
[13:10] changed and when I came back uh to uh
[13:13] Intel in 2001 TSMC was producing 5x the
[13:17] wafers of Intel.
[13:18] >> Wow.
[13:19] >> Right. Not 10% more 5x.
[13:22] >> Yeah. And all of a sudden that model of
[13:25] foundry became the model of the
[13:27] semiconductor industry with two
[13:29] exceptions Intel and memory. you know,
[13:32] memory design and manufacture, right,
[13:35] for you know, that is uniquely
[13:37] different. And obviously, you know,
[13:38] we're seeing the, you know, $3 trillion
[13:41] memory companies just extraordinary, you
[13:43] know, and, you know, trillion dollar
[13:45] foundry company uh in TSMC. You know,
[13:48] the industry has said, I want a lot of
[13:50] wafers. I want a lot of innovation of
[13:53] different designs. I have a layer of
[13:55] standardization and EDA tools. And the
[13:58] world changed. And obviously as I came
[13:59] back to Intel that was one of the core
[14:01] thesis of the new strategy. Yeah. We
[14:03] must become a foundry as well five to
[14:06] one and now it's more like seven to one
[14:07] in terms of wafers you know to TSMC to
[14:11] and
[14:11] >> are we going to be able to onshore that
[14:13] obviously we had the chips act and just
[14:15] give us broad strokes what you think is
[14:17] going to happen here in terms of
[14:19] obviously Taiwan is in play. Some people
[14:22] in the administration believe um it's
[14:25] going to happen the year after Trump's
[14:27] out unless he takes his third term.
[14:29] Other people believe like it was going
[14:32] to happen as early as 27 uh or maybe
[14:35] going into 28. So,
[14:38] are we going to be able to replicate
[14:41] that here in America in a reasonable
[14:42] amount of time, or is this like truly
[14:44] could be a cataclysmic event if,
[14:47] you know, god forbid, China decides,
[14:49] hey, we're going to blockade um Taiwan
[14:52] and and the Taiwanese decide, yeah,
[14:54] we're going to burn the fabs and we're
[14:56] going to fly out all of the engineers
[14:58] and ship them to America.
[15:00] >> Well, there's a lot in that question,
[15:02] you know. Do we have an hour to talk
[15:04] about this question? Well, I mean we
[15:05] have six minutes, but Okay. Um, yeah, do
[15:08] the best you can.
[15:09] >> Okay.
[15:11] >> I want to talk also about the AI bubble.
[15:13] >> So, super, you know, three things about
[15:14] this super quick. You know, one is the
[15:16] chips act is having benefit.
[15:17] >> Yeah.
[15:18] >> Right. You know, when we started the
[15:20] chips act in, you know, in 2001 when I
[15:22] came back, the US was building about 12%
[15:25] of leading edge. Today, that number is
[15:27] more like 18%.
[15:29] >> Okay. You know, we're making progress.
[15:31] It's not 50%. We have a long way to go,
[15:34] right? You know, Intel is starting to be
[15:36] a real foundry. Okay, that's real
[15:38] progress. Uh, and TSMC's factories are
[15:41] up and operating at scale, right? We
[15:43] have Samsung and, you know, uh, as well.
[15:46] But, you know, I say the Intel and the
[15:48] TSMC progress. Okay, that's meaningful.
[15:51] Now, let's make it ugly for a second.
[15:53] Uh, the island of Taiwan has less than 3
[15:56] weeks, a big article in the Wall Street
[15:58] Journal two weeks ago on this, less than
[16:00] 3 weeks of energy reserves.
[16:04] Wow. Okay, that should just put a chill
[16:08] in everybody's spine, right? Because the
[16:11] blockade after 3 weeks, the island
[16:14] browns out. When you turn off a fab, it
[16:17] doesn't come back on for 90 days, right?
[16:20] The economic impact of a brown out of
[16:23] Taiwan is greater than the Great
[16:25] Depression, right? Uh in the world,
[16:27] never do you need to do anything a shot
[16:30] to be fired. You just need to say,
[16:32] "Great, no energy for 3 weeks.
[16:33] >> No oil." Yes.
[16:35] >> Right. No LG, right? You know, that's
[16:36] how the island run. That is scary, you
[16:39] know, to me. We need more resilient
[16:41] supply chains uh associated, you know,
[16:44] with it. And I don't think this is an
[16:46] alternative for the world because if it
[16:48] really does become a risk, you know, and
[16:50] I'm, you know, I, you know, I don't sit
[16:52] in the situation room and get all the
[16:54] data and so on, but let's remind each
[16:56] other that I think China has blockaded
[16:58] the Taiwan Straits seven times over the
[17:00] last four years.
[17:02] >> Yeah.
[17:02] >> This isn't a theory.
[17:04] >> No, no. They're running exercises.
[17:05] They're being pernitious and
[17:08] >> right
[17:08] >> pretty provocative in terms of
[17:10] >> Is that 2027? Is that 2030? Is that
[17:12] 2035? their intentions have been clear
[17:15] over a sustained period of time. We need
[17:17] more resilient uh supply chains, you
[17:20] know, forward. So, something, you know,
[17:21] I put a lot of my time and energy into
[17:23] and we're making progress, but we need
[17:25] to go faster, need to go more
[17:27] meaningful.
[17:27] >> Yeah. And let's talk a little bit about
[17:29] the AI buildout. I mean, you watched the
[17:32] PC revolution, servers, the internet.
[17:35] These were all extraordinary buildouts.
[17:37] And then this is the buildout to end all
[17:41] buildouts. the amount of data centers,
[17:43] the amount of chips, the amount of
[17:44] inference needed.
[17:46] Do you think it's a bubble? I think I've
[17:48] heard you say like it's it's obviously a
[17:50] bubble, but what what's the risk factor
[17:54] here? That we build too much uh or that
[17:57] the technology doesn't solve enough
[17:59] problems and we are swimming in tokens?
[18:02] What what worries you about what you're
[18:04] seeing now? the valuations of these
[18:06] companies has gotten quite
[18:07] extraordinary. And you know, if they
[18:10] build too much and they spend too much
[18:12] money and they don't make enough money,
[18:14] well, based on your experience with
[18:17] running a company, a public one, that's
[18:20] a lot of tension on it. When you don't
[18:22] make as much money as you're spending,
[18:25] people tend to fall out of love with
[18:26] these stocks. Yeah.
[18:28] Well, I do think there, you know, there
[18:30] there is a silver lining here that
[18:33] guarantees we don't get too far ahead of
[18:36] oursel in terms of bubble, you know, and
[18:39] that is energy capacity.
[18:41] >> Right.
[18:41] >> Right. You know, energy capacity in the
[18:43] world is expanding four 5%. You know, in
[18:46] the US we had a decade at 1%. Right. You
[18:50] know, I mean, it's just hideous what we
[18:51] did to our energy grid, you know, over
[18:53] about a decade and a half. But now
[18:55] that's getting built out. But
[18:57] essentially, nobody's going to build and
[18:58] buy GPUs and build data centers if they
[19:01] don't have energy.
[19:02] >> So essentially, you have an upper bound
[19:04] on how aggressive and how hyped and
[19:07] bubbled that we get. So I take a lot of
[19:10] soloulless in that.
[19:11] >> Yeah.
[19:11] >> Right. You know, for it because what
[19:13] then is the incremental value of a token
[19:16] and if it's a measure of intelligence,
[19:18] it's somewhat infinite, right? You know,
[19:20] in the sense if I have more
[19:22] intelligence, I will do, you know,
[19:24] better supply chain. I will do better
[19:26] finance. I will do more, you know,
[19:28] efficient logistics. I will, you know,
[19:30] all of those things. So to me, the the
[19:33] potential value that we unleash in a
[19:36] token economic world is somewhat
[19:38] infinite, right? And particularly with
[19:40] labor shortages and so on that we see
[19:43] right in uh developed countries, I am an
[19:45] optimist, you know, that we're in a
[19:48] couple of decade buildout.
[19:50] >> Wow.
[19:50] >> Right. Not a couple of years, a couple
[19:52] of decades. One of the big objectives
[19:55] I've said is that I have to make AI
[19:58] 10,000x
[19:59] better,
[20:00] >> right? You know, it's way too expensive
[20:02] today. you know, we want to drop, you
[20:04] know, by five orders of magnitude the
[20:05] cost per token, you know, the energy,
[20:08] you know, per token so that we really do
[20:10] have Jevans law that we just explode the
[20:12] access to AI, right, in much more
[20:15] economic uh ways,
[20:17] >> which it does seem like Jevans uh
[20:19] paradox has been at play over the last
[20:21] year, like, oh my lord, these tokens are
[20:23] so cheap and the tools are getting so
[20:25] good. Yeah, I'm just going to start
[20:27] using these tools all day long until the
[20:29] bill comes in and you're like, "Okay,
[20:30] yeah, maybe I need to get some ROI out
[20:32] of this." But you do have these
[20:34] incredible companies, Cerebrris, Grock,
[20:36] etc. making inference
[20:37] >> dematrix
[20:39] silicon and so you know, and you know,
[20:40] if we accomplish right, you know, these
[20:43] orders of magnitude improving and token
[20:46] economics, availability, reduction in
[20:48] energy costs associated with it. You
[20:50] know, we just have a fantastic couple of
[20:53] decades in front of us. There has not
[20:55] been a time in human history where it's
[20:57] been better to be a technologist than
[20:59] the one we're in right now. We will
[21:01] solve chemistry. We will solve language.
[21:03] We will, you know, invent new materials,
[21:05] re, you know, new forms of, you know,
[21:08] uh, interaction, you know, uh, killing
[21:11] cancer, right? Lifting people out of
[21:13] poverty. There is not a better time to
[21:15] be alive than the one that we're in
[21:17] right now. And as technologists, we get
[21:19] to sit in the driver's seat of it.
[21:21] >> Pretty amazing. and you're investing uh
[21:23] and that's your passion. Now what do you
[21:26] think of these valuations? It's quite
[21:29] seems a you know if you live through the
[21:31] dotcom bubble we did see a disconnect
[21:33] there. These companies slightly
[21:35] different. We just had 11 labs up 600
[21:38] million in revenue. Lovable I think
[21:40] they're at five or 600 million. So
[21:42] that's quite different than the do
[21:44] speculation. Yeah.
[21:45] >> Yeah. Well fundamentally we have real
[21:47] revenues you know real margins coming
[21:49] out of these businesses as well. You
[21:51] know that said anytime the multiples get
[21:52] too high okay some corrections you know
[21:55] and to me periodic corrections that keep
[21:57] the multiple you know earnings multiples
[22:00] and you know so on in reasonable things
[22:02] is good because this will not be a
[22:03] smooth curve you know I'm predicting two
[22:06] decades of goodness and there's going to
[22:08] be lots of disruptions along the way
[22:10] it's not going to be a smooth curve and
[22:12] every time we have one of those
[22:13] corrections say thank you right we're
[22:15] not letting the bubble get ahead of
[22:16] itself right you know hey we had the SAS
[22:19] apocalypse there's going to be other
[22:20] apocalypses on that journey when when
[22:23] industries get impacted by the
[22:26] capabilities that will be unleashed and
[22:28] that's even before it gets exciting and
[22:30] what I call the trinity of computing
[22:32] classical computing AI computing and
[22:35] quantum computing and when those three
[22:37] come together okay that's when things
[22:39] get really exciting
[22:41] >> quantum's been about 5 years away for 25
[22:44] years um when is it actually going to do
[22:48] anything meaningful
[22:48] >> this decade this decade. So by 2030,
[22:51] >> yep,
[22:52] >> it'll be meaningful. What should we
[22:54] expect in terms of its impact in 2030?
[22:56] Like
[22:57] >> you know, you're going to be able to
[22:58] start doing things that cannot be
[22:59] computed today.
[23:01] >> You know, chemistry, you know, biology,
[23:03] there will be things that can't be
[23:05] computed today. You know, some of the
[23:06] easy things will be some of like the
[23:09] logistics where I will compute the best
[23:11] answer to get this thing to you, right?
[23:14] >> Traveling salesman problem,
[23:15] >> right? you know all of a sudden all of
[23:16] those problems uh obviously it's
[23:18] probably going to be you know 2020 2032
[23:21] 2033 when we solve you know things like
[23:24] encryption right you know where you know
[23:26] you'll have the fundamental Qday you
[23:29] know kind of implications but this
[23:30] decade we will see quantum supremacy uh
[23:33] results across multiple industries you
[23:36] know we know how to build cubits we know
[23:38] how to error correct cubits we now have
[23:40] algorithmics right against uh quantum
[23:43] and you know now it's just about
[23:45] engineering scale.
[23:46] >> Who's going to win?
[23:47] >> Well, obviously I'm a SI quantum guy,
[23:50] right? Since that's one of our portfolio
[23:51] companies. But the thing that you're
[23:53] seeing is that you now have like four,
[23:56] five, six modalities of quantum that are
[24:00] demonstrating pretty good results,
[24:02] right? You know, across trapped ions,
[24:04] across, you know, photonic uh
[24:06] approaches, spin uh approaches. So, you
[24:09] now say modality is not an issue. Error
[24:11] correction's been proven uh across them.
[24:14] And you know, I think the race will be
[24:16] on and my prediction is meaningful
[24:18] results before 2030.
[24:20] >> Wow. You realize that's about 40 months
[24:22] from now.
[24:23] >> Yeah. Okay. Meaningful results. Thanks
[24:25] so much, Pat, for sharing all this
[24:27] incredible uh
[24:28] >> information and knowledge. Great to see
[24:30] him.
[24:30] >> Very good.
[24:34] >> Your most valuable conversations rarely
[24:36] happen at a desk. The hallway sink, the
[24:38] dinner, the quick founder call. Plaude
[24:40] no pins clips on and captures all of it
[24:43] hands-free. Afterward, Plaude
[24:45] intelligence turns the recording into
[24:47] clean notes and clear next steps. You
[24:50] stay in the room, Plaude handles the
[24:52] rest. For people who live in meetings,
[24:54] that's real leverage. Wear your Plaude
[24:56] at plaud.ai.
[25:00] >> Oika is one of my favorite founders.
[25:03] He's the founder of Lovable. Why do I
[25:05] love this founder? Uh well he's built a
[25:08] product that people are addicted to
[25:10] primarily Anton the people who work for
[25:12] me and I uh love talking to you because
[25:16] as a founder you have a northstar you're
[25:19] incredibly laser focused on enabling
[25:22] anyone to build great software. Yeah
[25:25] it's the mission of the company. I'm
[25:27] paraphrasing here but
[25:28] >> essentially that's the mission of
[25:29] lovable
[25:30] >> mission I talk about empowering humans.
[25:32] empowering humans
[25:33] >> and the first gap is to build a product.
[25:36] >> The second gap is to build a business
[25:38] around that product, right?
[25:39] >> And now at everyone at Lovable, we're
[25:41] we're working on both of these two gaps,
[25:43] right?
[25:43] >> The first one, we got very far. We're
[25:46] seeing a million new projects built
[25:48] every single week on the
[25:49] >> incredible. And on the on the second
[25:51] one, we're investing a lot in making it
[25:53] easier to run your business and to get
[25:56] people to care, people to discover what
[25:58] you build and the entire business of
[26:00] what you're whatever you're doing as a
[26:01] small business. As if you're a large
[26:03] business, we're also getting a lot of
[26:05] traction. Um, and uh we're actually
[26:08] seeing as a proof of that more than 700
[26:12] million visits to the applications every
[26:14] month. So every month there's um extreme
[26:17] growth in in the surface area of of the
[26:19] entire
[26:21] more than 50 million apps built on the
[26:23] platform to date.
[26:24] >> How many years has lovable been in
[26:26] market or how many months now?
[26:28] >> 20 months since 20. Yeah. And and again
[26:31] we're seeing people who are first-time
[26:33] founders. We're seeing enterprise
[26:35] leaders move much faster together with
[26:38] their teams on this platform that has a
[26:40] lot of opinionated pieces in how you
[26:44] should uh create software and how to
[26:46] operate that software and how the
[26:48] different applications in your company
[26:50] connect to each other over time. So
[26:51] that's what why we're seeing so much
[26:53] growth also on the enterprise side which
[26:55] where where we actually growing fastest
[26:57] right now. This is really interesting
[26:58] because 10 years ago, uh, people were
[27:01] doing wizzywig software. Um, what was
[27:05] the name for it? Before Vive, no code,
[27:08] low code. Yes. And when I saw that 10
[27:11] years ago in my incubator, you know,
[27:14] every 20th company, somebody would come
[27:16] in who was an MBA or not a developer and
[27:19] they had vibecoded something and um not
[27:22] vibe coded, they had no coded and they
[27:25] were using these different software
[27:26] platforms and the software didn't look
[27:29] good. It didn't work perfectly well. It
[27:31] was slow, but the promise was there. And
[27:35] I guess it took LLMs and this new
[27:38] intelligence
[27:40] to make actually good software. So maybe
[27:42] you could talk a little bit about
[27:46] who is the customer because developers
[27:50] uh do developers use lovable or is it
[27:53] the other 95% of society that are your
[27:56] customers? How do you think about who
[27:57] your ideal customer profile is?
[27:59] >> Yeah, we're seeing people use Lovable
[28:02] both with a technical background. about
[28:04] 20% are technical or some type of
[28:06] engineer and they they love that we're
[28:09] quite opinionated. We put all the best
[28:11] practices into how the software is
[28:13] architected and we make it seamless to
[28:16] um with one prompt get payment set up in
[28:18] a very secure way and do things like run
[28:22] security scans after every change even
[28:24] now in the background monitoring the
[28:26] projects. So it's actually quite
[28:29] appreciated by the engineers in the
[28:31] technical community. Um also because
[28:33] it's a great bridge
[28:35] >> from the nontechnical people which is
[28:37] four out of five are are nontechnical uh
[28:39] and they're building uh often first to
[28:42] figure out what is the right thing to
[28:43] build
[28:44] >> which is where lovable has always been
[28:46] exceptionally good and uh now what we're
[28:49] seeing is that people are running
[28:53] businesses making more than million
[28:54] dollars of revenue on the on this
[28:56] platform. So it's it's this we're
[28:58] building for everyone. It's this entire
[28:59] spectrum. And what's what's exciting to
[29:01] see is often that if someone who
[29:03] discovers lovable from their colleagues
[29:04] at a large company, they go out and then
[29:07] run a side hustle and some of those
[29:10] those idols hustles really work. They
[29:11] make hundreds of thousands of dollars
[29:13] and then they become a founder after
[29:14] that. So there's this crosspollination
[29:16] from both.
[29:17] >> Yeah. And this is like the really
[29:18] interesting thing about vibe coding. If
[29:20] we were sitting here last year, people
[29:21] would look at it and say it's a great
[29:23] way to make a mockup. like you said, a
[29:25] great way to think about product and
[29:27] maybe create wireframes or a workable
[29:31] prototype. All of that's out the window
[29:33] now. The whole concept of building
[29:36] wireframes and building a mockup, well,
[29:40] you can just go right to building the
[29:41] product in a day or two days. And what
[29:46] people I think don't appreciate about
[29:47] what you're doing at Lovable is after
[29:50] you've made a product that you're proud
[29:52] of and that has some product market fit,
[29:55] there are many more steps that are
[29:57] required. You mentioned payments, you
[30:00] mentioned security, uh making sure that
[30:02] the data isn't lost or that it's not
[30:04] leaked.
[30:06] That's changed dramatically over the
[30:08] last 12 months. Yeah,
[30:09] >> very much so. So um I would say many
[30:12] engineers they don't look at the code
[30:14] they don't write code anymore and that
[30:17] means that you don't need to be an
[30:19] engineer to create software right um but
[30:21] the the thing that lovable does for any
[30:24] anyone also the non-technical people is
[30:26] that it it um takes uh creates a
[30:30] structure for the architecture of the
[30:31] software that you build and it makes
[30:33] sure that you don't go off a cliff um
[30:36] and that things like setting up payments
[30:38] emails things like getting discovered by
[30:41] other AI chat engines and uh by Google
[30:45] search those things are kind of taken
[30:47] care of. So you don't have to know how
[30:49] all these things work in the details you
[30:52] trust you can trust the platform to take
[30:53] care care of data security connecting to
[30:57] other tools that you might be using in a
[30:59] secure way and and that's really where
[31:01] um us being opinionated from day one and
[31:04] being focused on making this for the
[31:06] 99%. It's a it's a vast market right
[31:09] from the from day one is what made us
[31:11] very successful.
[31:12] >> Yeah. And I can tell you internally I
[31:14] gave my team all the different tools
[31:16] they could possibly want to use and
[31:19] somebody had started with lovable. I
[31:21] think I told you the story when you were
[31:23] on this week in startups a year ago like
[31:25] and they made some interesting websites
[31:28] and they were trying to make an
[31:29] internet. They couldn't quite get it
[31:31] done. Then I had some people who started
[31:34] using you know cursor or clawed code.
[31:37] They started vibe coding stuff but they
[31:39] couldn't finish the product. And then
[31:41] people tried to solve some problems with
[31:43] codework. I really like perplexity
[31:45] computer. And then my team came to me
[31:48] and for one of our projects I was
[31:51] talking to you about founder university
[31:52] our pre-acelerator. They wanted to make
[31:54] an internet.
[31:56] Um, now this is something I would have
[31:58] never okayed because it would have cost
[32:00] $500,000
[32:02] 10 years ago to make it and we don't
[32:04] have that kind of budget. You know, we
[32:05] would rather put that towards the
[32:06] founders in the program and getting more
[32:09] people into the program. And in 4 to 8
[32:12] hours, they made the whole internet and
[32:15] they made a bunch of things I hadn't
[32:16] asked for. And it was the person running
[32:19] the um this founder university who made
[32:23] it and she did it on her own without
[32:27] uh permission in lovable. I said, "Whoa,
[32:29] where how did you build this?" She said,
[32:30] "Loveable." I was like, "Oh, we still
[32:33] have lovable." And they're like, she's
[32:34] like, "I just put it on my corporate
[32:36] card." To your point, she made it. Now
[32:39] that software is driving the program and
[32:45] the reason people do the
[32:48] uh the the program in their country, we
[32:50] have it in Saudi and in Japan is because
[32:52] it has economic impact.
[32:54] >> So I said, "Hey, I have an idea. Can you
[32:56] make for me an economic impact of the 50
[32:59] companies that are in the program?"
[33:02] She asked Lovable to do it. I gave her
[33:04] some, you know, prompting, human
[33:07] prompting boss to now it has the
[33:09] economic impact in there and it
[33:12] considered, you know, with our
[33:13] prompting, well, how many people work at
[33:15] each company? What are they paying
[33:17] taxes? How much do they rent their home
[33:19] for? What is their average salary? And
[33:21] it built something that I would have
[33:23] never been able to afford to build. And
[33:25] lovable is 50 bucks a month, I think. I
[33:28] don't know how much you charge, but it's
[33:30] far too little. Like
[33:32] >> $50 a month, I think.
[33:33] >> Yeah. That's if you're on a business
[33:35] plan. Yeah. It starts at 25.
[33:37] >> Yeah. So, uh, the economic impact of
[33:40] what you're building is I would equate
[33:44] for what you built to us, it would have
[33:45] cost me $500,000 2 years ago. It was
[33:48] built in 4 hours by an employee, which
[33:51] if you just put employees at 50, 60,
[33:53] whatever, $70, uh, plus the cost of your
[33:56] software, it got made for less than
[33:58] $2,000
[33:59] >> in a year. It's extraordinary.
[34:02] I I'd love to hear more about the
[34:04] progress of the of the internet.
[34:06] Anything that you asked for that you
[34:08] want to forward directed to me.
[34:10] >> Uh well, right now, you know, my concern
[34:13] was security and making sure that data
[34:16] didn't leak and they talked to your team
[34:18] and they went through it and
[34:20] >> it's secure. So, we feel good about it.
[34:23] >> Well, look, um I'm now asking people who
[34:27] do penetration testing to say, I want
[34:29] you to compare all the tools. Yeah. and
[34:31] uh make sure that there's a all the work
[34:33] that we're doing that's not visible on
[34:35] security and trust. Yeah, there's a lot
[34:37] a lot a lot of things um where we we
[34:41] invest and spend money on that every
[34:43] also free users get a lot of security
[34:45] scanning running in the background that
[34:47] that actually um translates to something
[34:50] that security experts can can see. And
[34:53] >> a year ago we were at mock-ups. Now
[34:55] we're at functionality and secure and
[34:58] super viable for deployment. Where will
[35:01] you be in a year?
[35:03] >> Yeah. So, what we're seeing is that
[35:05] there's a gap in build being able to
[35:07] build the product, right? And and you
[35:09] built an entire internet on the
[35:11] platform. That's great. Um, what we've
[35:13] done since then is to have a new product
[35:16] line basically the hosting part which is
[35:18] both the AI and you know all the normal
[35:20] hosting and that product line has been
[35:22] going faster than the building uh thing.
[35:24] I mention
[35:25] >> AWS competitor.
[35:27] >> It it lets you it lets you run all your
[35:30] software and then we're working with
[35:31] companies like AWS under the hood as
[35:33] well. But but what you also want to have
[35:36] is um to use lovable we're seeing by our
[35:40] customers as an AI co-founder,
[35:42] >> a partner that you talk to about
[35:43] everything in your business. And if
[35:45] you're running your apps, your tools are
[35:48] on the platform, then just talking to
[35:51] Lovable has access to all the data that
[35:53] you might want to know about your about
[35:56] your company, how it's doing. So, we're
[35:59] we're working with some of our customers
[36:02] in pre-release to give them access to a
[36:04] co-founder that works for you even when
[36:06] you're sleeping
[36:07] >> and comes back to you in the morning and
[36:09] says like, "Here are some strategic
[36:11] directions you could go. here's some
[36:13] optimizations you can do go in terms of
[36:15] growing your business faster serving
[36:16] your customers better uh faster uh and
[36:20] and and that's um that evolution towards
[36:23] operation and intelligence for towards
[36:26] driving towards outcome
[36:28] >> for your business
[36:30] to build the software but you stay to
[36:32] build the business
[36:33] >> yes to operate your business and um what
[36:36] we're already doing I've been doing for
[36:38] a very long time is to compound from
[36:40] everything we're learning every time
[36:42] lovable makes a mistake. Uh it goes to a
[36:44] gentic system with our engineers in it
[36:46] improving it. That compounding
[36:48] intelligence is of course applicable to
[36:52] our our customers, our users running
[36:54] their business on our platform as well.
[36:56] >> Is software going to become
[36:59] 100% bespoke even like the internal
[37:02] tools. I was looking at Slack and our
[37:06] bill for Slack even on the highest
[37:07] version is maybe $10,000 a year. It's
[37:10] not a lot of money. It's well worth it.
[37:13] But I was starting to think, well, maybe
[37:15] I should vibe code my own Slack so it's
[37:18] integrated into everything we do at a
[37:20] deeper level. So how do you think the f
[37:23] what do you think the future will look
[37:24] like in terms of some of these
[37:28] you know uh foundational pieces of
[37:31] software that every startup every
[37:32] enterprise uses Salesforce HubSpot
[37:36] Slack
[37:38] uh the Google suite Microsoft Office
[37:41] will bespoke software
[37:44] start to replace those do you believe
[37:46] >> I I like this question let let me ask
[37:48] answer it but I'll just give you a story
[37:51] about someone I recently heard who's
[37:52] going on this journey. They're quite
[37:54] advanced. So,
[37:56] >> NAD, he works at a pretty large company
[37:59] in the US, Nursa, and uh he came to our
[38:02] platform because he wanted to build out
[38:04] a new product lines, nurse study for
[38:06] educating more nurses, right? And and he
[38:09] built out all the admin tools for the
[38:12] program, the scheduling for the nurses
[38:14] getting getting their licenses and their
[38:16] certification management. and he was
[38:18] able to build that into a product and to
[38:20] take it to market because they have they
[38:22] have had all that access to nurses
[38:24] wanting their certification. What he
[38:26] also did was he took it into the back
[38:29] office internally and they've now
[38:31] replaced more than 10 tools that they
[38:33] had bespoke applications and um I think
[38:38] in terms of your question you can do
[38:40] that for multiple reasons. In their case
[38:42] they're saving more than a million
[38:43] dollars per year,
[38:44] >> right?
[38:44] >> So that's that's huge, right? But it's
[38:47] also the case that in some cases you
[38:50] have specific requirements where the
[38:52] tools that you've been using to date
[38:53] they aren't suited for those requirement
[38:56] exactly and in those cases I think yes
[38:59] >> you will have more more bespoke
[39:01] solutions. Yeah,
[39:02] >> but we're I also expect us to see that
[39:06] lovable continues to interoperate with
[39:09] all of those tools. And uh I'm not sure
[39:12] if you tried this if if you ask for
[39:14] connecting to anything in the Google
[39:16] suite or now to anything in the
[39:18] Microsoft suite or or Slack lovable
[39:21] guides you through all the steps to do
[39:22] that in a way where you can get a a very
[39:25] good overview of exactly how the data
[39:27] flows which is of course very important
[39:29] that you don't give access to the wrong
[39:31] person to the wrong data and you can
[39:32] continue to use Salesforce um HubSpot
[39:35] and all the tools that you kind of like
[39:37] to use under the hood but with a bespoke
[39:39] interface on top of H how have these new
[39:41] frontier models they're in some ways
[39:44] competitive but in some ways you can use
[39:48] them to power lovable. So how do you
[39:50] think about the competition with them
[39:54] opensource
[39:55] a and the future of lovable because
[39:57] people have announced that lovable's
[39:59] dead every 6 months since you started
[40:02] and then every 6 months you go from 100
[40:05] to 200 to 300 I think you're at 400
[40:07] million in revenue something crazy. We
[40:09] we we reached 500 in May.
[40:11] >> Okay. Growth is a phenomenal.
[40:13] >> So you're dying again by another 100
[40:16] million in annual revenue.
[40:17] >> Exactly.
[40:18] >> So but underneath the hood you're using
[40:21] some of these.
[40:22] >> Yeah. Let me explain. Yeah. So we've
[40:24] always had this strategy that we do
[40:26] whatever is best for our customers. And
[40:28] in terms of the intelligence that means
[40:31] that we're using multiple models. And so
[40:35] if you ask Lovable now, it's actually
[40:37] routed to the model that's most suitable
[40:40] to whatever you want to do. And that's
[40:42] both the commercial frontier models. So
[40:45] from multiple vendors
[40:46] >> and increasingly it's open weight models
[40:50] where our team when whenever it's get
[40:52] gets routed to an to our own model that
[40:55] model becomes more intelligent for our
[40:57] agent harness. Yeah. especially on the
[40:59] mistakes that it might be making in some
[41:01] cases on which which tool to call, which
[41:04] integration to create and how to guide
[41:06] you through uh success for your
[41:08] business.
[41:09] >> Right. So, you're all in on open source.
[41:12] You believe that's the future of
[41:14] Lovable. I I'm reading into it. So we
[41:17] have multiple partnerships and we're
[41:18] investing heavily to be close with those
[41:20] partners and it's the big the big labs
[41:22] and it's also to make sure that um we
[41:26] get the fastest performance at the
[41:28] lowest cost for our customers when we
[41:31] know that we can do that with our own
[41:32] models
[41:33] >> right
[41:34] >> and uh we have a really really strong
[41:36] research team up in Stockholm who is
[41:38] working on what's called post training
[41:40] so and we're applying all the best
[41:42] practices to do that and scaling up that
[41:44] team uh quite significantly since we
[41:47] also believe it's a it's a part of the
[41:49] European ecosystem to have that
[41:51] capability in Europe specifically.
[41:53] >> Are you doing or are you using any of
[41:55] the data labeling data training
[41:58] companies to help you understand the
[42:01] most common businesses and build that
[42:03] proprietary data? So, so what we're
[42:06] doing is that we're looking at um the
[42:08] mistakes that any of the models do right
[42:10] now and then we we prioritize them by
[42:13] what drives most impact for our
[42:15] customers and then we make the models we
[42:17] create data sets or we um we do did
[42:19] something called reinforcement learning
[42:21] specifically for the problems where the
[42:23] frontier models are making mistakes for
[42:24] us right now and um we have this
[42:28] enormous token distribution right from
[42:31] um a million new products being built
[42:33] every every single
[42:34] You're burning a lot of tokens.
[42:36] >> We are. Yes. And that's and that's a lot
[42:38] of signals for making the system both
[42:41] the agent harness
[42:43] >> and um what we've been refining over the
[42:46] last two years which is the skills that
[42:49] we have have this like internal type of
[42:50] skills that the agent knows when to
[42:53] remember the facts from our software
[42:55] engineers that know how to build really
[42:56] really good software. We're modifying
[42:59] both of those on every every single
[43:00] week.
[43:01] >> It makes total sense. And somebody told
[43:02] me some companies are doing token
[43:05] dumping. They're, you know, selling $100
[43:08] worth of tokens for $50. Um, you know,
[43:12] basically they become token resellers in
[43:14] some ways and they're money losing
[43:16] businesses. You have to you're money
[43:19] you're profitable I believe now or close
[43:21] to it. Um, we we always monitor our
[43:24] margins, but again um we're doing what's
[43:26] best for our customers and that means
[43:28] that often means more intelligence. So
[43:30] we're not we're not looking at oh let's
[43:31] use a we've never had the decision to
[43:34] say let's use a cheaper model here if
[43:37] it's measurably worse for our customers
[43:40] and we can measure that what's best for
[43:41] >> but are they is it unlimited for the 50
[43:43] or you have caps now
[43:44] >> we have caps
[43:46] overages and caps are people starting to
[43:48] hit them
[43:49] >> yeah our customers definitely hit caps
[43:51] and then you can top up you can have a
[43:54] we have multiple subscription tiers
[43:56] >> what number I'm just curious like what
[43:57] percentage of people need to top up.
[43:59] They're so addicted to it that they're
[44:01] blowing past the the
[44:02] >> so from the lowest subscription tier.
[44:05] >> Yeah.
[44:05] >> Um I I think it's the
[44:08] >> uh m
[44:10] it's something like 60% of our customers
[44:12] I think
[44:13] >> I'm hearing that more and more often
[44:14] that people are willing to pay the
[44:17] overages because they're getting so much
[44:19] value. And I think that's the future of
[44:21] the business is people are looking at it
[44:23] going like I am. Well, if I'm paying
[44:26] $600 and if you token max to 6,000 a
[44:30] year, but this is a $500,000 piece of
[44:34] software, I don't care. I'm still paying
[44:35] somewhere between.1%
[44:38] and 1% of what I would have paid 3 years
[44:40] ago. Who cares?
[44:43] >> Go for it. Um, so
[44:45] >> yeah, what we're seeing is everything is
[44:47] about moving moving fast these days and
[44:49] and AI more AI usually lets you move
[44:52] much faster. So the spend is usually
[44:54] worth it.
[44:55] >> Do your customers a final question for
[44:57] you because I'm starting to see this now
[45:00] where multiple people in the
[45:01] organization try to solve the same
[45:03] software problem and they're competing
[45:06] with each other. So like this internet
[45:08] I'm talking about, we built one for
[45:10] Japan.
[45:11] >> Yeah.
[45:11] >> But somebody built the US one. So now I
[45:13] have two pieces of software. So I said
[45:15] to the two different people
[45:18] or do we have did you guys fork each
[45:20] other's code or they're like no we just
[45:22] built two different lovable projects.
[45:25] And I'm like is that the right thing to
[45:28] do because you went faster and I had two
[45:30] swings at bat two different intelligent
[45:32] brilliant people making their version of
[45:35] the software.
[45:36] >> But you would never have done that
[45:38] >> in the previous way of building
[45:40] software. You would have one track of
[45:41] software and you would be building
[45:43] Franken software where you'd be trying
[45:46] to get all the needs into it from the
[45:48] two different groups. I Yeah, I I'm
[45:50] actually a huge fan of very rapid
[45:53] experimentation and I I have a story
[45:56] where for a while I worked at a a place
[45:59] called CERN where they do particle
[46:01] physics. It's it's pretty here in
[46:04] Europe, right? Uh and that's where I was
[46:06] introduced to this concept of
[46:07] co-opetition where they have two
[46:10] actually quite isolated teams working on
[46:13] the same um particle accelerator but
[46:15] different places on it and then they
[46:17] don't share the results until they
[46:18] publish and that way they uh they can
[46:22] kind of over time learn what's working
[46:23] best in the different organizations but
[46:25] you don't get stuck in a local minimum
[46:27] and it's you know free markets work
[46:28] extremely well because of competition
[46:30] and they they they do that in academia
[46:31] as well and now since the engineering is
[46:34] less of the bottleneck. It's more the
[46:35] question of what is the right thing to
[46:37] build. I think it's a great thing to
[46:39] have if you have the sufficiently many
[46:41] humans right to do to try to attempt
[46:45] solving the same problem in different
[46:46] ways. And then if you do that on
[46:48] lovable, what I like to do is I I take I
[46:50] bring up a new project or one of the
[46:51] projects and I I say, "Hey, can you go
[46:53] and just check out this other one and
[46:55] take this these three things that I
[46:57] really like and and bring them bring
[46:59] them over here and maybe even run an a
[47:02] split test, run an experiment to see if
[47:04] it's if it's improves improving the
[47:06] metrics for for our customers we're
[47:07] trying to serve.
[47:08] >> Did you see somebody used Fable to build
[47:10] Fortnite
[47:12] >> and uh
[47:12] >> I've seen the 3D some of the 3D games?"
[47:14] Yeah.
[47:15] >> Yeah. What is your take on, you know,
[47:18] this latest version from Anthropic
[47:19] Fable? I know they're a part or I assume
[47:21] they're a partner. I don't know that.
[47:22] >> Yeah, we use Fable as well as one of the
[47:24] models.
[47:25] >> What do you think of it in terms of
[47:28] compared to the last generation faster,
[47:30] better, both?
[47:32] >> Yeah. Is it a massive step function?
[47:35] Yeah.
[47:35] >> What I've seen is that it can in the
[47:38] first attempt create very sophisticated
[47:41] things that look really good. Then when
[47:43] as you're evolving right it's it's still
[47:45] the same thing where you as a human you
[47:47] have to think you often should be
[47:49] planning together with your agent about
[47:52] what is the right thing to do and and
[47:54] that's more of that's again more of the
[47:55] bottleneck uh whereas more intelligence
[47:58] is on some tasks it's great yeah like it
[48:01] creates really beautiful things 3D games
[48:03] for example but on figuring what to what
[48:06] to build figuring out figuring out what
[48:08] are the right strategic directions or
[48:10] experiments you should run to improve
[48:12] outcomes for your business. That's um uh
[48:15] that's not changing as fast is the
[48:17] humans knowing how to use the tool to
[48:19] get and to plug in all the right data to
[48:22] be able to take the right decisions for
[48:23] taking your product forward and to take
[48:25] your business forward.
[48:26] >> Um listen, I love the product, but even
[48:30] more than I love the product and you as
[48:31] a founder, I love the outcome. The
[48:34] outcome for business is extraordinary.
[48:35] So, anybody who's listening, Lovable is
[48:39] absolutely worth your time. Don't wait.
[48:41] Just put it on your corporate card and
[48:43] start building. That's my message. Just
[48:45] start building with Lovable. It's an
[48:47] incredible product. And uh
[48:49] congratulations on being reborn six
[48:53] times cuz every 6 months you add 100
[48:56] million in revenue it seems. And then
[48:57] everybody says Lovable's dead because
[49:00] the new foundation model is so good. But
[49:02] you keep studying your customer and and
[49:05] you keep somehow surviving and thriving.
[49:07] So congratulations as an entrepreneur.
[49:10] Thank you so much, Jason. I enjoyed that
[49:12] chat. I hope you enjoy the rest of your
[49:14] stay here in Paris.
[49:15] >> It's pretty great. And the Palace of
[49:16] Versailles is so impressive, huh? Uh
[49:18] someday we'll be building this with
[49:20] lovable and optimist robots. I
[49:22] >> I'm looking forward to it. I'm going all
[49:24] in.
[49:39] I'm going all in.
