#490 – State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI
#490 – State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI
Podcast
Listen to Episode
Note: AI-generated summary based on third-party content. Not financial advice. Read more.
Quick Insights

Consider NVIDIA (NVDA) as the primary "picks and shovels" investment for the AI boom, capitalizing on its dominant hardware and entrenched CUDA software ecosystem. Google (GOOGL) is a strong, vertically integrated competitor poised for long-term success by using its own TPU chips to control costs and power its Gemini models. Exercise caution with Meta Platforms (META), as its AI strategy appears to be faltering, creating uncertainty around its future in the foundation model space. The rise of powerful open-weight AI models from China is a key trend that could commoditize the market, favoring companies with unique hardware or data advantages. Finally, look for opportunities in application-layer companies like Box (BOX) or Shopify (SHOP) that are successfully integrating AI into their core products.

Detailed Analysis

NVIDIA (NVDA)

  • The podcast highlights NVIDIA's dominant position in the AI hardware market, referring to the profit margin on their chips as "insane".
  • Their primary competitive advantage, or "moat", is not just their hardware but the CUDA software ecosystem, which has been developed over two decades and is deeply integrated into AI development.
  • Leadership is seen as a key strength, with CEO Jensen Huang's vision and operational focus compared to Steve Jobs at Apple. It's suggested that without him, the deep learning revolution might have been delayed by decades.
  • NVIDIA continues to innovate, developing new chips like the Vera Rubin GPU, which is designed for specific parts of the AI inference process to lower costs.
  • The company's continued success is seen as being tied to the rapid pace of AI innovation. If progress were to slow, competitors with their own custom chips (like Google, Amazon, Microsoft) would have more time to catch up.

Takeaways

  • NVIDIA is presented as the quintessential "picks and shovels" investment for the AI boom. An investment in NVDA is a bet on the continued, rapid expansion of AI compute needs across the entire industry.
  • The CUDA platform represents a powerful and durable competitive advantage that makes it difficult for competitors to displace them, even with comparable hardware.
  • The primary risk for investors is a potential slowdown in the overall pace of AI advancement, which could reduce the urgency for customers to buy NVIDIA's top-of-the-line, flexible chips and allow more specialized, in-house chips to gain ground.

Google (GOOGL)

  • Google's Gemini models are described as "fantastic" and are expected to continue gaining market share on OpenAI's ChatGPT in 2026.
  • A key point is Google's "structural advantage" in AI. By developing its own custom chips (TPUs), it can avoid paying NVIDIA's high margins and control its entire infrastructure stack from the ground up.
  • The hosts note that the Gemini app is improving and is particularly good for fast queries and explanations, positioning it as a strong consumer-facing product.
  • Google Cloud is a major competitor to Amazon's AWS and Microsoft's Azure, and its deep integration with its own AI models and hardware is a significant strength.

Takeaways

  • Google is a vertically integrated AI giant. An investment in GOOGL is a bet on their ability to leverage massive scale, in-house chip design, and vast data resources to compete effectively in the AI race.
  • Their ability to control costs by not relying solely on NVIDIA could become a significant long-term financial advantage, especially as the cost of running AI models at scale becomes a primary concern.
  • Investors should watch the adoption and performance of the Gemini family of models against competitors as a key indicator of their success in capitalizing on their structural advantages.

Anthropic (Private)

  • The hype around their latest model, Claude Opus 4.5, is described as "insane", particularly for its advanced coding capabilities.
  • The company is perceived as the "least chaotic" of the major AI labs, with a strong, focused culture and a clear strategy targeting enterprise customers.
  • The hosts expect Anthropic to see continued success in the enterprise software market due to the quality of its models.
  • A significant risk factor mentioned is a $1.5 billion lawsuit they lost for using copyrighted books to train their models, highlighting the legal and financial risks associated with training data.
  • The CEO, Dario Amodei, is not expected to sell the company, suggesting it will remain a key independent player for the foreseeable future.

Takeaways

  • While Anthropic is a private company, it is a crucial player for public market investors to watch. Its success, particularly in the high-value enterprise and coding markets, puts direct competitive pressure on public companies like Google.
  • The performance of its models serves as a benchmark for the entire industry. If public companies fail to keep pace, they could lose market share in key segments.
  • The legal battles over training data are a systemic risk for the entire AI industry, and the outcome of cases like Anthropic's could set important precedents.

OpenAI (Private)

  • As the creator of ChatGPT, OpenAI benefits from being the incumbent, with strong brand recognition and user "muscle memory."
  • The company is described as operationally "chaotic" but "very good at landing things," with a world-class research division that has produced breakthroughs like Sora (video generation) and O1 (reasoning models).
  • A key challenge mentioned is that OpenAI is "GPU deprived," meaning they are constantly struggling with compute constraints, which impacts their ability to serve a massive user base.
  • Their strategy includes releasing open-weight models (like GPT-OSS) partly to offload compute costs to the community ("we can use your GPUs").

Takeaways

  • OpenAI is the innovation leader in the space, but its reliance on outside partners for compute (Microsoft for cloud, NVIDIA for chips) could be a long-term vulnerability compared to a vertically integrated player like Google.
  • As a private company, its financial success is most directly reflected in its key partner, Microsoft (MSFT). Microsoft's ability to integrate OpenAI's cutting-edge technology into its products (Azure, Office, etc.) is a core part of its investment thesis.
  • The tension between their immense research talent and their operational/compute challenges will be a key dynamic to watch.

Meta Platforms (META)

  • The hosts were critical of Meta's recent AI strategy, declaring "R.I.P. Llama" in reference to their open-source model.
  • Llama was a pioneering and beloved open-weight model, but the discussion suggests Meta fumbled its advantage by focusing too much on chasing benchmarks with huge models that the community couldn't run, leading to backlash.
  • The future of Meta's commitment to open-source AI is considered uncertain due to rumored internal political struggles and a shift in strategy. The hosts speculate there may not be an open-weight Llama 5.

Takeaways

  • Meta's position in the AI race appears to be faltering, specifically on the open-source front where they were once the clear leader.
  • This has created a vacuum that Chinese AI labs have eagerly filled, which could have long-term strategic consequences for the US AI ecosystem.
  • Investors should be cautious and look for a clear and consistent AI strategy from Meta's leadership before assuming they will be a dominant player in the foundation model space.

Investment Theme: US vs. China AI Competition

  • A central theme of the discussion is the escalating competition between US-based AI labs (OpenAI, Anthropic, Google) and a growing number of powerful Chinese labs (DeepSeek, Zhipu AI, Moonshot AI).
  • While US models are currently considered to have a better user experience, Chinese companies are catching up quickly on a technical level.
  • The key strategy for Chinese firms is releasing powerful open-weight models. This allows them to gain global influence and developer adoption, bypassing security concerns that would prevent US companies from subscribing to their APIs directly.

Takeaways

  • This is not just a corporate competition but a geopolitical one. The rise of high-quality, permissively licensed open-source models from China is a direct threat to the business models of Western companies that charge for API access.
  • This trend could lead to the commoditization of AI models, putting pressure on the high valuations of private AI labs and the revenue streams of public tech giants.
  • Investors should monitor the adoption of these open-weight models and the US response, such as the "Atom Project," which aims to bolster the American open-source AI ecosystem.

Investment Theme: AI Applications & Specialization

  • The discussion highlights that massive value will be created not just by the general-purpose models, but by specialized applications built on top of them.
  • Coding is a prime example. Tools like Cursor (Private) and Anthropic's Claude Code are creating huge productivity gains. A survey showed that 80% of professional developers find their work more enjoyable with AI, and senior developers are heavily using AI-generated code.
  • Enterprise AI is another key area. The ability of companies like Box (BOX) and Shopify (SHOP) to effectively integrate AI into their core products is seen as a critical factor for future success.
  • The podcast suggests that the next wave of value will come from specializing models for high-value industries like finance, law, and pharmaceuticals using proprietary data.

Takeaways

  • Beyond the big model builders, investors should look at companies that are effectively using AI to enhance their existing products and workflows. These "application layer" companies could be major beneficiaries of the AI boom.
  • The productivity gains in software development are real and happening now. This could lead to faster innovation cycles and lower operating costs for all tech companies.
  • The real "moat" for many businesses in the future may be their unique, proprietary data, which they can use to fine-tune foundation models for a competitive advantage that general-purpose models cannot replicate.
Ask about this postAnswers are grounded in this post's content.
Episode Description
Nathan Lambert and Sebastian Raschka are machine learning researchers, engineers, and educators. Nathan is the post-training lead at the Allen Institute for AI (Ai2) and the author of The RLHF Book. Sebastian Raschka is the author of Build a Large Language Model (From Scratch) and Build a Reasoning Model (From Scratch). https://lexfridman.com/sponsors/ep490-sc Transcript: https://lexfridman.com/ai-sota-2026-transcript CONTACT LEX: Feedback – give feedback to Lex: https://lexfridman.com/survey AMA – submit questions, videos or call-in: https://lexfridman.com/ama Hiring – join our team: https://lexfridman.com/hiring Other – other ways to get in touch: https://lexfridman.com/contact SPONSORS: Box: Intelligent content management platform. https://box.com/ai Quo: Phone system (calls, texts, contacts) for businesses. https://quo.com/lex UPLIFT Desk: Standing desks and office ergonomics. https://upliftdesk.com/lex Fin: AI agent for customer service. https://fin.ai/lex Shopify: Sell stuff online. https://shopify.com/lex CodeRabbit: AI-powered code reviews. https://coderabbit.ai/lex LMNT: Zero-sugar electrolyte drink mix. https://drinkLMNT.com/lex Perplexity: AI-powered answer engine. https://perplexity.ai/ OUTLINE: (00:00) – Introduction (01:39) – Sponsors, Comments, and Reflections (16:29) – China vs US: Who wins the AI race? (25:11) – ChatGPT vs Claude vs Gemini vs Grok: Who is winning? (36:11) – Best AI for coding (43:02) – Open Source vs Closed Source LLMs (54:41) – Transformers: Evolution of LLMs since 2019 (1:02:38) – AI Scaling Laws: Are they dead or still holding? (1:18:45) – How AI is trained: Pre-training, Mid-training, and Post-training (1:51:51) – Post-training explained: Exciting new research directions in LLMs (2:12:43) – Advice for beginners on how to get into AI development & research (2:35:36) – Work culture in AI (72+ hour weeks) (2:39:22) – Silicon Valley bubble (2:43:19) – Text diffusion models and other new research directions (2:49:01) – Tool use (2:53:17) – Continual learning (2:58:39) – Long context (3:04:54) – Robotics (3:14:04) – Timeline to AGI (3:21:20) – Will AI replace programmers? (3:39:51) – Is the dream of AGI dying? (3:46:40) – How AI will make money? (3:51:02) – Big acquisitions in 2026 (3:55:34) – Future of OpenAI, Anthropic, Google DeepMind, xAI, Meta (4:08:08) – Manhattan Project for AI (4:14:42) – Future of NVIDIA, GPUs, and AI compute clusters (4:22:48) – Future of human civilization
About Lex Fridman Podcast
Lex Fridman Podcast

Lex Fridman Podcast

By Lex Fridman

Conversations about science, technology, history, philosophy and the nature of intelligence, consciousness, love, and power. Lex is an AI researcher at MIT and beyond.