Former Intel CEO Pat Gelsinger on Google AI chips: Competition is good for all

Key Concepts

• Tensor Processing Unit (TPU): Google's custom-designed hardware accelerator for machine learning.
• Artificial Intelligence (AI): The simulation of human intelligence processes by machines, especially computer systems.
• Large Language Models (LLMs): AI models trained on vast amounts of text data to understand and generate human-like language.
• Proprietary Data Center: A data center owned and operated by a specific company for its exclusive use.
• Commercial Chip Provider: A company that designs and sells chips to external customers.
• Circular Transactions: Business deals where companies invest in each other, often to secure supply chains or partnerships.
• Moat (in AI): A sustainable competitive advantage that protects a company's market position.
• Multi-modal Experiences: AI systems that can process and understand information from multiple types of data (e.g., text, images, audio).
• Mixture of Experts (MoE): An AI architecture where multiple specialized models work together to solve a problem.

Google's TPU and Market Entry

Pat Gelsinger discusses the significant development of Google potentially challenging NVIDIA's dominance in the AI chip market, specifically with their Tensor Processing Unit (TPU). He highlights that Google has developed seven generations of TPUs, which are custom-designed for machine learning tasks. A crucial aspect of this development is Google's partnership with Broadcom. Gelsinger emphasizes that while Google has historically used TPUs for its proprietary data centers, the partnership with Broadcom is critical for bringing these chips to market at scale. This move signifies Google's intent to become a commercial chip provider, a departure from their previous focus on internal use. Gelsinger views this as positive news for Google, Broadcom, and ultimately for the AI market by fostering competition.

NVIDIA's Competitive Stance and Broadcom's Role

The conversation touches upon NVIDIA's aggressive and competitive nature. Gelsinger acknowledges that Google's TPU is a viable competitor, and it's natural for companies to seek alternatives to NVIDIA. He interprets NVIDIA's tweet as an indication that they welcome competition. However, Gelsinger stresses the difference between building a proprietary chip for a private data center and making it commercially available for others. He notes that commercializing a chip for external data centers involves significantly more work, drawing from his 35 years of experience. This is why the Broadcom relationship is deemed essential for Google to fully realize the potential of its TPUs for Meta and other major data centers.

Circular Transactions and Revenue Quality

The discussion shifts to "circular transactions," exemplified by the investments involving Anthropic, Microsoft, Amazon, and Google. Anthropic, a leading large language model, has received investments from all three major tech players. Gelsinger expresses skepticism about the quality of revenue generated from these circular deals. He argues that if a company is committing its capital to buy your product, it's not as valuable as a direct commitment of capital to purchase your product. While acknowledging that these companies have strong balance sheets and are using them creatively to enable capital deployment in AI, Gelsinger believes markets need to assess the true value of such revenue due to its circular nature. He cautions that if these transactions become too extreme, it could be detrimental to markets overall, but they have served as a creative mechanism for capital deployment in the AI sector.

The Future of Large Language Models (LLMs)

The conversation explores the competitive landscape of new large language models (LLMs), with mentions of Google's Gemini, Anthropic's upgrades, and anticipated developments from ChatGPT. Marc Benioff's endorsement of Gemini is noted. Gelsinger expresses excitement about these innovations but suggests that simply making larger LLMs may be encountering diminishing returns. He posits that dedicated models, multi-modal experiences, and "Mixture of Experts" (MoE) architectures are more likely areas for future breakthroughs. Gelsinger believes that LLMs themselves will eventually give way to other forms of knowledge representation, which will drive the next wave of advancements in AI learning.