LLMs as the New CPUs: The Core of the Generative AI Era

In the same way that the Central Processing Unit (CPU) became the foundational element of personal computing, Large Language Models (LLMs) are emerging as the core component driving the generative AI era. Just as CPUs revolutionized computing by performing a vast array of tasks efficiently and reliably, LLMs are transforming the landscape by enabling machines to understand and generate human-like text with unprecedented sophistication. This analogy helps frame the broader technological ecosystem evolving around LLMs, shaping the future of innovation and industry.

The Operating System: AI Platforms

If LLMs are the new CPUs, AI platforms serve as the operating systems (OS) that provide the necessary environment for these models to operate efficiently and interact with other components. These platforms manage resources, ensure compatibility, and offer a suite of tools and services that facilitate the development, deployment, and integration of AI solutions. Key players in this space include OpenAI's platform, Google's TensorFlow Extended (TFX), and Microsoft's Azure AI, among others.

Satya Nadella, in a recent Stratechery interview, compared this to the traditional debate between vertical integration and modularization. He referenced historical business models, contrasting the vertically integrated approach of Apple (and now Google in AI) with the more modular strategies of companies like Microsoft.

Integrated AI: Google’s Approach

Google represents the vertically integrated approach in the AI space, analogous to Apple in the consumer electronics market. Google controls every aspect of its AI stack, from the custom TPU processors on which its Gemini models run, to the Vertex AI platform that developers use to access these models. This integration allows Google to optimize performance across the entire stack, exemplified by Gemini 1.5’s industry-leading 2 million token context window, likely a result of deep collaboration between Google’s infrastructure and model-building teams.

This strategy mirrors the early days of computing, where firms like IBM had to integrate design and manufacturing to meet performance demands. As Professor Clayton Christensen noted, when product functionality is not yet sufficient to meet customer needs, integrated systems have a competitive edge because they can optimize performance beyond what modular systems can achieve.

Modular AI: AWS and Microsoft

In contrast, AWS embodies the modular approach. AWS offers flexibility through its Bedrock managed development platform, which supports a variety of models, including its own Titan family and third-party models running primarily on Nvidia GPUs. This strategy leverages the modularity of different AI components, allowing customers to mix and match based on their needs.

Microsoft's position is somewhat intermediate. While it heavily integrates with OpenAI’s GPT models through its Azure cloud, it also supports a broad range of AI services, positioning itself between the fully integrated approach of Google and the modular strategy of AWS. This hybrid model reflects a balance between leveraging specialized integrations and maintaining flexibility.

Implications for Big Tech

Google: As Nadella noted, Google’s strategy of deep integration mirrors Apple’s successful model in hardware. However, the challenge lies in whether this approach will resonate in AI, especially given that Google’s products are not physical devices but internet services. Google’s potential to leverage its integrated approach in consumer AI, such as through its Pixel phones, could redefine its market positioning.

AWS: Amazon bets on modularization, believing that data gravity—where data resides—will drive the adoption of its AI services. This aligns with AWS's strength in providing scalable, flexible infrastructure, though it remains to be seen how this strategy will fare against more integrated competitors.

Microsoft: Straddling both approaches, Microsoft’s collaboration with OpenAI allows it to benefit from integration while offering a broad array of AI services. This strategy provides a hedge against the risks of relying too heavily on a single partner and allows Microsoft to adapt as the AI landscape evolves.

Meta and Others: Meta’s open-source approach with models like Llama focuses on product optimization and broad usage, leveraging community contributions without diverting from its core product goals. Databricks, with its focus on data integration and custom model training, highlights the importance of data as a critical point of integration in AI.

Conclusion

The debate between vertical integration and modularity in the generative AI era reflects broader themes in technology and business strategy. As LLMs become central to AI development, the platforms that support them—acting as operating systems—will shape the future of innovation. Whether through the integrated approaches of Google and Apple, the modular strategies of AWS and Microsoft, or hybrid models, the key to success will be balancing performance optimization with flexibility and scalability. As AI continues to evolve, these strategic decisions will determine which companies lead the next wave of technological transformation.

Acknowledgements

Special thanks to Ben Thomson for his clarity and insights through Stratechery, which greatly shaped my understanding of the generative AI landscape. Your work is invaluable, Ben. Thank you!