Toward Recursive Improvement in Artificial Intelligence

One of the most fascinating developments in AI is what researchers call recursive self-improvement (RSI), a process in which artificial intelligence itself contributes to the development of its next generations. We are already beginning to see its effects, with Anthropic, for example, revealing in its IPO prospectus that more than 90% of contributions to its computer code in September were written by Claude.

But while a model can optimize its algorithms to improve them and reduce the amount of computing power required, it remains constrained by the physical characteristics of the chips on which it operates, including their memory capacity, bandwidth and the speed at which they communicate with one another.
This explains the growing interest among AI labs in also mastering semiconductor design. The goal is to create a feedback loop between software and hardware, in which more intelligent models help design higher-performance chips, which in turn accelerate the development of even more intelligent models.

This is where Nvidia's business model could show its limitations, because the group must design processors that are sufficiently versatile to meet the needs of thousands of customers, while a lab developing its own chips can optimize them around the characteristics of its future models, even before those models are operational.

Beyond the potential savings from partially bypassing Nvidia's margins, which are close to 75%, the goal is therefore to develop a technological advantage that competitors could not simply buy. OpenAI plans to deploy 10 gigawatts of accelerators developed with Broadcom, while Anthropic was already using more than 1 million of Amazon's Trainium2 chips in April. According to TrendForce, shipments of application-specific integrated circuits for AI (ASICs) could even exceed those of GPUs as early as 2027.

Google in pole position in the race for vertical integration

It is from this perspective that Google's position appears particularly interesting. The group has been developing its own accelerators, known as TPUs, for more than a decade and can design its chips around Gemini's future needs, while adapting its models to the characteristics of its next-generation chips.
Because each semiconductor has a limited area and energy budget, Google can devote more resources to the components that constrain the performance of its models. Its TPU 8i, specifically designed for inference, for example has three times more integrated memory and, according to the group, delivers a performance-to-price ratio up to 80% higher than that of the previous generation.

This feedback loop is no longer purely theoretical. AlphaEvolve, an agent developed by DeepMind, contributed to the design of a circuit intended for a future generation of TPUs and optimized a computing operation by 23%, reducing Gemini's total training time by 1%. Google has also been using AlphaChip since 2020 to optimize the physical arrangement of components in its TPUs.

In addition, advances in artificial intelligence could weaken CUDA, the software ecosystem that has historically represented one of Nvidia's main barriers to entry. A scientific study published in July showed that specialized models could automatically convert CUDA code to AMD's software environment with an 88.2% success rate, achieving performance comparable to native code in 85% of the cases studied.

Nvidia forced to compete with AI labs

Nvidia's investments in its Nemotron models and its planned $12.9bn acquisition of Hugging Face make perfect sense in this context. If proprietary large language models dominate the market, companies could essentially consume AI as a service, with the computing performed on the labs' accelerators. By supporting the emergence of sufficiently capable open-source models, Nvidia is instead seeking to preserve an ecosystem in which companies continue to purchase or lease its GPUs.

This is probably where Jensen Huang's absence could be felt most strongly. At the end of October, Nvidia's multiyear commitments related to cloud services and data-center leases had reached $102bn, up from $74bn three months earlier, with some of this infrastructure intended to support the development of its own models. Will his successor be equally willing to sacrifice part of the group's exceptional profitability to invest downstream in the value chain and support an ecosystem on which demand for its semiconductors ultimately depends?

Google also has a distribution force that is difficult to match

What makes Google's case even more interesting is that the group does not necessarily need to win the technological battle to emerge as the winner of the AI race. With 13 products each boasting more than 1 billion users, Google can integrate its agents directly into the applications where we already perform a large portion of the tasks they are supposed to automate.

While Nvidia still has to establish itself among the leading AI labs, and OpenAI and Anthropic still have to industrialize the design of their own semiconductors, Google already controls the main links in the value chain, from TPU chips to Gemini models, as well as its cloud infrastructure and distribution channels. And if recursive improvement does indeed amplify the advantages of this vertical integration, Alphabet could ultimately capture an increasing share of the economic rents that currently drive Nvidia's fortune.

https://www.marketscreener.com/news/why-nvidia-could-lose-without-jensen-huang-and-google-become-ai-s-biggest-winner-ce785ed2dc81f42d