AI: ASIC Chips Reshape Computing Power
The battle over artificial intelligence is no longer just about the models; it's now about the infrastructure that runs them. The rise of ASIC chips, designed for specific uses, is gradually reshaping the technological balance that has been dominated by GPUs up until now.
The Strategic Shift in Specialized Chips
In April 2025, Google unveiled the seventh generation of its Tensor Processing Unit (TPU), an ASIC-type chip designed specifically for AI-related computations. Used in its own data centers, these chips offer high processing capacity while consuming significantly less energy than traditional graphics processors. Long reserved for internal use, they are now fueling speculation about a potential opening to other cloud players.
This move illustrates a profound shift in the architecture of digital infrastructures. GPUs, versatile and powerful, have played a central role in the rise of generative AI. However, their energy consumption and operational costs are becoming major concerns as computing volumes explode. ASICs, optimized for specific tasks, are emerging as a more efficient solution to these constraints.
An Economic and Energy Logic
For hyperscalers, the equation is becoming increasingly clear. Investing in custom chips carries a high initial cost but allows for a long-term reduction in the operational expenses of data centers. According to Pictet AM, Google's TPUs display two to three times greater energy efficiency than GPUs for certain AI tasks. Meanwhile, Amazon Web Services highlights its Trainium chips, which reportedly offer a price-performance ratio 30% to 40% superior to solutions relying solely on GPUs.
This pursuit of efficiency takes on a strategic dimension in a context of growing constraints on access to affordable and sustainable energy. At scale, performance is no longer measured solely by raw power but by the ability to perform more computations with fewer resources.
Broadcom and Marvell: Unseen Winners of the Transition
Behind this technological evolution, certain industry players are finding their place. Broadcom has established itself as Google's main partner for the design and manufacturing of TPUs. While Google defines the architecture of its chips, Broadcom supplies essential components as part of a close and hard-to-replicate partnership. This position places the group at the heart of the AI ecosystem, helping to reduce the sector's reliance on Nvidia.
Marvell, on the other hand, is experiencing a more mixed period in the markets. Investors are questioning its dependence on a few major clients, which has weighed on its stock price in 2025. However, the company maintains solid technological assets and direct exposure to the growth of the ASIC market. According to Pictet AM, the market is currently applying a waiting discount, pending more tangible evidence of the ramp-up of its solutions.
Toward a More Fragmented Ecosystem
The rise of ASICs does not mark the end of GPUs. Nvidia maintains a dominant position in many applications, particularly for the development and training of new models. However, the AI ecosystem is becoming more fragmented, with growing complementarity between GPUs and specialized chips. Ultimately, companies' ability to turn massive AI investments into measurable productivity gains will be crucial.
In this technological race, the competitive advantage will not only depend on power but also on efficiency, cost control, and vertical integration of infrastructures.
This content has been automatically translated using artificial intelligence. While we strive for accuracy, some nuances may differ from the original French version.
